Mustafa Turker’s research while affiliated with Hacettepe University and other places

Timely and accurate crop mapping is crucial for yield prediction, food security assessment and agricultural management. Convolutional neural networks (CNNs) have become powerful state-of-the-art methods in many fields, including crop type detection from satellite imagery. However, existing CNNs generally have large number of layers and filters that increase the computational cost and the number of parameters to be learned, which may not be convenient for the processing of time-series images. To that end, we propose a light CNN model in combination with parcel-based image analysis for crop classification from time-series images. The model was applied on two areas (Manisa and Kırklareli) in Türkiye using Sentinel-2 data. Classification results based on all bands of the time-series data had overall accuracies (OA) of 89.3% and 88.3%, respectively for Manisa and Kırklareli. The results based on the optimal bands selected through the Support Vector Machine–Recursive Feature Elimination (SVM-RFE) method had OA of 86.6% and 86.5%, respectively. The proposed model outperformed the VGG-16, ResNet-50, and U-Net models used for comparison. For Manisa and Kırklareli respectively, VGG-16 achieved OA of 86.0% and 86.5%, ResNet-50 achieved OA of 84.1% and 84.8%, and U-Net achieved OA of 82.2% and 81.9% based on all bands. Based on the optimal bands, VGG-16 achieved OA of 84.2% and 84.7%, ResNet-50 achieved OA of 82.4% and 83.1%, and U-Net achieved OA of 80.5% and 80.2%. The results suggest that the proposed model is promising for accurate and cost-effective crop classification from Sentinel-2 time-series imagery.

Tree detection from space imagery is important in the agriculture and forestry industries. However, very high spatial resolution satellite imagery represents fine details on the ground making the object detection task more challenging. Most of the existing tree detection methods use multispectral bands, including an infrared (IR) band, which provides distinct information about vegetation areas making the detection task more manageable. However, the sheer amount of optical data employed as input to information extraction procedures contains only three bands, and IR bands may not always be available. Thus, this study presents automatic tree detection by only using the Red, Green, and Blue (RGB) bands of very high spatial resolution satellite images through deep learning. The proposed method was built on top of a U-Net architecture whose ability to detect different types of trees is explored. The U-Net architecture was trained using WorldView-3 RGB images. In addition to RGB bands, vegetation indices were computed and used as additional bands to investigate their effects on the results. In this respect, six models were generated, and each model was trained and tested individually. The models used include (RGB, RGB + VARI, RGB + GLI, RGB + GRVI, RGB + all indices, and only vegetation indices). Four accuracy assessment equations were calculated for each model, and the results were compared.

An accurate detection of agricultural fields is often needed for agricultural-related applications, such as subsidies monitoring, field-based crop yield estimation and agricultural statistics extraction. High-resolution space images have become the fundamental source to extract agricultural field boundaries. Manual boundary delineation is not practical. In this study, we present an approach to detect agricultural fields from satellite images on the basis of agricultural field blocks. An agricultural field block consists of one or more fields that are owned by the farmers. The approach combines the Otsu’s thresholding algorithm and marker-controlled watershed (MCW)-based segmentation. First, the well-separated field segments within a field block being considered are detected through recursive processing of the Otsu’s thresholding algorithm. Then, these distinct field segments are used to generate a marker image, and further extraction of individual fields is carried out through a marker-controlled watershed (MCW)-based segmentation. The approach was tested using 10-m resolution Satellite Pour l’Observation de la Terre (SPOT)-5 multi-spectral (XS) image, 4-m resolution IKONOS XS image, 2.40-m resolution QuickBird XS image, and 0.60-m resolution QuickBird pan-sharpened (PS) image. The results were evaluated using the reference field boundary dataset. The achieved overall accuracies were 89.7, 83.2, 81.0, and 77.4% for the IKONOS XS, QuickBird XS, SPOT-5 XS, and QuickBird PS images, respectively. The results are promising and indicate that the approach can be used for the extraction of agricultural fields from space imagery.

