ArticlePDF Available

Abstract and Figures

In the context of monitoring and assessment of water consumption in the agricultural sector, the objective of this study is to build an operational approach capable of detecting irrigation events at plot scale in a near real-time scenario using Sentinel-1 (S1) data. The proposed approach is a decision tree-based method relying on the change detection in the S1 backscattering coefficients at plot scale. First, the behavior of the S1 backscattering coefficients following irrigation events has been analyzed at plot scale over three study sites located in Montpellier (southeast France), Tarbes (southwest France), and Catalonia (northeast Spain). To eliminate the uncertainty between rainfall and irrigation, the S1 synthetic aperture radar (SAR) signal and the soil moisture estimations at grid scale (10 km × 10 km) have been used. Then, a tree-like approach has been constructed to detect irrigation events at each S1 date considering additional filters to reduce ambiguities due to vegetation development linked to the growth cycle of different crops types as well as the soil surface roughness. To enhance the detection of irrigation events, a filter using the normalized differential vegetation index (NDVI) obtained from Sentinel-2 optical images has been proposed. Over the three study sites, the proposed method was applied on all possible S1 acquisitions in ascending and descending modes. The results show that 84.8% of the irrigation events occurring over agricultural plots in Montpellier have been correctly detected using the proposed method. Over the Catalonian site, the use of the ascending and descending SAR acquisition modes shows that 90.2% of the non-irrigated plots encountered no detected irrigation events whereas 72.4% of the irrigated plots had one and more detected irrigation events. Results over Catalonia also show that the proposed method allows the discrimination between irrigated and non-irrigated plots with an overall accuracy of 85.9%. In Tarbes, the analysis shows that irrigation events could still be detected even in the presence of abundant rainfall events during the summer season where two and more irrigation events have been detected for 90% of the irrigated plots. The novelty of the proposed method resides in building an effective unsupervised tool for near real-time detection of irrigation events at plot scale independent of the studied geographical context.
Content may be subject to copyright.
A preview of the PDF is not available
... Remote sensing is a unique and valuable tool, capable of addressing the lack of large-scale precise information over irrigation practices, and overcoming the limitations of analyses based on in situ observations, which are often prone to inconsistencies and gaps in the information collected. Current results in the field of remote sensing for irrigation practices featured the creation of global or regional scale maps of irrigated areas [14], [15], [16], [17], [18], [19], irrigation timings [20], [21], [22], and quantification of irrigation amounts at variable resolutions [23], [24], [25], [26], [27], [28], [29], [30]. In particular, for studies oriented on the mapping of irrigated areas, remote sensing data are often coupled with machine learning (ML) models, proving to be successful with both supervised [14], [15] and unsupervised approaches [16]. ...
... Fawaz et al. [40] proved how ResNET outperforms traditional models for classification tasks in a comparison study performed using a large variety of multidisciplinary datasets. Moreover, in the field of irrigation mapping, Bazzi et al. [22] confirmed how deep learning models outperform traditional ML techniques. ...
... ResNET main architecture is used as in Wang et al. [83]. This model was adapted in this study to be run with multiple variables (multivariate model), through the use of a late fusion of parallel networks [22], [84], [85], [86], [87]. ...
Article
Full-text available
Maps of irrigation systems are of critical value for a better understanding of the human impact on the water cycle, while they also present a very useful tool at the administrative level to monitor changes and optimize irrigation practices. This study proposes a novel approach for classifying different irrigation systems at field level by using remotely sensed data at sub-field scale as inputs of different supervised Machine Learning (ML) models for time-series classification. The ML models were trained using ground-truth data from more than 300 fields collected during a field campaign in 2020 across an intensely cultivated region in Catalunya, Spain. Two hydrological variables retrieved from satellite data, actual evapotranspiration ( $ET_{a}$ ) and soil moisture ( $SM$ ), showed the best results when used for classification, especially when combined together, retrieving a final accuracy of $90.1 \pm 2.7\%$ . All the three ML models employed for the classification showed that they were able to distinguish different irrigation systems, regardless of the different crops present in each field. For all the different tests, the best performances were reached by ResNET, the only Deep Neural Network model among the three tested. The resulting method enables the creation of maps of irrigation systems at field level and for large areas, delivering detailed information on the status and evolution of irrigation practices.
