Conference Paper

Deep Learning Neural Networks for Land Use Land Cover Mapping

Authors:
To read the full-text of this research, you can request a copy directly from the authors.

No full-text available

Request Full-text Paper PDF

To read the full-text of this research,
you can request a copy directly from the authors.

... 3,7 To resolve the semantic segmentation task of LULC mapping, Long et al. 8 and Badrinarayananet al. 9 introduced an encoder-decoder architecture, which extracts features from the imagery using CNN in the form of an encoder and then decodes those features back to the original size using a reversed CNN. The implementation of encoder-decoder networks in remote sensing, [10][11][12][13][14] has achieved great success with accurate and consistent results between 80% and 90% for LULC mapping. These successes have led to the implementation of DL models for mapping of LULC within commercial software, for example, ArcPro (Esri). ...
... One of the first to successfully use semantic segmentation architectures for remote sensing applications were. [12][13][14]39 Their solution is based on an FCN structure with different encoders (i.e. VG-GNet, GoogleNet, ResNet) used on Landsat 5/7 satellite images. ...
... By creating tiles of 224 x 224 pixels, there is a suitable number of images to train a DL model, and images of this size can be processed by our computer hardware. Consequently, we used the data augmentation technique called tiling, which was used in, [12][13][14]39 with some minor improvements. ...
Preprint
Full-text available
Deep learning semantic segmentation algorithms have provided improved frameworks for the automated production of Land-Use and Land-Cover (LULC) maps, which significantly increases the frequency of map generation as well as consistency of production quality. In this research, a total of 28 different model variations were examined to improve the accuracy of LULC maps. The experiments were carried out using Landsat 5/7 or Landsat 8 satellite images with the North American Land Change Monitoring System labels. The performance of various CNNs and extension combinations were assessed, where VGGNet with an output stride of 4, and modified U-Net architecture provided the best results. Additional expanded analysis of the generated LULC maps was also provided. Using a deep neural network, this work achieved 92.4% accuracy for 13 LULC classes within southern Manitoba representing a 15.8% improvement over published results for the NALCMS. Based on the large regions of interest, higher radiometric resolution of Landsat 8 data resulted in better overall accuracies (88.04%) compare to Landsat 5/7 (80.66%) for 16 LULC classes. This represents an 11.44% and 4.06% increase in overall accuracy compared to previously published NALCMS results, including larger land area and higher number of LULC classes incorporated into the models compared to other published LULC map automation methods.
... The PSPNet and UNet performed best. On the other hand, Storie and Henry (2018) obtained the best results with a FCN-8 (Long et al., 2015). Additionally, Stivaktakis et al. (2019) showed that CNN models with prior augmentation outperform all CNN models without augmentation. ...
... To compare the classification quality of different CNN models, we implemented UNet, PSPNet, SegNet, and FCN-8. As mentioned above, these models already achieved good results in LULC classification (Zhang et al., 2020;Storie and Henry, 2018). We implemented them For training, we used 80% of randomly selected image pairs from the overall data. ...
... This result is also in line with the publication of Zhang et al. (2020) on the LULC classification, where UNet also performed best. In the study of Storie and Henry (2018), the FCN-8 model performed best, but it is not known exactly with which other models the comparison was made. ...
Article
Full-text available
We present the results from evaluating various Convolutional Neural Network (CNN) models to compare their usefulness for forest type classification. Machine Learning based on CNNs is known to be suitable to identify relevant patterns in remote sensing imagery. With the availability of free data sets (e.g. the Copernicus Sentinel-2 data), Machine Learning can be utilized for forest monitoring, which provides useful and timely information helping to measure and counteract the effects of climate change. To this end, we performed a case study with publicly available data from the federal state of North Rhine-Westphalia in Germany. We created an automated pipeline to preprocess and filter this data and trained the CNN models UNet, PSPNet, SegNet, and FCN-8. Since the data contained large rural areas, we augmented the imagery to improve classification results. We reapplied the trained models to the data, compared the results for each model, and evaluated the effect of augmentation. Our results show that UNet performs best with a categorical accuracy of 73% when trained with augmented imagery.
... The dataset EuroSAT contains 27,000 labeled and geo-referenced images made public. Furthermore, the study proposed in [101] exploits deep learning neural networks for remote sensing to improve the production of land cover maps and land use. In particular, the authors of [101] consider the southern agricultural region of Manitoba in Canada. ...
... Furthermore, the study proposed in [101] exploits deep learning neural networks for remote sensing to improve the production of land cover maps and land use. In particular, the authors of [101] consider the southern agricultural region of Manitoba in Canada. ...
... Deep learning has recently been widely used by researchers to respond to the space sector's countless challenges. For example, neural networks are used to improve the land use for agriculture [101,109], mitigate the consequences of climate change [138,140], and monitor water quality [147]. Another example is the use of CNNs [67,68] to assist astronauts during a mission by facilitating the astronauts' work and lives through an intelligent assistant. ...
Article
Full-text available
Artificial intelligence is applied to many fields and contributes to many important applications and research areas, such as intelligent data processing, natural language processing, autonomous vehicles, and robots. The adoption of artificial intelligence in several fields has been the subject of many research papers. Still, recently, the space sector is a field where artificial intelligence is receiving significant attention. This paper aims to survey the most relevant problems in the field of space applications solved by artificial intelligence techniques. We focus on applications related to mission design, space exploration, and Earth observation, and we provide a taxonomy of the current challenges. Moreover, we present and discuss current solutions proposed for each challenge to allow researchers to identify and compare the state of the art in this context.
... Deep learning has exhibited remarkable accuracy in computer vision tasks and holds tremendous potential for efficiently processing vast amounts of earth observation satellite image data in automated workflows. [24]. ...
... This will be evaluated by using the F1-score Accuracy and Intersection over Union The evaluation metrics mentioned in section 4.3 will be used to evaluate the first requirement stated in above Table 3.2 to measure the accuracy of the proposed model of detecting spatial information with highly accurate results and a high pixel-wise accuracy from satellite imagery. 24 ...
Thesis
Full-text available
Detecting spatial information from satellite imagery using deep learning for semantic segmentation is an important field that is significantly growing due to its importance in applications such as the automated generation of vector maps, urban planning, and geographic information systems. In this research, the utilization of deep learning for the semantic segmentation of spatial information from satellite imagery is explored. The objective is to devise an efficient and precise method for detecting and categorizing diverse features on the Earth's surface, including road networks, building footprints, water bodies, vegetation, and land cover which can be used in automatic map production. The proposed technique entails training a deep convolutional neural network to detect spatial features from a small dataset of satellite imagery, followed by a segmentation process to classify the various spatial features. This study conducts various experiments on satellite imagery to achieve high accuracy rates that outperform traditional image processing techniques. In addition, this project also compares various models such as networks with U-shaped architecture U-Net and modified U-Net (Inception ResNetV2U-Net) with various spatial features. Both Implemented models achieved higher results than other relevant research papers. Although the Inception ResNetV2U-Net model produced slightly better results than U-Net, with a validation accuracy of 87.5% and a validation coefficient of 87%, the U-Net model achieved also high validation accuracy and coefficient of 86.5% and 84%, respectively. Additionally, the U-Net model exhibited significantly improved and better training and validation loss than ResNetV2U-Net. Furthermore, the U-Net model showed a shorter average prediction time of satellite imagery. Therefore, the U-Net model is proven to be more suitable for detecting spatial information from small satellite datasets.
... Considering the advantage of DL in dealing with spectral information and spatial land cover features, a possible alternative to address spectral-spatial errors lies in end-to-end, image-based, and DL-based SRM (EIDS) models. Attempts have included CNNs [38][39][40][41][42], GANs [43,44], and deep residual networks (DRNs) [45]. Consequently, EIDSs avoid spectral errors associated with the production by spectral unmixing. ...
... First, most EIDS models, including CNNs [39][40][41], GAN [44], and DRN [45], directly learn the nonlinear relationship between land cover classes and the reflectance of the remotely sensed imagery, and view land cover mapping as an imagesegmentation task. In this way, the effect of spatial variability in spectral properties of land covers in remotely sensed imagery has not been fully considered. ...
Article
Full-text available
Super-resolution mapping (SRM) can effectively predict the spatial distribution of land cover classes within mixed pixels at a higher spatial resolution than the original remotely sensed imagery. The uncertainty of land cover fraction errors within mixed pixels is one of the most important factors affecting SRM accuracy. Studies have shown that SRM methods using deep learning techniques have significantly improved land cover mapping accuracy but have not coped well with spectral-spatial errors. This study proposes an end-to-end SRM model using a spectral-spatial generative adversarial network (SGS) with the direct input of multispectral remotely sensed imagery, which deals with spectral-spatial error. The proposed SGS comprises three parts: (1) Cube-based convolution for spectral unmixing is adopted to generate land cover fraction images. (2) A residual-in-residual dense block fully and jointly considers spectral and spatial information and reduces spectral errors. (3) A relativistic average GAN is designed as a backbone to further improve super-resolution performance and reduce spectral-spatial errors. SGS was tested in one synthetic and two realistic experiments with multi-/hyper-spectral remotely sensed imagery as the input, comparing the results with those of hard classification and several classic SRM methods. The results showed that SGS performed well at reducing land cover fraction errors, reconstructing spatial details, removing unpleasant and unrealistic land cover artifacts, and eliminating false recognition.