Agricultural field boundary information is important and often required for the geosciences and the agricultural sector. In this paper, a novel method is developed to extract sub-boundaries within the permanent boundaries of agricultural land parcels from high-resolution optical satellite imagery using an improved cluster-based snake model. The method takes the advantage of the results of an automatic fuzzy c-means (FCM) clustering and edge detection to compute external forces for an improved gradient vector flow (GVF) snake model. The GVF snake algorithm is improved by using an automatic seeding model based on clustering results and image moment functions. To seed the improved GVF algorithm, an ellipse is automatically delineated for each cluster within agricultural parcel by utilizing image moment functions (in particular silhouette moments). The GVF snake model is then implemented for each seed, one seed at a time. Active contours tend to have curve shapes rather than straight lines due to their structure that consists of several connected nodes within each contour. Therefore, the final accurate results are obtained after performing a three-stage line simplification operation. The experiments of the method were conducted on 20 test fields in a study area located near to the town of Karacabey, Turkey, using the 4-m resolution IKONOS multispectral (xs) image, the 2.44-m resolution QuickBird xs image, and the 0.61-m resolution QuickBird pan-sharpened (PS) image. Experimental results demonstrate that using both the clustering and edge detection results as external forces for the improved GVF snake model increases the accuracy of the results. In addition, the developed method showed a fairly good performance in extracting sub-boundaries for the fields comprising crops with an inherent high inner heterogeneity, such as rice and corn. The method can potentially be applied in the extraction of within-field sub-boundaries from high-resolution satellite imagery in agricultural areas.

This paper presents an integrated approach for the automatic extraction of rectangular- and circular-shape buildings from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping. The building patches are detected from the image using the binary SVM classification. The generated normalized digital surface model (nDSM) and the normalized difference vegetation index (NDVI) are incorporated in the classification process as additional bands. After detecting the building patches, the building boundaries are extracted through sequential processing of edge detection, Hough transformation and perceptual grouping. Those areas that are classified as building are masked and further processing operations are performed on the masked areas only. The edges of the buildings are detected through an edge detection algorithm that generates a binary edge image of the building patches. These edges are then converted into vector form through Hough transform and the buildings are constructed by means of perceptual grouping. To validate the developed method, experiments were conducted on pan-sharpened and panchromatic Ikonos imagery, covering the selected test areas in Batikent district of Ankara, Turkey. For the test areas that contain industrial buildings, the average building detection percentage (BDP) and quality percentage (QP) values were computed to be 93.45% and 79.51%, respectively. For the test areas that contain residential rectangular-shape buildings, the average BDP and QP values were computed to be 95.34% and 79.05%, respectively. For the test areas that contain residential circular-shape buildings, the average BDP and QP values were found to be 78.74% and 66.81%, respectively.

An integrated approach is presented for the automatic extraction of photorealistic facade textures from single street-level building images. The initial facade texture is extracted using Watershed segmentation. The seed pixels (markers) to trigger the segmentation are located automatically both for the foreground (facade) and the background regions, and the segmentation is carried out repetitively until the facade texture is extracted. The extracted facade image is geometrically rectified using a developed automatic technique based on Hough transformation and interest point detection. The occluded areas on facade textures are restored by employing an image matching-based procedure. The approach was tested on two different datasets captured from the residential areas of Ankara, the capital of Turkey. The datasets contain a total of 40 building facade images that were taken from the street-level. The results indicate that the facade textures are extracted adequately. For facade image extraction, an average quantitative accuracy of 83% was achieved. For rectification, 24 out of 40 buildings provided the positional error under 10 pixels at 95% confidence level. The subjective assessment of facade restoration yielded the mean rating value of 2.46 for the datasets used, in which the rating values are ranked between 1 for "Excellent" and 6 for "Unusable".