... Where Sentinel-1 ensures continuous SAR imaging for six days of temporal resolution, the Sentinel-2 offers four days of revisit time for optical images. Several studies showed that the C-band (5.405 GHz) SAR time series from the Sentinel-1 (S1) satellite is efficient for mapping irrigated areas and for detecting irrigation frequency (irrigation episodes) [7,16,33]. Models built using the S1 C-band data rely on detecting the changes in the backscattering of the SAR coefficient between consecutive images (at a time difference of six days), such as the one developed by Bazzi et al. [32,33], Le Page et al. [34] and Ouaadi et al. [35], that is basically related to changes in the surface soil moisture (about 3 to 5 cm depth in the C-band). A sharp increase in the backscattering coefficients between two S1 images can be attributed to a sharp increase in surface soil moisture, caused either by an irrigation episode or by a rainfall event. ...
... Several studies showed that the C-band (5.405 GHz) SAR time series from the Sentinel-1 (S1) satellite is efficient for mapping irrigated areas and for detecting irrigation frequency (irrigation episodes) [7,16,33]. Models built using the S1 C-band data rely on detecting the changes in the backscattering of the SAR coefficient between consecutive images (at a time difference of six days), such as the one developed by Bazzi et al. [32,33], Le Page et al. [34] and Ouaadi et al. [35], that is basically related to changes in the surface soil moisture (about 3 to 5 cm depth in the C-band). A sharp increase in the backscattering coefficients between two S1 images can be attributed to a sharp increase in surface soil moisture, caused either by an irrigation episode or by a rainfall event. ...
... In a recent study, Bazzi et al. [33] proposed an algorithm called the IEDM (Irrigation Event Detection Model) to detect irrigation events at the plot scale. This algorithm mainly relies on the change detection in the S1 C-band backscattering coefficients where the increase in the SAR backscattering coefficient is mainly due to the increase in the soil moisture values caused either by a rainfall event or by an irrigation episode. ...
Article
Full-text available
Irrigation monitoring is of great importance in agricultural water management to guarantee better water use efficiency, especially under changing climatic conditions and water scarcity. This study presents a detailed assessment of the potential of the Sentinel-1 (S1) Synthetic Aperture Radar (SAR) data to detect irrigation events at the plot scale. The potential of the S1 data to detect the irrigation events was carried out using the Irrigation Event Detection Model (IEDM) over semi-arid and temperate oceanic climates in five study sites in south Europe and the Middle East. The IEDM is a decision tree model initially developed to detect irrigation events using the change detection algorithm applied to the S1 time series data. For each study site and at each agricultural plot, all available S1 images during the period of irrigation were used to construct an S1 time series and apply the IEDM. Different types of major summer irrigated crops were analyzed in this study, including Maize, Soybean, Sorghum and Potato, mainly with the sprinkler irrigation technique. The irrigation detection accuracy was evaluated using S1 images and the IEDM against the climatic condition of the studied area, the vegetation development (by means of the normalized difference vegetation index, NDVI) and the revisit time of the S1 sensor. The main results showed generally good overall accuracy for irrigation detection using the S1 data, reaching 67% for all studied sites together. This accuracy varied according to the climatic conditions of the studied area, with the highest accuracy for semi-arid areas and lowest for temperate areas. The analysis of the irrigation detection as a function of the crop type showed that the accuracy of irrigation detection decreases as the vegetation becomes well developed. The main findings demonstrated that the density of the available S1 images in the S1 time series over a given area affects the irrigation detection accuracy, especially for temperate areas. In temperate areas the irrigation detection accuracy decreased from 70% when 15 to 20 S1 images were available per month to reach less than 56% when less than 10 S1 images per month were available over the study sites.
... The objective here is to measure the spatial and temporal anomalies of the soil moisture values (RD p and TA p ) at a plot scale independent of the rainfall events. Removing the effect of rainfall in calculating RD p and TA p may help reduce the ambiguity between the soil moisture variation due to rainfall or irrigation [40,43] especially in regions that encounter frequent rainfalls in the summer irrigation season such as the Orléans study site. Thus, instead of integrating all the SSM estimations provided by the S 2 MP to calculate RD p and TA p , the estimation dates encountering rainfall episodes are excluded. ...
... In other words, when the average soil moisture of all the agricultural plots increases more than 5 vol.% between the time t i−1 at time t i (SSM t i − SSM t i−1 > 5 vol.%), the S 2 MP map at time t i is excluded from the calculation of both RD p and TA p . The global increase in the soil moisture at the basin scale could be evidence of an existing rainfall event that occurred two to three days before the S1 acquisition date (the date of the S 2 MP estimation) [40,43,44]. ...