... An important drawback when dealing with satellite images by means of machine learning methods is the need for labeled data [24]. Most of the works have been focused on supervised classification techniques. ...
... Finally, it is worth mentioning that the clustering and classification of SITS is closely related with the automatic generation of land cover maps, which has experienced a rapid development during the last decade [36], [37] due to the increasing availability and quality of satellite imagery data. Also in this more specific context, the development of innovative methods based on deep learning techniques opens up new research opportunities [24], [7], [38]. ...
Preprint
Full-text available
Satellite images constitute a highly valuable and abundant resource for many real world applications. However, the labeled data needed to train most machine learning models are scarce and difficult to obtain. In this context, the current work investigates a fully unsupervised methodology that, given a temporal sequence of satellite images, creates a partition of the ground according to its semantic properties and their evolution over time. The sequences of images are translated into a grid of multivariate time series of embedded tiles. The embedding and the partitional clustering of these sequences of tiles are constructed in two iterative steps: In the first step, the embedding is able to extract the information of the sequences of tiles based on a geographical neighborhood, and the tiles are grouped into clusters. In the second step, the embedding is refined by using the neighborhood defined by the clusters, and the final clustering of the sequences of tiles is obtained. We illustrate the methodology by conducting the semantic clustering of a sequence of 20 satellite images of the region of Navarra (Spain). The results show that the clustering of multivariate time series is robust and contains trustful spatio-temporal semantic information about the region under study. We unveil the close connection that exists between the geographic and embedded spaces, and find out that the semantic properties attributed to these kinds of embeddings are fully exploited and even enhanced by the proposed clustering of time series.
... However, these methods utilizing spatial and semantic information have lower accuracy and weaker segmentation capabilities, which prevent ultrahigh resolution image segmentation. In recent years, deep learning image analysis has made great progress [12][13][14] in facilitating automated interpretation of high-resolution remote sensing images. ...
... Zhang et al. [33] reported a robust 3-D medical watermarking based on wavelet transform for data protection, Sun et al. [34] reported robust reversible audio watermarking scheme for telemedicine and privacy protection. Although the foregoing research yielded robust results in building feature extraction, two serious challenges remained to be addressed [14]: 1 The building feature segmentation method via post-processing steps is too complex and the integration between modules is difficult; 2 The method of extracting different features via multiple different networks and combining these features is hindered by complex networks, increased need for hardware equipment and a long learning curve. ...
Article
Full-text available
In order to accurately segment architectural features in high-resolution remote sensing images, a semantic segmentation method based on U-net network multi-task learning is proposed. First, a boundary distance map was generated based on the remote sensing image of the ground truth map of the building. The remote sensing image and its truth map were used as the input in the U-net network, followed by the addition of the building ground prediction layer at the end of the U-net network. Based on the ResNet network, a multi-task network with the boundary distance prediction layer was built. Experiments involving the ISPRS aerial remote sensing image building and feature annotation data set show that compared with the full convolutional network combined with the multi-layer perceptron method, the intersection ratio of VGG16 network, VGG16 + boundary prediction, ResNet50 and the method in this paper were increased by 5.15%, 6.946%, 6.41% and 7.86%. The accuracy of the networks was increased to 94.71%, 95.39%, 95.30% and 96.10% respectively, which resulted in high-precision extraction of building features.
... Recent advances in deep learning have greatly enhanced remote sensing image analysis, especially for land cover mapping [12,13], disaster monitoring [14,15], and change detection [16,17]. Deep learning techniques, particularly convolutional neural networks (CNNs) [18], have become a leading approach for automated cultivated land extraction [19]. ...
Article
Full-text available
Approximately 24% of the global land area consists of mountainous regions, with 10% of the population relying on these areas for their cultivated land. Accurate statistics and monitoring of cultivated land in mountainous regions are crucial for ensuring food security, creating scientific land use policies, and protecting the ecological environment. However, the fragmented nature of cultivated land in these complex terrains challenges the effectiveness of existing extraction methods. To address this issue, this study proposed a cascaded network based on an improved semantic segmentation model (DeepLabV3+), called Cascade DeepLab Net, specifically designed to improve the accuracy in the scenario of fragmented land features. This method aims to accurately extract cultivated land from remote sensing images. This model enhances the accuracy of cultivated land extraction in complex terrains by incorporating the Style-based Recalibration Module (SRM), Spatial Attention Module (SAM), and Refinement Module (RM). The experimental results using high-resolution satellite images of mountainous areas in southern China show that the improved model achieved an overall accuracy (OA) of 92.33% and an Intersection over Union (IoU) of 82.51%, marking a significant improvement over models such as U-shaped Network (UNet), Pyramid Scene Parsing Network (PSPNet), and DeepLabV3+. This method enhances the efficiency and accuracy of monitoring cultivated land in mountainous areas and offers a scientific basis for policy formulation and resource management, aiding in ecological protection and sustainable development. Additionally, this study presents new ideas and methods for future applications of cultivated land monitoring in other complex terrain regions.
... Each neuron's value is modified by the weights of the connecting nodes to determine how the input values are converted into output values [19]. A neural network with many more layers and parameters than a particular neural network is referred to as "deep learning" [20]. The structure of deep learning is shown in Figure 1. ...
Article
Full-text available
Information on land use and cover needs to be gathered due to the growing urban population, city growth, and urbanization. Applications for this data include environmental protection, urban planning, planning for urban infrastructure, and strategic planning to guarantee the sustainable growth of urban areas. The primary source of data on land cover and land use at the moment is remote sensing imagery. Information about land cover and land use can be retrieved from remote sensing images using image classification techniques. In terms of classification accuracy, deep learning techniques recently outperformed other methods for classifying land use and cover. Convolutional neural networks (CNNs), which are quite popular in this field, are one of the significant deep learning classification architectures frequently used in land cover and land use classification. Recently, the convolutional neural network technique known as ResNet has been applied to remote sensing applications, particularly for the classification of land use and cover. ResNet models are an effective choice for classifying land cover and land use because they can handle the vanishing gradient issue. The primary objective of this study is to assess the performance of the Glorot Uniform and Random Uniform weight initializers in the ResNet50, ResNet101, and ResNet152 architectures for extracting the land cover and land use of the EuroSat dataset. The weighted F1 score, IoU indexes, overall accuracy, and kappa coefficient were used to evaluate the accuracy of the results. ResNet101's corresponding values for these indexes were, in turn, 0.8869, 0.7951, 0.8871, and 0.8743. These results indicate that, in terms of classification accuracy, ResNet101 has outperformed the ResNet50 and ResNet152 methods.
... DL models are known for their high accuracy, scalability, and ability to generalize spatial and temporal features. However, DL models are less interpretable, computationally expensive, prone to overfitting, and demand substantial volumes of labeled training data, which can often be limited or costly to acquire for some regions or LULC types (Storie & Henry, 2018;Parente et al., 2019;Ienco et al., 2019;Ma et al., 2019). ...
Article
Multitemporal imagery offers a critical advantage by capturing seasonal variations, which are essential for differentiating between land use and land cover (LULC) types. While these types may appear similar when examined at one specific time, they exhibit distinct phenological patterns across different seasons. This temporal depth is crucial for enhancing model accuracy, particularly in heterogeneous landscapes where LULC transitions are frequent and complex. This paper made use of spectral bands and indices of Sentinel-2 from April to September 2020 were utilized for LULC classification using two advanced machine learning models: Random forest (RF) and support vector machine (SVM). The spectral indices include the normalized difference vegetation index (NDVI), normalized difference built-up index (NDBI), and modified normalized water index (MNDWI). The dataset was divided into four temporal feature sets: April-May, June-July, August-September, and the entire period from April-September. For each two-month period, the median values of the spectral bands and indices were used. Both models were evaluated based on overall accuracy, F1-score, Kappa coefficient, precision, and recall. Results indicate that incorporating temporal features enhanced the performance of the chosen models, with overall accuracy increasing from 82.4% to 94.03% for RF and from 75.4% to 93.54% for SVM. Additionally, the RF algorithm demonstrated higher accuracy than the SVM model across various time periods, with notable increases in F1 scores, Kappa statistic, precision, and recall. These improvements underscore the ability of the models to leverage rich temporal and spectral data provided by Sentinel-2 for accurate LULC classification. This study highlights the importance of incorporating temporal dynamics in remote sensing applications to enhance the precision and reliability of LULC classification.
... The application of GANs in Environmental Sciences and Ecology has developed innovative solutions for species distribution modelling, land cover prediction, environmental data analysis, and water quality monitoring, thereby enhancing our understanding and management of ecosystems and the environment [131][132][133]. ...