This study aims to find building patches from pan-sharpened IKONOS imagery using two-class support vector machines (SVM) classification. In addition to original bands of the image, the normalized digital surface model, normalized difference vegetation index, and several texture measures (mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation) are also used in the classification. The study illustrates the performance of the binary SVM classification in building detection from IKONOS imagery. Moreover, the effect of additional bands in building detection is examined. The approach was tested in three test sites that are located in the Batikent district of Ankara, Turkey. The SVM classification provided quite accurate results with the building detection percentage (BDP) values in the range 81.27-96.26% and the quality percentage (QP) values in the range 41.01-74.83%. It was found that the usage of additional bands in SVM classification had a significant effect in building detection accuracy. When compared to results obtained using solely the original bands, the additional bands increased the accuracy up to 10.44% and 8.45% for BDP and QP, respectively.

This study presents an approach for the automatic extraction of dynamic sub-boundaries within existing agricultural fields from remote sensing imagery using perceptual grouping. We define sub-boundaries as boundaries, where a change in crop type takes a place within the fixed geometry of an agricultural field. To perform field-based processing and analysis operations, the field boundary data and satellite imagery are integrated. The edge pixels are detected using the Canny edge detector. The edge pixels are then analyzed to find the connected edge chains and from these chains the line segments are detected using the graph-based vectorization method. The spurious line segments are eliminated through a line simplification process. The perceptual grouping of the line segments is employed for detecting sub-boundaries and constructing sub-fields within the fixed geometry of agricultural fields. Our strategy for perceptual grouping involves the Gestalt laws of proximity, continuation, symmetry and closure. The processing and analysis operations are carried out on field-by-field basis. For each field, the geometries of sub-boundaries are determined through analyzing the line segments that fall within the field and the extracted sub-boundaries are integrated with the fixed geometry of the field.

We propose a new approach for building detection using high-resolution satellite imagery based on an adaptive fuzzy-genetic algorithm. This novel approach improves object detection accuracy by reducing the premature convergence problem encountered when using genetic algorithms. We integrate the fundamental image processing operators with genetic algorithm concepts such as population, chromosome, gene, crossover and mutation. To initiate the approach, training samples are selected that represent the specified two feature classes, in this case “building” and “non-building”. The image processing operations are carried out on a chromosome-by-chromosome basis to reveal the attribute planes. These planes are then reduced to one hyperplane that is optimal for discriminating between the specified feature classes. For each chromosome, the fitness values are calculated through the analysis of detection and mis-detection rates. This analysis is followed by genetic algorithm operations such as selection, crossover and mutation. At the end of each generation cycle, the adaptive-fuzzy module determines the new (adjusted) probabilities of crossover and mutation. This evolutionary process repeats until a specified number of generations has been reached. To enhance the detected building patches, morphological image processing operations are applied. The approach was tested on ten different test scenes of the Batikent district of the city of Ankara, Turkey using 1 m resolution pan-sharpened IKONOS imagery. The kappa statistics computed for the proposed adaptive fuzzy-genetic algorithm approach were between 0.55 and 0.88. The extraction performance of the algorithm was better for urban and suburban buildings than for buildings in rural test scenes.

This article presents an approach for automatic building database updating from high-resolution space imagery based on the support vector machine (SVM) classification and building models. The developed approach relies on an idea that the buildings are similar in shape within an urban block or a neighbourhood unit. First, the building patches are detected through classification of the pan-sharpened high-resolution space imagery along with the normalized digital surface model (nDSM) and the normalized difference vegetation index (NDVI) using the binary SVM classifier. Then, the buildings that exist in the vector database but not seen in the image are detected through the analyses of the detected building patches. The buildings, which were constructed after the compilation date of the existing vector database, are extracted through the proposed model-based approach that utilizes the existing building database as a building model library. The approach was implemented in selected urban areas of the Batikent district of Ankara, the capital city of Turkey, using the IKONOS images and the existing building database. The results validated the success of the developed approach, with the building extraction accuracy computed to be higher than 80%.