... In the first step of the S 2 IM, selecting reference data depends on two irrigation metrics, one derived from the S1 time series and the other derived from the S2-NDVI time series. The first S1 metric relies on applying an unsupervised change detection algorithm called the irrigation event detection model (IEDM) proposed by Bazzi et al. [43] on S1 time series for both VV and VH polarizations. The IEDM provides, for each plot and at each S1 image in each polarization (VV and VH), an irrigation possibility weight which represents a proxyprobability of the presence or absence of irrigation as: 0 (no irrigation), 25 (low irrigation possibility), 50 (medium irrigation possibility) and 100 (high irrigation possibility). ...
Article
Full-text available
Accurate information about the irrigated surface is essential to help assess the impact of irrigation on water consumption, the hydrological cycle and regional climate. In this study, we compare recently developed operational and spatially transferrable classification models proposed for irrigation mapping. The first model suggests the use of spatio-temporal soil moisture indices derived from the Sentinel-1/2 soil moisture product (S2MP) at plot scale to map irrigated areas using the unsupervised K-means clustering algorithm (Dari model). The second model called the Sentinel-1/2 Irrigation mapping (S2IM) is a classification model based on the use the Sentinel-1 (S1) and Sentinel-2 (S2) time series data. Five study cases were examined including four studied years in a semi-oceanic area in north-central France (between 2017 and 2020) and one year (2020) in a Mediterranean context in south France. Main results showed that the soil-moisture based model using K-means clustering (Dari model) performs well for irrigation mapping but remains less accurate than the S2IM model. The overall accuracy of the Dari model ranged between 72.1% and 78.4% across the five study cases. The Dari model was found to be limited over humid conditions as it fails to correctly distinguish rain-fed plots from irrigated plots with an accuracy of the rain-fed class reaching 24.2% only. The S2IM showed the best accuracy in the five study cases with an overall accuracy ranging between 72.8% and 93.0%. However, for humid climatic conditions, the S2IM had an accuracy of the rain-fed class reaching 62.0%. The S2IM is thus superior in terms of accuracy but with higher complexity for application than the Dari model that remains simple yet effective for irrigation mapping.
... Compared to multispectral sensors, microwave observations have the advantage of being mostly insensitive to weather conditions. Various remotely sensed soil moisture products can effectively detect the irrigation timing at the field-and regional-scale (Bazzi et al., 2020;Lawston et al., 2017;Le Page et al., 2020;Zappa et al., 2021). The estimation of irrigation water amounts based on remotely sensed soil moisture has been carried out either by adapting the SM2RAIN algorithm (Brocca et al., 2014) or by quantifying the difference between satellite and modeled soil moisture (Zaussinger et al., 2019). ...
... Soil moisture products with sub-kilometric spatial resolution, such as those derived from Sentinel-1 (Bauer-Marschallinger et al., 2019;El Hajj et al., 2017), have the potential to provide accurate information about irrigation timing at (quasi) field-scale (Bazzi et al., 2020;Le Page et al., 2020;Zappa et al., 2021). Despite the adequate temporal resolution ensured by the Sentinel-1 mission (1.5-4 days over Europe), low detection accuracy could be obtained because of low irrigation rates. ...