Preprint
Full-text available
Generative adversarial networks (GANs) have become a recent and rapidly developing research topic in Machine Learning. Since their inception in 2014, a significant number of variants have been proposed to address various topics across many fields, and has particularly excelled not only in image and language processing, but also in the medical and data science domains. In this paper, we aim to highlight the significance and advance that these GAN models can introduce in the field of Business Economics, where they have yet to be fully developed. To this end, a review of the literature of GANs is presented in general together with a more specific review in the field of Business Economics wherein only a few papers can be found. Furthermore, the most relevant papers are analysed in order to provide an approach the opportunity to research into GANs in the field of Business Economics.
... Several studies (Ben Hamida et al., 2018;Corbane et al., 2021;Helber, Bischke, Dengel, & Borth, 2019;Sharma, Liu, Yang, & Shi, 2017;Storie & Henry, 2018;Syrris et al., 2019) have been exploring DL, through CNNs, for automatically extracting features and classifying land cover using images from Unmanned Aerial Vehicle (UAV), and satellite images. Sharma et al. (2017) accomplished an accuracy of 85.60% by training a small CNN model on patch-based samples of 5x5 pixels from Landsat-8 imagery to classify 8 distinct classes. ...
Preprint
Full-text available
The focus of this paper is to present an innovative approach to classify land cover, with specific emphasis on Land Use, Land-Use Change and Forestry (LULUCF) monitoring. LULUCF is a sector in greenhouse gas inventory that tracks changes in greenhouse gas levels in the atmosphere due to land use and land-use change. In this study, we employed Deep Learning classifiers and Random Forest to classify land cover/land use in Czechia, adhering to LULUCF regulations. We evaluated the effectiveness of 2D and 3D Convolutions in Convolutional Neural Networks, with varying filter sizes and training methods, alongside the use of Random Forest classifier. We used Sentinel-2 bands with 10 m and 20 m spatial resolution, NDVI, NDVI variance, and SRTM altitude data to create input paths of 5×5 pixels. The results indicate that the 3D model trained with classical training and 3×3-pixel filters achieved the best F1 Score of 0.84. One significant advantage of using convolutional neural networks is their ability to include information from a pixel’s neighbourhood in the classification process, in contrast to solely considering the pixel itself.
... Long et al. tested the Fully Convolutional Network (FCN) for semantic segmentation and scene parsing on different datasets, such as PASCAL VOC, NYUDv2, SIFT Flow, and found that using classification networks and multi-resolution layer combinations can improve accuracy in image segmentation and speed up the task [32]. Luus [34,35]. Chen et al. demonstrated that Deeplabv3 is an effective tool for LULC segmentation, providing excellent accuracy and precision [36]. ...
Article
Full-text available
Flood is the most frequently occurring and dangerous natural disaster, which leads to loss of human life, economic loss, and agricultural loss. It also has an impact on a variety of services, including health, education, and transportation etc. So, in order to give assistance and conduct rescue operations promptly, flood detection, its mapping, and flood damage assessment are crucial duties. Additionally, they support urban planning, building design, and other future endeavors. This study focuses on generating flood maps using synthetic aperture radar images from the Sentinel-1 (COPERNICUS/S1_GRD) satellite. Further study includes damage assessments in seven different sectors: urban land, agricultural land, forest land, barren land, range land, permanent water bodies, and unknown. This forecasts how much of the land in these 7 areas was affected by flooding. For the aforementioned land use and land cover classifications, the study proposes the best-fitting ensemble model, which is the aggregate of 3 image segmentation models that are Resnet34, InveptionV3, and VGG16. These three models are trained on the DeepGlobe dataset to give a mean Intersection over Union score of 75.84% and an F1 score of 0.76. A further proposed damage assessment technique is validated on a selected study area, i.e., village Vasagade from Kolhapur district of Maharashtra, which was severely affected in the year 2021s flood.
... Dealing with SITS data involves significant methodological and technical challenges. A crucial issue is a need for labeled data (Storie and Henry 2018), which are particularly difficult to obtain for satellite images. Moreover, even if the ground truth is available, the land cover class could change over time, especially when the time series is long (Guyet and Hervé 2016). ...
Article
Full-text available
Temporal sequences of satellite images constitute a highly valuable and abundant resource for analyzing regions of interest. However, the automatic acquisition of knowledge on a large scale is a challenging task due to different factors such as the lack of precise labeled data, the definition and variability of the terrain entities, or the inherent complexity of the images and their fusion. In this context, we present a fully unsupervised and general methodology to conduct spatio-temporal taxonomies of large regions from sequences of satellite images. Our approach relies on a combination of deep embeddings and time series clustering to capture the semantic properties of the ground and its evolution over time, providing a comprehensive understanding of the region of interest. The proposed method is enhanced by a novel procedure specifically devised to refine the embedding and exploit the underlying spatio-temporal patterns. We use this methodology to conduct an in-depth analysis of a 220 km22^2 region in northern Spain in different settings. The results provide a broad and intuitive perspective of the land where large areas are connected in a compact and well-structured manner, mainly based on climatic, phytological, and hydrological factors.
... This field grew in traditional statistics and artificial intelligence communities [70]. In this study, classical ML like SVM, RF, etc., were used to map lower order LULC classes to reduce the computation cost over the large geographical area with lower-resolution data because deep learning requires huge computation costs and highresolution data and is mostly suitable for higher order LULC classification [71][72][73]. ...
Article
Full-text available
Land use and land cover (LULC) classification plays a significant role in the analysis of climate change, evidence-based policies, and urban and regional planning. For example, updated and detailed information on land use in urban areas is highly needed to monitor and evaluate urban development plans. Machine learning (ML) algorithms, and particularly ensemble ML models support transferability and efficiency in mapping land uses. Generalization, model consistency, and efficiency are essential requirements for implementing such algorithms. The transfer-ensemble learning approach is increasingly used due to its efficiency. However, it is rarely investigated for mapping complex urban LULC in Global South cities, such as India. The main objective of this study is to assess the performance of machine and ensemble-transfer learning algorithms to map the LULC of two metropolitan cities of India using Landsat 5 TM, 2011, and DMSP-OLS nightlight, 2013. This study used classical ML algorithms, such as Support Vector Machine-Radial Basis Function (SVM-RBF), SVM-Linear, and Random Forest (RF). A total of 480 samples were collected to classify six LULC types. The samples were split into training and validation sets with a 65:35 ratio for the training, parameter tuning, and validation of the ML algorithms. The result shows that RF has the highest accuracy (94.43%) of individual models, as compared to SVM-RBF (85.07%) and SVM-Linear (91.99%). Overall, the ensemble model-4 produces the highest accuracy (94.84%) compared to other ensemble models for the Kolkata metropolitan area. In transfer learning, the pre-trained ensemble model-4 achieved the highest accuracy (80.75%) compared to other pre-trained ensemble models for Delhi. This study provides innovative guidelines for selecting a robust ML algorithm to map urban LULC at the metropolitan scale to support urban sustainability.
... Numerous research studies have shown that deep learning is able to obtain better scores as a general rule compared with more "traditional" machine learning algorithms. [91][92][93][94] However, in the case of this article, the pixel approach shows slightly higher scores. This may be due to the architecture of the neural network, which could be improved by integrating more convolution and hidden layers. ...
Conference Paper
Full-text available
Wetlands are one of the most important ecosystems in the world. Today, however, their fate is under serious threat, and their alarming decline highlights the urgent need to preserve these areas rich in biodiversity. The aim of this work is to spatially map and mapping the wetlands of the Crozon peninsula in Brittany France. The methodology is divided into two parts; the first part is to map the wetlands as a whole, while the second part is to map the wetlands using a adapted typology. Several databases were used to spatialize the wetlands: 12 Sentinel-2 images in L3A format, 23 Sentinel-1 VV and VH images and the RGE Alti (DTM at 1 metre resolution). The images were processed and stacked alone or in synergy. A Random Forest (RF) machine learning algorithm was then trained to predict wetlands in our study area using binary training data. The training data were obtained from a wetland inventory conducted in Brittany, distributed at the scale of a Sentinel-2 tile (30UUU). Post-processing was then carried out on the best result: binary morphological erosion and thresholding based on the DTM to remove outliers. We carried out two classifications, which we later merged. The classifications were carried out using a Pleiades time series (five dates) to achieve a very fine scale classification. A classification of 13 land cover classes with six different wetland types (mudflats, salt marshes, coastal lagoons, wet meadows, wet forest, swamps/bogs) was performed using three methods: pixel-by-pixel random forest, object-based random forest and Convolutional Neural Network (CNN). The best results obtained was for the pixel-based classification: kappa = 0.89, overall accuracy = 0.90, F1-score = 0.90.
... Machine learning (ML) techniques such as random forest (RF), neural networks (NN), support vector machine (SVM), maximum likelihood classifier (MLC), decision trees, and K-Nearest Neighbors (KNN) (Gislason et al., 2006;Svoboda et al, 2022;Kulkarni and Lowe, 2016;Storie and Henry, 2018;Alshari et al., 2023;Taati, A. et al. 2015;Norovsuren, B. et al., 2019;Friedl and Brodley, 1997;Srimani and Prasad, 2012;Upadhyay et al., 2016) have been used for land use/land cover mapping due to its high accuracy and efficiency. Moreover, NN and Artificial Neural Networks (ANN) are often used interchangeably on machine learning applications. ...
Article
Full-text available
The balance between food security and energy security is a national issue of extreme importance. A more stable supply of electricity could be achieved as solar farms expand but at the expense of losing some of the prime agricultural lands which endangers availability of sufficient agricultural produce. This study aims to use ANN-Cellular Automata (CA) via the Geographic Information System (GIS) platform and remote sensing (RS) data to assess the impact of cropland transition to solar farms and other land use/land cover (LULC). Several remotely sensed data were processed including MODIS land cover data (MCD12Q1), VIIRS nighttime lights (VNL v2.1), Advanced Himawari Imager Shortwave Radiation (AHI-SWR) product, and population density (LandScan) as inputs to the Cellular Automata-Artificial Neural Network (CA-ANN) model to simulate LULC changes in Tarlac Province, Philippines via the Modules for Land Use Change Evaluation (MOLUSCE) plugin in QGIS. For years 2019, 2023 and 2027 with 2015 as the base year, results showed an increasing trend for savannas and grassland with ΔLULC values of +11.4% to +15.1% and +0.2% to 3.5%, respectively. Meanwhile, a decreasing trend is observed for built-up/water, forest, and cropland with ΔLULC values of −3.0% to −6.3%, −8.5% to −21.1%, and −3.9% to −4.2%, respectively. Results also showed a conversion of a 100-ha area of croplands to solar farm from year 2019 to 2023 which translates to an estimated monetary loss from agricultural produce due to solar farm conversion amounting to Php 7,584,720.00 (~USD 138,000) which is equivalent to the total average annual income of about 67 families in Tarlac. Lastly, the simulated 2027 LULC map showed pixels with unrealistic conversions from solar farm (year 2023) to cropland (year 2027). To improve the model, it is recommended to add more spatial data to effectively capture factors that may contribute to the expansion of solar farms in the future. Moreover, high resolution LULC maps (vector maps if available) can be used instead of a course resolution satellite-derived raster data. Nonetheless, this study has demonstrated the use of RS, GIS and machine learning techniques to model cropland conversion to solar farms and other LULC classes. Results from this study can provide scientific data to policy makers, solar industry players and other relevant stakeholders in doing technoeconomic assessment of solar farm development and expansion considering its effect on energy security and food security towards national sustainable development.
... The result of their research showed better performance of the 2D CNN in overall accuracy over the random forest, ensemble of multilayer perceptron, and 1D-CNN methods. Storie and Henry, 2018 [19] investigated deep learning neural networks (FCN-8 and VGG-16 network) to analyze LULC types using Landsat 5 and Landsat 7. The results showed high average accuracy and quickly produced land use mapping. Later, Akhassan et al., 2020 [20] evaluated a deep learning framework for LULC mapping using multispectral satellite imagery. ...
Article
Full-text available
Using deep learning semantic segmentation for land use extraction is the most challenging problem in medium spatial resolution imagery. This is because of the deep convolution layer and multiple levels of deep steps of the baseline network, which can cause a degradation problem in small land use features. In this paper, a deep learning semantic segmentation algorithm which comprises an adjustment network architecture (LoopNet) and land use dataset is proposed for automatic land use classification using Landsat 8 imagery. The experimental results illustrate that deep learning semantic segmentation using the baseline network (SegNet, U-Net) outperforms pixel-based machine learning algorithms (MLE, SVM, RF) for land use classification. Furthermore, the LoopNet network, which comprises a convolutional loop and convolutional block, is superior to other baseline networks (SegNet, U-Net, PSPnet) and improvement networks (ResU-Net, DeeplabV3+, U-Net++), with 89.84% overall accuracy and good segmentation results. The evaluation of multispectral bands in the land use dataset demonstrates that Band 5 has good performance in terms of extraction accuracy, with 83.91% overall accuracy. Furthermore, the combination of different spectral bands (Band 1-Band 7) achieved the highest accuracy result (89.84%) compared to individual bands. These results indicate the effectiveness of LoopNet and multispectral bands for land use classification using Landsat 8 imagery.
... Semantic segmentation of satellite images using deep learning (DL) algorithms is becoming an important tool that remote sensing scientists use for land use land cover (LULC) map generation (Storie and Henry 2018;Henry et al. 2019;Alhassan et al. 2020;Liu et al. 2022). These DL techniques are relatively inexpensive, rapid to generate, and highly accurate. ...
Preprint
Full-text available
The availability of frequent and cost-free satellite images is in growing demand in the research world. Such satellite constellations as Landsat 8 and Sentinel-2 provide a massive amount of valuable data daily. However, the discrepancy in the sensors' characteristics of these satellites makes it senseless to use a segmentation model trained on either dataset and applied to another, which is why domain adaptation techniques have recently become an active research area in remote sensing. In this paper, an experiment of domain adaptation through style-transferring is conducted using the HRSemI2I model to narrow the sensor discrepancy between Landsat 8 and Sentinel-2. This paper's main contribution is analyzing the expediency of that approach by comparing the results of segmentation using domain-adapted images with those without adaptation. The HRSemI2I model, adjusted to work with 6-band imagery, shows significant intersection-over-union performance improvement for both mean and per class metrics. A second contribution is providing different schemes of generalization between two label schemes - NALCMS 2015 and CORINE. The first scheme is standardization through higher-level land cover classes, and the second is through harmonization validation in the field.
... While basic machine learning models do become progressively better at performing their specific functions as they take in new emergent data, they still need some human intervention. Deep learning algorithms in layers can build an "artificial neural network" (Figure 7) that is able to learn and make intelligent classification decisions on its own [102]. Figure 8 illustrates the differences between traditional machine learning and deep learning. ...
Article
Full-text available
We are currently living in the era of big data. The volume of collected or archived geospatial data for land use and land cover (LULC) mapping including remotely sensed satellite imagery and auxiliary geospatial datasets is increasing. Innovative machine learning, deep learning algorithms, and cutting-edge cloud computing have also recently been developed. While new opportunities are provided by these geospatial big data and advanced computer technologies for LULC mapping, challenges also emerge for LULC mapping from using these geospatial big data. This article summarizes the review studies and research progress in remote sensing, machine learning, deep learning, and geospatial big data for LULC mapping since 2015. We identified the opportunities, challenges, and future directions of using geospatial big data for LULC mapping. More research needs to be performed for improved LULC mapping at large scales.
... This, in turn, was used in a wide spectrum of computer vision applications [9,10]. The success of CNN-based segmentation algorithms led to their use in automated development of LULC maps, with very good results [11,12,13]. However, the process of annotating a large number of images needed for training CNN-based semantic segmentation models is a significant challenge. ...
Preprint
Full-text available
Deep learning has become one of remote sensing scientists' most efficient computer vision tools in recent years. However, the lack of training labels for the remote sensing datasets means that scientists need to solve the domain adaptation problem to narrow the discrepancy between satellite image datasets. As a result, image segmentation models that are then trained, could better generalize and use an existing set of labels instead of requiring new ones. This work proposes an unsupervised domain adaptation model that preserves semantic consistency and per-pixel quality for the images during the style-transferring phase. This paper's major contribution is proposing the improved architecture of the SemI2I model, which significantly boosts the proposed model's performance and makes it competitive with the state-of-the-art CyCADA model. A second contribution is testing the CyCADA model on the remote sensing multi-band datasets such as WorldView-2 and SPOT-6. The proposed model preserves semantic consistency and per-pixel quality for the images during the style-transferring phase. Thus, the semantic segmentation model, trained on the adapted images, shows substantial performance gain compared to the SemI2I model and reaches similar results as the state-of-the-art CyCADA model. The future development of the proposed method could include ecological domain transfer, {\em a priori} evaluation of dataset quality in terms of data distribution, or exploration of the inner architecture of the domain adaptation model.
... In the 18 papers dealing with mapping, the most commonly used method was Random Forests (a share of 50%), followed by Neural Networks and Support Vector Machines (both with a share of 16.67%). The mapping approaches focused on land use/land cover (Huang et al. 2018;Storie and Henry 2018;Yin et al. 2018;Pavri and Farrell 2020), land and vegetation cover (Evans and Cushman 2009;Samarkhanov et al. 2019), or land cover change (Keshtkar et al. 2017), by applying ML methods to satellite imagery. More focused applications dealt with vegetation communities (Helmer et al. 2008;Henderson et al. 2014;Mishra et al. 2020a). ...