Article
Full-text available
While ensuring food security worldwide, irrigation is altering the water cycle and generating numerous environmental side effects. As detailed knowledge about the timing and the amounts of water used for irrigation over large areas is still lacking, remotely sensed soil moisture has proved potential to fill this gap. However, the spatial resolution and revisit time of current satellite products represent a major limitation to accurately estimating irrigation. This work aims to systematically quantify their impact on the retrieved irrigation information, hence assessing the value of satellite soil moisture for estimating irrigation timing and water amounts. In a real-world experiment, we modeled soil moisture using actual irrigation and meteorological data, obtained from farmers and weather stations, respectively. Modeled soil moisture was compared against various remotely sensed products differing in terms of spatio-temporal resolution to test the hypothesis that high-resolution observations can disclose the irrigation signal from individual fields while coarse-scale satellite products cannot. Then, in a synthetic experiment, we systematically investigated the effect of soil moisture spatial and temporal resolution on the accuracy of irrigation estimates. The analysis was further elaborated by considering different irrigation scenarios and by adding realistic amounts of random errors in the soil moisture time series. We show that coarse-scale remotely sensed soil moisture products achieve higher correlations with rainfed simulations, while high-resolution satellite observations agree significantly better with irrigated simulations, suggesting that high-resolution satellite soil moisture can inform on field-scale (∼40 ha) irrigation. A thorough analysis of the synthetic dataset showed that satisfactory results, both in terms of detection (F-score > 0.8) and quantification (Pearson’s correlation > 0.8), are found for noise-free soil moisture observations either with a temporal sampling up to 3 days or if at least one-third of the pixel covers the irrigated field(s). However, irrigation water amounts are systematically underestimated for temporal samplings of more than one day, and decrease proportionally to the spatial resolution, i.e., coarsening the pixel size leads to larger irrigation underestimations. Although lower spatial and temporal resolutions decrease the detection and quantification accuracies (e.g., R between 0.6 and 1 depending on the irrigation rate and spatio-temporal resolution), random errors in the soil moisture time series have a stronger negative impact (Pearson R always smaller than 0.85). As expected, better performances are found for higher irrigation rates, i.e. when more water is supplied during an irrigation event. Despite the potentially large underestimations, our results suggest that high-resolution satellite soil moisture has the potential to track and quantify irrigation, especially over regions where large volumes of irrigation water are applied to the fields, and given that low errors affect the soil moisture observations.
... Remote sensing offers a powerful tool for irrigation monitoring at large scales. Several studies have shown the significant potential of remote sensing data to quantify both the extent of irrigation and irrigation timing using passive optical [6][7][8][9] and/or active radar remote sensing data [10][11][12][13][14]. While optical remote sensing provides the difference in the spectral signature between irrigated and rain-fed crops caused by higher levels of photosynthesis and biomass for irrigated plots, SAR (Synthetic Aperture Radar) data provides the wetness information of the soil (soil moisture), which varies between irrigated and rain-fed plots. ...
... In a recent study, Bazzi et al. [10] proposed a decision tree algorithm for irrigation events' detection called the IEDM (Irrigation Event Detection Model), which mainly relies on the change in the SAR backscattering signal at the plot scale between consecutive S1 C-band images at a 6-day revisit time. They separated the increase in the SAR signal due to an irrigation event from the increase in the SAR backscattered signal due to rainfall by using rainfall information derived from the grid-scale (10 km × 10 km) SAR backscattered signal [23]. ...
Article
Full-text available
Comprehensive knowledge about irrigation timing is crucial for water resource management. This paper presents a comparative analysis between C- and L-band Synthetic Aperture Radar (SAR) data for the detection of irrigation events. The analysis was performed using C-band time series data derived from the Sentinel-1 (S1) satellite and two L-band images from the PALSAR-2 (ALOS-2) sensor acquired over irrigated grassland plots in the Crau plain of southeast France. The S1 C-band time series was first analyzed as a function of rainfall and irrigation events. The backscattering coefficients in both the L and C bands were then compared to the time difference between the date of the acquired SAR image and the date of the last irrigation event occurring before the SAR acquisition (Δt). Sensitivity analysis was performed for 2 classes of the Normalized Difference Vegetation Index (NDVI ≤0.7 and NDVI >0.7). The main results showed that when the vegetation is moderately developed (NDVI ≤0.7), the C-band temporal variation remains sensitive to the soil moisture dynamics and the irrigation events could be detected. The C-VV signal decreases due to the drying out of the soil when the time difference between the S1 image and irrigation event increases. For well-developed vegetation cover (NDVI >0.7), the C-band sensitivity to irrigation events becomes dependent on the crop type. For well-developed Gramineae grass with longs stalks and seedheads, the C band shows no correlation with Δt due to the absence of the soil contribution in the backscattered signal, contrary to the legume grass type, where the C band shows a good correspondence between C-VV and Δt for NDVI > 0.7. In contrast, analysis of the L-band backscattering coefficient shows that the L band remains sensitive to the soil moisture regardless of the vegetation cover development and the vegetation characteristics, thus being more suitable for irrigation detection than the C band. The L-HH signal over Gramineae grass or legume grass types shows the same decreasing pattern with the increase in Δt, regardless of the NDVI-values, presenting a decrease in soil moisture with time and thus high sensitivity of the radar signal to soil parameters. Finally, the co-polarizations for both the C and L bands (L-HH and C-VV) tend to be more adequate for irrigation detection than the HV cross-polarization, as they show higher sensitivity to soil moisture values.