Article
Full-text available
ContextArtificial Intelligence (AI) has rapidly developed over the past several decades. Several related AI approaches, such as Machine Learning (ML), have been applied to research on landscape patterns and ecological processes.Objectives Our goal was to review the methods of AI, particularly ML, used in studies related to landscape ecology and the main topics addressed. We aimed to assess the trend in the number of ML papers and the methods used therein, and provide a synopsis and prospectus of current use and future applications of ML in landscape ecology.Methods We conducted a systematic literature search and selected 125 papers for review. These were examined and scored according to multiple criteria regarding methods and topic. We applied quantitative statistical methods, including cluster analysis based on titles, abstracts, and keywords and a non-metric multidimensional scaling based on attributes assigned during the review. We used Random Forests machine learning to describe the differences between identified clusters in terms of the topics and methods they included.ResultsThe most frequent method found was Random Forests, but it is noteworthy to mention the increasing popularity of tools related to Deep Learning. The topics cover both ecologically oriented issues and the landscape-human interface. There has been a rapid increase in ML and AI methods in landscape ecology research, with Deep Learning and complex multi-step pipeline AI methods emerging in the last several years.Conclusions The rapid increase in the number of ML papers in landscape ecology research, and the range of methods employed in them, suggest explosive growth in application of these methods in landscape ecology. The increase of Deep Learning approaches in the most recent years suggest a major change in analytical paradigms and methodologies that we feel may transform the field and enable analyses of more complex pattern process relationships across vaster data sets than has been possible previously.
... Overfitting and the need for large training datasets [41][42][43] The high computational cost of training [75,76] [ 34,62,76,90] RNN Hybrid/ Semi-supervised Determine patterns and other significant features present in the dataset [15,16] Predict future developments [22] It remembers each piece of information through time [27] RNN is even used with convolutional layers to extend the effective pixel neighborhood [97] Training an RNN is a very difficult task [62] Gradient vanishing and exploding problems [98] [5] ...
Article
Full-text available
Agriculture land is playing a vital role in developing the economy of Indian states and contributes ~ 15% of India’s gross domestic product (GDP). Moreover, agriculture is a major source of livelihood by engaging two-third (~ 66%) of the nation’s population in various activities such as food supply, the raw material to the industries, internal and external trade. Therefore, the continuous monitoring and mapping of agricultural land are crucial for the sustainable life and development of the country. Most of the agriculture monitoring solutions are based on field observations or conventional strategies which are time-consuming and costlier. However, remote sensing delivers a cost-effective solution of acquiring information regarding the healthy or unhealthy vegetation in agricultural land with the help of a diverse range of advanced geospatial techniques such as classification, change detection, and pan-sharpening. In the present paper, we have performed a systematic survey with respect to recent advancements made in the classification algorithm, especially for agricultural land. These emerging methods incorporated in classifiers are machine learning and deep learning to enhance and detect the various features of vegetation parameters. It is expected that such studies will provide effective guidance to the researchers in better understanding the features, limitations, and specific importance of emerging classifiers in the Agriculture domain.
... Besides, Deep Neural Network (DNN) is also another powerful and efficient algorithm of AI. Nowadays, DNN has become popular and widely used to solve practical engineering problems and provide reliable results [37][38][39][40]. ...
Article
Full-text available
This study focuses on the use of deep neural network (DNN) to predict the soil friction angle, one of the crucial parameters in geotechnical design. Besides, particle swarm optimization (PSO) algorithm was used to improve the performance of DNN by selecting the best structural DNN parameters, namely, the optimal numbers of hidden layers and neurons in each hidden layer. For this aim, a database containing 245 laboratory tests collected from a project in Ho Chi Minh city, Vietnam, was used for the development of the proposed hybrid PSO-DNN model, including seven input factors (soil state, standard penetration test value, unit weight of soil, void ratio, thickness of soil layer, top elevation of soil layer, and bottom elevation of soil layer) and the friction angle was considered as the target. The data set was divided into three parts, namely, the training, validation, and testing sets for the construction, validation, and testing phases of the model. Various quality assessment criteria, namely, the coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE), were used to estimate the performance of PSO-DNN models. The PSO algorithm showed a remarkable ability to find out an optimal DNN architecture for the prediction process. The results showed that the PSO-DNN model using 10 hidden layers outperformed the DNN model, in which the average correlation improvement increased R2 by 1.83%, MAE by 5.94%, and RMSE by 8.58%. Besides, a global sensitivity analysis technique was used to detect the most important inputs, and it showed that, among the seven input variables, the elevation of top and bottom of soil played an important role in predicting the friction angle of soil.
... In many instances, when multiple land use classes were considered, the accuracy of the results was compromised, with values less than 0.5 being common [9]. Therefore, significant efforts were placed on developing algorithms that improve accuracy, such as [10], which obtained data for 16 classes with an accuracy of 88%, or [11], which obtained data for 10 classes with an accuracy of 98%. Increases in the accuracy of these methods are often not combined with increases in the number of classes, which would raise the utility of the results. ...
Article
Full-text available
The Willamette Valley, bounded to the west by the Coast Range and to the east by the Cascade Mountains, is the largest river valley completely confined to Oregon. The fertile valley soils combined with a temperate, marine climate create ideal agronomic conditions for seed production. Historically, seed cropping systems in the Willamette Valley have focused on the production of grass and forage seeds. In addition to growing over two-thirds of the nation’s cool-season grass seed, cropping systems in the Willamette Valley include a diverse rotation of over 250 commodities for forage, seed, food, and cover cropping applications. Tracking the sequence of crop rotations that are grown in the Willamette Valley is paramount to answering a broad spectrum of agronomic, environmental, and economical questions. Landsat imagery covering approximately 25,303 km2 were used to identify agricultural crops in production from 2004 to 2017. The agricultural crops were distinguished by classifying images primarily acquired by three platforms: Landsat 5 (2003–2013), Landsat 7 (2003–2017), and Landsat 8 (2013–2017). Before conducting maximum likelihood remote sensing classification, the images acquired by the Landsat 7 were pre-processed to reduce the impact of the scan line corrector failure. The corrected images were subsequently used to classify 35 different land-use classes and 137 unique two-year-long sequences of 57 classes of non-urban and non-forested land-use categories from 2004 through 2014. Our final data product uses new and previously published results to classify the western Oregon landscape into 61 different land use classes, including four majority-rule-over-time super-classes and 57 regular classes of annually disturbed agricultural crops (19 classes), perennial crops (20 classes), forests (13 classes), and urban developments (5 classes). These publicly available data can be used to inform and support environmental and agricultural land-use studies
... In recent years, deep learning (DL) has become the fastest-growing trend in big data analysis. DL models, especially the convolutional neural networks (CNNs), achieved significant improvement in RS image analysis tasks including scene classification [10][11][12], land use and land cover (LULC) classification [13,14], and urban object extraction [15][16][17]. The advantage of CNNs is that hierarchical deep features from a low-level to high-level can be automatically extracted via a common end-to-end learning process, instead of manually designing or handcrafting features [18]. ...
Article
Full-text available
Automated extraction of buildings from earth observation (EO) data has long been a fundamental but challenging research topic. Combining data from different modalities (e.g., high-resolution imagery (HRI) and light detection and ranging (LiDAR) data) has shown great potential in building extraction. Recent studies have examined the role that deep learning (DL) could play in both multimodal data fusion and urban object extraction. However, DL-based multimodal fusion networks may encounter the following limitations: (1) the individual modal and cross-modal features, which we consider both useful and important for final prediction, cannot be sufficiently learned and utilized and (2) the multimodal features are fused by a simple summation or concatenation, which appears ambiguous in selecting cross-modal complementary information. In this paper, we address these two limitations by proposing a hybrid attention-aware fusion network (HAFNet) for building extraction. It consists of RGB-specific, digital surface model (DSM)-specific, and cross-modal streams to sufficiently learn and utilize both individual modal and cross-modal features. Furthermore, an attention-aware multimodal fusion block (Att-MFBlock) was introduced to overcome the fusion problem by adaptively selecting and combining complementary features from each modality. Extensive experiments conducted on two publicly available datasets demonstrated the effectiveness of the proposed HAFNet for building extraction.
... Improving LC classification accuracy with the help of Machine Learning (ML) algorithms to meet users' needs has drawn considerable attention from the RS community [13][14][15][16]; however, ML methods provide inferior performance for the infrequent LC classes [17,18]. This is related to the fact that most of the ML classifiers try to decrease the overall error rate during the training phase, which leads to a higher level of accuracy for the main classes and lower level of accuracy for the infrequent classes [19][20][21]. ...
Article
Full-text available
Timely and accurate Land Cover (LC) information is required for various applications, such as climate change analysis and sustainable development. Although machine learning algorithms are most likely successful in LC mapping tasks, the class imbalance problem is known as a common challenge in this regard. This problem occurs during the training phase and reduces classification accuracy for infrequent and rare LC classes. To address this issue, this study proposes a new method by integrating random under-sampling of majority classes and an ensemble of Support Vector Machines, namely Random Under-sampling Ensemble of Support Vector Machines (RUESVMs). The performance of RUESVMs for LC classification was evaluated in Google Earth Engine (GEE) over two different case studies using Sentinel-2 time-series data and five well-known spectral indices, including the Normalized Difference Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), Soil-Adjusted Vegetation Index (SAVI), Normalized Difference Built-up Index (NDBI), and Normalized Difference Water Index (NDWI). The performance of RUESVMs was also compared with the traditional SVM and combination of SVM with three benchmark data balancing techniques namely the Random Over-Sampling (ROS), Random Under-Sampling (RUS), and Synthetic Minority Over-sampling Technique (SMOTE). It was observed that the proposed method considerably improved the accuracy of LC classification, especially for the minority classes. After adopting RUESVMs, the overall accuracy of the generated LC map increased by approximately 4.95 percentage points, and this amount for the geometric mean of producer's accuracies was almost 3.75 percentage points, in comparison to the most accurate data balancing method (i.e., SVM-SMOTE). Regarding the geometric mean of users' accuracies, RUESVMs also outperformed the SVM-SMOTE method with an average increase of 6.45 percentage points.
... In the 18 papers dealing with mapping, the most commonly used method was Random Forests (a share of 50%), followed by Neural Networks and Support Vector Machines (both with a share of 16.67%). The mapping approaches focused on land use/land cover (Huang et al. 2018;Storie and Henry 2018;Yin et al. 2018;Pavri and Farrell 2020), land and vegetation cover (Evans and Cushman 2009;Samarkhanov et al. 2019), or land cover change (Keshtkar et al. 2017), by applying ML methods to satellite imagery. More focused applications dealt with vegetation communities (Helmer et al. 2008;Henderson et al. 2014;Mishra et al. 2020a). ...
Presentation
Individual tree detection from remote sensing data is one of the main research areas concerning the exploitation of satellite and drone imagery for forestry and related ecological fields. While traditional techniques for tree detection for this kind of data exist, they are demonstrated to have many drawbacks related to pre-processing steps, data heterogeneity, spatial scale or raster resolution. In this study we take a look at implementing deep learning techniques for tree detection using drone-generated raster products. Deep learning is a subfield of machine learning that builds neural networks for high-level decision-making processes in data through model architectures. For this study, we trained and tested the Single Shot Detector algorithm using the ArcGIS API for Python and Jupyter Notebook. The drone data was acquired for an orchard, called Moara Domnească, in Ilfov county, Romania. Following the standard processing steps using Drone2Map, we obtained the ortophotomap and the Digital Surface Model (DSM). We then further derived slope and hillshade models for the same site. The first results show a poor performance (<40% detection) for ortophotomap, largely due to the similarity in spectral response between tree crowns and grass patches around the trees. DSM and slope raster data have similar detection values, while hillshade gave the best results, with a detection score of over 80%, we believe, mainly because the shape of trees is visible and distinct in comparison with the ground. Further testing and tuning is to be done on combinations of these raster products and other deep learning algorithms.
Article
Full-text available
Land-use maps containing crop rotational information are very important in land management and physical planning. Such maps are usually generated using multitemporal data. Although recent technology allows analysts to process multitemporal information effectively, the use of single date imagery for such purpose is more efficient. This study aimed to map detailed agricultural land-use with crop rotational information based on a single date Landsat 8 imagery and SRTM-derived terrain attributes. A landscape ecological approach assuming the influence of terrain characteristics on the existence of crop and land-use types was implemented in multisource classification using random decision forest (RDF) machine learning algorithm. The use of seven optical bands and five terrain attributes could provide a land-use map at 88.03% accuracy, compared to seven optical bands only that generate 82.45% accuracy. These results are also better than those of maximum likelihood. The most influential variables in the achieved accuracy are elevation and thermal band.
Chapter
Land use and land change (LULC) has several widely discussed research challenges. However, there is a significant research gap in detecting the direction of change of any remotely sensed object. Detection of such deviation has several environmental and social consequences. Further, it has direct correspondence for disaster management, early indication of natural phenomenon, and planning. In the past, different machine and deep learning techniques have been employed for LULC detection using different remote sensing paradigms such as aerial, multi-spetral, and hyper-spectral images. However, few works have studied the shape features of these land class changes to comprehend the deviation. Majority of them use complex shape features and employ multi-modal analysis. Therefore, the primary objective of this work is to quantify the direction of such change using multi-spectral images explicitly. Characteristics of intersecting lines are studied here to achieve this. Thus, the proposed technique can detect the deviation of any concave and convex remote sensing objects without holes using linear equations.
Article
Full-text available
Deep learning has become one of remote sensing scientists' most efficient computer vision tools in recent years. However, the lack of training labels for the remote sensing datasets means that scientists need to solve the domain adaptation problem to narrow the discrepancy between satellite image datasets. As a result, image segmentation models that are then trained, could better generalize and use an existing set of labels instead of requiring new ones. This work proposes an unsupervised domain adaptation model that preserves semantic consistency and per-pixel quality for the images during the style-transferring phase. This paper's major contribution is proposing the improved architecture of the SemI2I model, which significantly boosts the proposed model's performance and makes it competitive with the state-of-the-art CyCADA model. A second contribution is testing the CyCADA model on the remote sensing multi-band datasets such as WorldView-2 and SPOT-6. The proposed model preserves semantic consistency and per-pixel quality for the images during the style-transferring phase. Thus, the semantic segmentation model, trained on the adapted images, shows substantial performance gain compared to the SemI2I model and reaches similar results as the state-of-the-art CyCADA model. The future development of the proposed method could include ecological domain transfer, a priori evaluation of dataset quality in terms of data distribution, or exploration of the inner architecture of the domain adaptation model.
Thesis
Im Zuge des globalen Wandels gewinnt die Analyse des Zustands der Erdoberfläche und ihrer Veränderungen immer mehr an Bedeutung. Die Bereitstellung geeigneter Informationen hinsichtlich der Landbedeckung und Landnutzung erfolgt dabei basierend auf der Auswertung fernerkundungsgestützter Bilddaten. Für die Auswertung werden gegenwärtig überwiegend hybride semi-automatische Klassifizierungsmethoden genutzt. Die aktuellen Entwicklungen im Hinblick auf eine steigende Datenqualität und stetig zunehmende verfügbare Datenmengen lassen jedoch den Bedarf nach geeigneteren Verfahren entstehen, um den Ansprüchen und dem Potenzial dieser Daten gerecht werden zu können. Durch entscheidende Fortschritte im Hardwarebereich und der Prozessparallelisierung rücken automatisierte, auf künstlicher Intelligenz basierende Klassifizierungsmethoden zunehmend in den Fokus der Wissenschaft. Diese Arbeit hat das Ziel, das Potenzial von Deep Learning Anwendungen für die Klassifizierung von Landbedeckung und Landnutzung zu evaluieren. Dazu wird ein neuronales Netzerk implementiert und mit einem fernerkundungsgestützten Bilddatensatz für Deutschland trainiert und getestet. Die gewählte Methode zeigt großes Potenzial hinsichtlich einer automatisierten Klassifizierung von Landbedeckung mit klassenspezifischen f1-Werten bis zu 0,93.
Article
With Deep Learning (DL) outperforming previous Machine Learning (ML) techniques in classifying images, the remote sensing community has recently shown an increased interest in using these algorithms to classify Land Use and Land Cover (LULC) using multispectral and hyperspectral data. Land Use (LU) and Land Cover (LC) are two types of cartographic data that are used to develop smart cities and monitor the environment. LULC classification can benefit greatly from successfully applying remote sensing Image Classification (IC) using high spatial resolution data. The acquisition of spatiotemporal data for LULC classification has been made more accessible because of recent improvements in spatial analysis and Deep Learning (DL) technology. Considering the quality of Deep Neural Networks (DNN) in related Computer Vision (CV) tasks and the enormous volume of remotely sensed data accessible, DL methods appear to be particularly promising for modelling many remote sensing problems. However, there are several issues with ground-truth, resolution, and the nature of data that have a significant impact on categorization performance. We propose a Reversible Residual Network (RAVNet), a hybrid residual attention sensitive segmentation approach, to precisely categorize LULC in this study. The suggested network is based on the VNet model, which extracts relevant information by mixing low-level and high-level Feature Maps (FM). The attention-aware features change adaptively to the integration of residual modules. Our system was tested on the National Agriculture Imagery Program (NAIP) dataset, and the findings demonstrate that our architecture is competitive against other learning models.
Article
Full-text available
With the development of deep learning, semantic segmentation technology has gradually become the mainstream technical method in large-scale multi-temporal landcover classification. Large-scale and multi-temporal are the two significant characteristics of Landsat imagery. However, the mainstream single-temporal semantic segmentation network lacks the constraints and assistance of pre-temporal information, resulting in unstable results, poor generalization ability, and inconsistency with the actual situation in the multi-temporal classification results. In this paper, we propose a multi-temporal network that introduces pre-temporal information as prior constrained auxiliary knowledge. We propose an element-wise weighting block module to improve the fine-grainedness of feature optimization. We propose a chained deduced classification strategy to improve multi-temporal classification’s stability and generalization ability. We label the large-scale multi-temporal Landsat landcover classification dataset with an overall classification accuracy of over 90%. Through extensive experiments, compared with the mainstream semantic segmentation methods, our proposed multi-temporal network achieves state-of-the-art performance with good robustness and generalization ability.
Article
Analyzing satellite images and remote sensing (RS) data using artificial intelligence (AI) tools and data fusion strategies has recently opened new perspectives for environmental monitoring and assessment. This is mainly due to the advancement of machine learning (ML) and data mining approaches, which facilitate extracting meaningful information at a large scale from geo-referenced and heterogeneous sources. This paper presents the first review of AI-based methodologies and data fusion strategies used for environmental monitoring, to the best of the authors’ knowledge. The first part of the article discusses the main challenges of geographical image analysis. Thereafter, a well-designed taxonomy is introduced to overview the existing frameworks, which have been focused on: (i) detecting different environmental impacts, e.g. land cover land use (LULC) change, gully erosion susceptibility (GES), waterlogging susceptibility (WLS), and land salinity and infertility (LSI); (ii) analyzing AI models deployed for extracting the pertinent features from RS images in addition to data fusion techniques used for combining images and/or features from heterogeneous sources; (iii) describing existing publicly-shared and open-access datasets; (iv) highlighting most frequent evaluation metrics; and (v) describing the most significant applications of ML and data fusion for RS image analysis. This is followed by an overview of existing works and discussions highlighting some of the challenges, limitations and shortcomings. To provide the reader with insight into real-world applications, two case studies illustrate the use of AI for classifying LULC changes and monitoring the environmental impacts due to dams’ construction, where classification accuracies of 98.57% and 97.05% have been reached, respectively. Lastly, recommendations and future directions are drawn.
Chapter
Land cover (LC) refers to the ground’s surface cover, which can be plant, urban development, or water, among other things. Land cover detection, delineation, and mapping using ongoing advances in sensor advances have seen an immense measure of exceptionally fine spatial goal (Sentinel-2) distantly detected symbolism being gathered for like clockwork (every five days). The Geospatial Data Abstraction Library (GDAL) is a geospatial programming library that is open source. Working in Python with spatial images enables us to channel the data, fetch the spatial data, catch, circle, and charge the raster or vector datasets with a successful use of the computational power giving a greater degree on information investigation. In the first place, we gather two date changes of Sentinel inferred information to research and guide locales that have gone through the land cover advances that you are keen on checking and planning. After collecting, we clip the Sentinel images to our study area using Rasterio and Fiona. By using Normalized Difference Vegetation Index (NDVI), we create maps of two collected dates of Sentinel-2. Finally, by differencing the resulted NDVI outputs, we create a final land cover change map. This study provided an important addition to the field, land cover change detection through GDAL; it has a lot of potential in a variety of geospatial applications.
Article
The study aims to analyse the long-term impacts of mining activities in Jharia coalfield (JCF) on land-use (LU) patterns using transfer learning of the deep convolutional neural network (Deep CNN) model. A new database was prepared by extracting 10,000 image samples of 6 × 6 size for five LU types (barren land, built-up area, coal mining region, vegetation and waterbody) from Landsat data to train and validate the model. The satellite data from 1987 to 2021 at an interval of two years was used for change analysis. The study results revealed that the model offers 95 and 88% accuracy on the training and the validation dataset. The results indicate that barren land, coal mining region, and waterbody have been decreased from 237.30 sq. km. (=39.88%) to 171.25 sq. km (=28.78%), 118.77 sq. km. (=19.96%) to 68.73 sq. km (=11.55%), and 35.58 sq. km (=5.98%) to 18.68 sq. km (=3.14%) during 1987–2021, respectively. On the other hand, the built-up area and vegetation have been increased from 120.14 sq. km (=20.19%) to 233.02 sq. km (=39.16%) and 83.19 sq. km (=13.98%) to 103.36 sq. km (=17.37%) during 1987–2021. The time-series correlation results indicate that coal mining is the most sensitive LU type from 1987 to 2021, whereas barren land is least sensitive up to 2011, and thereafter vegetation is the least sensitive.
Chapter
The Coronavirus disease 2019 SARS-CoV-2 is a disease which causes fear to human lives that has taken thousands and hundreds of lives globally. The pandemic which has resulted in a global health emergency is currently a much sought-after research topic. The frequently mutating virus which has originated from Chiroptera and subsequently got transmitted to other mammals including humans. However, at the genomic level, it is yet to be unraveled what makes humans more prone to getting infected by the coronaviruses. Here, we have implemented a Machine Learning model known as K-means Clustering that uses the combination of different features to determine the risk of infection. In this research paper, the K-means clustering method is used since it is a good performer for Clustering analysis. The algorithm can group the sequences of the dataset into five clusters based on the Elbow plot and co-linearity of co-efficient. Using dimensional reduction technique PCA is used with a 3D visualization and a heat map to showcase the correlation efficiency between the mutated and original sequence considered.KeywordsCovid-19SARS-CoV-2GenomeK-means clusteringElbow plotPrincipal component analysis
Article
The ability to extract roads, detect buildings, and identify land cover types from satellite images is critical for sustainable development, agriculture, forestry, urban planning, and climate change research. Semantic segmentation with satellite images to extract vegetation covers and urban planning is essential for sustainable development and is a need for the hour. In this paper, Deep Unet, the modified version of Unet, is used for semantic segmentation with pre-processing of the image using FAAGKFCM and SLIC Superpixel to establish mapping for classifying different landfills based on satellite imagery. The research aims to train and test convolutional models for mapping land cover and testing the usability of land cover and identification of changes in land cover. Using mIoU and global accuracy as the evaluation metrics, the proposed model is compared with other methods, namely SegNet, UNet, DeepUNet. It is found that the proposed model outperforms other methods with mIoU of 89.51 and 90.6% global accuracy.
Article
The importance of timely and accurate information about the land resources and the natural resources increased rapidly. Due to the impact of urbanization, the we face hasty climatic change. To mitigate the urban heat island in the developed and developing cities, a very accurate land cover classification has to be developed. Through which we can identify the changes in build-up areas, water bodies and vegetation index. In this paper, a hybrid hot encoding VGG19 deep learning method has been proposed. And a transfer learning method has been used to transfer the training data trained by the RestNet50 method to the proposed HGVGG19 method. The satellite images and aerial images are collected from various sources and classified based on the features. And the image dataset has been pre-processed using the image augmentation technique. Through which the image has been resized and processed for training it with the proposed mode. The categorical data cannot be processed directly, so we use one hot encoding method to find the borders of the class. Then the data has been trained using VGG19 method. Then using the MLR classifier we classify the images and using decision tree the class prediction has been predicted. After testing the model an accuracy of 98.5% has been achieved. Using the proposed algorithm, the analysis has been made with the historical images of many regions. And eight different class values have been obtained and stored as the textual data. Using the data, the land cover changes and the prediction of the land cover has been obtained with an accuracy of 98.5%.
Thesis
Full-text available
With the growth of computational power, the deep learning algorithms have achieved remarkable results in several areas. Agriculture is one of the areas that are using these algorithms for the most varied domains. Therefore, this work presents a systematic literature review to consolidate the state-of-the-art about the use of deep learning applied to agricultural challenges. Papers published between January 2012 and April 2019 were considered. From the 819 papers found, 230 papers were classified. We evaluated the deep learning techniques used, crops covered, data sets used, deep learning and agriculture challenges, and among other important insights. The results have shown that deep learning is successfully used for several crops in agriculture. In the livestock branch, for example, most of the works achieved an accuracy above 95%. In total, 47.2% of the papers achieved an accuracy above 95%. Consequently, there is a lot of work to be done in the area of deep learning for agriculture. Our analysis are very important to support new research that seeks to apply deep learning in agriculture and highlight the research gaps.
Conference Paper
Full-text available
State-of-the-art models for semantic segmentation are based on adaptations of convolutional networks that had originally been designed for image classification. However, dense prediction problems such as semantic segmentation are structurally different from image classification. In this work, we develop a new convolutional network module that is specifically designed for dense prediction. The presented module uses dilated convolutions to systematically aggregate multi-scale contextual information without losing resolution. The architecture is based on the fact that dilated convolutions support exponential expansion of the receptive field without loss of resolution or coverage. We show that the presented context module increases the accuracy of state-of-the-art semantic segmentation systems. In addition, we examine the adaptation of image classification networks to dense prediction and show that simplifying the adapted network can increase accuracy.
Article
Full-text available
We present an analysis of three possible strategies for exploiting the power of existing convolutional neural networks (ConvNets) in different scenarios from the ones they were trained: full training, fine tuning, and using ConvNets as feature extractors. In many applications, especially including remote sensing, it is not feasible to fully design and train a new ConvNet, as this usually requires a considerable amount of labeled data and demands high computational costs. Therefore, it is important to understand how to obtain the best profit from existing ConvNets. We perform experiments with six popular ConvNets using three remote sensing datasets. We also compare ConvNets in each strategy with existing descriptors and with state-of-the-art baselines. Results point that fine tuning tends to be the best performing strategy. In fact, using the features from the fine-tuned ConvNet with linear SVM obtains the best results. We also achieved state-of-the-art results for the three datasets used.
Article
Full-text available
Pixel-level labelling tasks, such as semantic segmentation and depth estimation from single RGB image, play a central role in image understanding. Recent approaches have attempted to harness the capabilities of deep learning techniques for image recognition to tackle pixel-level labelling tasks. To solve this problem, we introduce a new form of convolutional neural network, called CRF-RNN, which expresses a Conditional Random Field (CRF) as a Recurrent Neural Network (RNN). Our short network can be plugged in as a part of a deep Convolutional Neural Network (CNN) to obtain an end-to-end system that has desirable properties of both CNNs and CRFs. Importantly, our system fully integrates CRF modelling with CNNs, making it possible to train the whole system end-to-end with the usual back-propagation algorithm. We apply this framework to the problem of semantic image segmentation, obtaining competitive results with the state-of-the-art without the need of introducing any post-processing method for object delineation.
Article
Central to the looming paradigm shift toward data-intensive science, machine-learning techniques are becoming increasingly important. In particular, deep learning has proven to be both a major breakthrough and an extremely powerful tool in many fields. Shall we embrace deep learning as the key to everything? Or should we resist a black-box solution? These are controversial issues within the remote-sensing community. In this article, we analyze the challenges of using deep learning for remote-sensing data analysis, review recent advances, and provide resources we hope will make deep learning in remote sensing seem ridiculously simple. More importantly, we encourage remote-sensing scientists to bring their expertise into deep learning and use it as an implicit general model to tackle unprecedented, large-scale, influential challenges, such as climate change and urbanization.
Conference Paper
State-of-the-art models for semantic segmentation are based on adaptations of convolutional networks that had originally been designed for image classification. However, dense prediction and image classification are structurally different. In this work, we develop a new convolutional network module that is specifically designed for dense prediction. The presented module uses dilated convolutions to systematically aggregate multi-scale contextual information without losing resolution. The architecture is based on the fact that dilated convolutions support exponential expansion of the receptive field without loss of resolution or coverage. We show that the presented context module increases the accuracy of state-of-the-art semantic segmentation systems. In addition, we examine the adaptation of image classification networks to dense prediction and show that simplifying the adapted network can increase accuracy.
Conference Paper
Rectified activation units (rectifiers) are essential for state-of-the-art neural networks. In this work, we study rectifier neural networks for image classification from two aspects. First, we propose a Parametric Rectified Linear Unit (PReLU) that generalizes the traditional rectified unit. PReLU improves model fitting with nearly zero extra computational cost and little overfitting risk. Second, we derive a robust initialization method that particularly considers the rectifier nonlinearities. This method enables us to train extremely deep rectified models directly from scratch and to investigate deeper or wider network architectures. Based on our PReLU networks (PReLU-nets), we achieve 4.94% top-5 test error on the ImageNet 2012 classification dataset. This is a 26% relative improvement over the ILSVRC 2014 winner (GoogLeNet, 6.66%). To our knowledge, our result is the first to surpass human-level performance (5.1%, Russakovsky et al.) on this visual recognition challenge.
Conference Paper
We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 dif- ferent classes. On the test data, we achieved top-1 and top-5 error rates of 37.5% and 17.0% which is considerably better than the previous state-of-the-art. The neural network, which has 60 million parameters and 650,000 neurons, consists of five convolutional layers, some of which are followed by max-pooling layers, and three fully-connected layers with a final 1000-way softmax. To make training faster, we used non-saturating neurons and a very efficient GPU implemen- tation of the convolution operation. To reduce overfitting in the fully-connected layers we employed a recently-developed regularization method called dropout that proved to be very effective. We also entered a variant of this model in the ILSVRC-2012 competition and achieved a winning top-5 test error rate of 15.3%, compared to 26.2% achieved by the second-best entry
Article
The recent emergence of deep learning for characterizing complex patterns in remote sensing imagery reveals its high potential to address some classic challenges in this domain, e.g. scene classification. Typical deep learning models require extremely large datasets with rich contents to train a multilayer structure in order to capture the essential features of scenes. Compared with the benchmark datasets used in popular deep learning frameworks, however, the volumes of available remote sensing datasets are particularly limited, which have restricted deep learning methods from achieving full performance gains. In order to address this fundamental problem, this article introduces a methodology to not only enhance the volume and completeness of training data for any remote sensing datasets, but also exploit the enhanced datasets to train a deep convolutional neural network that achieves state-of-the-art scene classification performance. Specifically, we propose to enhance any original dataset by applying three operations – flip, translation, and rotation to generate augmented data – and use the augmented dataset to train and obtain a more descriptive deep model. The proposed methodology is validated in three recently released remote sensing datasets, and confirmed as an effective technique that significantly contributes to potentially revolutionary changes in remote sensing scene classification, empowered by deep learning.
Article
Deep-learning (DL) algorithms, which learn the representative and discriminative features in a hierarchical manner from the data, have recently become a hotspot in the machine-learning area and have been introduced into the geoscience and remote sensing (RS) community for RS big data analysis. Considering the low-level features (e.g., spectral and texture) as the bottom level, the output feature representation from the top level of the network can be directly fed into a subsequent classifier for pixel-based classification. As a matter of fact, by carefully addressing the practical demands in RS applications and designing the input"output levels of the whole network, we have found that DL is actually everywhere in RS data analysis: from the traditional topics of image preprocessing, pixel-based classification, and target recognition, to the recent challenging tasks of high-level semantic feature extraction and RS scene understanding.
Article
Convolutional networks are powerful visual models that yield hierarchies of features. We show that convolutional networks by themselves, trained end-to-end, pixels-to-pixels, improve on the previous best result in semantic segmentation. Our key insight is to build "fully convolutional" networks that take input of arbitrary size and produce correspondingly-sized output with efficient inference and learning. We define and detail the space of fully convolutional networks, explain their application to spatially dense prediction tasks, and draw connections to prior models. We adapt contemporary classification networks (AlexNet, the VGG net, and GoogLeNet) into fully convolutional networks and transfer their learned representations by fine-tuning to the segmentation task. We then define a skip architecture that combines semantic information from a deep, coarse layer with appearance information from a shallow, fine layer to produce accurate and detailed segmentations. Our fully convolutional network achieves improved segmentation of PASCAL VOC (30% relative improvement to 67.2% mean IU on 2012), NYUDv2, SIFT Flow, and PASCAL-Context, while inference takes one tenth of a second for a typical image.
Article
State-of-the-art models for semantic segmentation are based on adaptations of convolutional networks that had originally been designed for image classification. However, dense prediction and image classification are structurally different. In this work, we develop a new convolutional network module that is specifically designed for dense prediction. The presented module uses dilated convolutions to systematically aggregate multi-scale contextual information without losing resolution. The architecture is based on the fact that dilated convolutions support exponential expansion of the receptive field without loss of resolution or coverage. We show that the presented context module increases the accuracy of state-of-the-art semantic segmentation systems. In addition, we examine the adaptation of image classification networks to dense prediction and show that simplifying the adapted network can increase accuracy.
Article
Rectified activation units (rectifiers) are essential for state-of-the-art neural networks. In this work, we study rectifier neural networks for image classification from two aspects. First, we propose a Parametric Rectified Linear Unit (PReLU) that generalizes the traditional rectified unit. PReLU improves model fitting with nearly zero extra computational cost and little overfitting risk. Second, we derive a robust initialization method that particularly considers the rectifier nonlinearities. This method enables us to train extremely deep rectified models directly from scratch and to investigate deeper or wider network architectures. Based on our PReLU networks (PReLU-nets), we achieve 4.94% top-5 test error on the ImageNet 2012 classification dataset. This is a 26% relative improvement over the ILSVRC 2014 winner (GoogLeNet, 6.66%). To our knowledge, our result is the first to surpass human-level performance (5.1%, Russakovsky et al.) on this visual recognition challenge.
Article
We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 dif-ferent classes. On the test data, we achieved top-1 and top-5 error rates of 37.5% and 17.0% which is considerably better than the previous state-of-the-art. The neural network, which has 60 million parameters and 650,000 neurons, consists of five convolutional layers, some of which are followed by max-pooling layers, and three fully-connected layers with a final 1000-way softmax. To make train-ing faster, we used non-saturating neurons and a very efficient GPU implemen-tation of the convolution operation. To reduce overfitting in the fully-connected layers we employed a recently-developed regularization method called "dropout" that proved to be very effective. We also entered a variant of this model in the ILSVRC-2012 competition and achieved a winning top-5 test error rate of 15.3%, compared to 26.2% achieved by the second-best entry.