... Currently, the IoT for small orchards has several new and cheaper sensors that can be used during optimal irrigation scheduling [23][24][25][26]. Remotely sensed data with drones [27,28] and satellites can estimate crop or plant water status to plan plant irrigation at greater scales [29][30][31]. ...
Article
Full-text available
Water is getting scarce and irrigation practices should become more efficient. Mango orchards require great quantities of water, and policies in developing countries are substituting surface gravity irrigation by pressurized systems. A commercial orchard having mature 25-year-old trees and a 10-year-old HD high-density section were irrigated with micro sprinklers using 100% ETc (crop evapotranspiration) and reduced deficit irrigation treatments of 75% and 50% ETc. Water soil measurements were made with EC-5 probes at 10 and 35 cm in depth to study the effect of the different irrigation treatments. After the 2020 harvest, mature trees were trimmed without achieving pruning severity greater than 1.3. Canopy volume, mango size, fruit yield and water-use efficiency WUE were analyzed during 2020 and 2021. Sporadic storms produced sprinkler watering problems as weeds proliferated within trees. A controller with a fuzzy algorithm optimized orchard management and saved water in trees without decreasing yield and fruit size. It was found that one year after mature trees were trimmed by taking away the larger internal branch, more light penetrated the canopy, increasing yield by 60%; pruning in HD trees presented a yield increase of 5.37%. WUE (water-use efficiency) also increased with pruning and its value increased to 87.6 when the fuzzy controller and the 50% DI treatments were used in mature trees. This value was 260% greater than the one obtained in pruned trees without the controller. HD trees presented a lower WUE and yield per hectare than mature trees.
... Some of these studies have used supervised decision tree, random forest, and support vector machine (SVM) based classification approaches [23,46,60]. In contrast, others have used unsupervised decision tree-based classification [4], achieving a good overall performance of over 80% accuracy. Other studies have achieved comparable accuracy with a -means clustering algorithm [15,16] and a deep learning approach of a convolution neural network (CNN) architecture [3]. ...
... Another lingering issue in land process mapping is determining the conditions under which a model can be utilized in locations beyond where it was trained. Site-specific methods may not be easily transferable to other places or climes (Ozdogan et al., 2010;Bazzi et al., 2020), and the performance of transferred models can often only be assessed after full implementation in a novel setting (de Lima and Marfurt, 2020). Therefore, processes that yield insights about model transferability before training and inference offer benefits to researchers seeking to understand the maximum spatial applicability of their approaches. ...
Article
Full-text available
In presenting an irrigation detection methodology that leverages multiscale satellite imagery of vegetation abundance, this paper introduces a process to supplement limited ground-collected labels and ensure classifier applicability in an area of interest. Spatiotemporal analysis of MODIS 250 m enhanced vegetation index (EVI) timeseries characterizes native vegetation phenologies at regional scale to provide the basis for a continuous phenology map that guides supplementary label collection over irrigated and non-irrigated agriculture. Subsequently, validated dry season greening and senescence cycles observed in 10 m Sentinel-2 imagery are used to train a suite of classifiers for automated detection of potential smallholder irrigation. Strategies to improve model robustness are demonstrated, including a method of data augmentation that randomly shifts training samples; and an assessment of classifier types that produce the best performance in withheld target regions. The methodology is applied to detect smallholder irrigation in two states in the Ethiopian Highlands, Tigray and Amhara, where detection of irrigated smallholder farm plots is crucial for energy infrastructure planning. Results show that a transformer-based neural network architecture allows for the most robust prediction performance in withheld regions, followed closely by a CatBoost model. Over withheld ground-collection survey labels, the transformer-based model achieves 96.7% accuracy over non-irrigated samples and 95.9% accuracy over irrigated samples. Over a larger set of samples independently collected via the introduced method of label supplementation, non-irrigated and irrigated labels are predicted with 98.3 and 95.5% accuracy, respectively. The detection model is then deployed over Tigray and Amhara, revealing crop rotation patterns and year-over-year irrigated area change. Predictions suggest that irrigated area in these two states has decreased by approximately 40% from 2020 to 2021.
Article
Full-text available
Abstract: The feasibility of soil moisture retrieval from C-band Sentinel-1 data has been widely acknowledged, with pre-operational 1-km products currently available at regional and/or continental scale using the long-term (LTCD) or short-term change detection (STCD) methods. Both algorithms share the same assumptions of time-invariant roughness and vegetation, which can be questionable even for a short period of 4 Sentinel-1 acquisitions (18 – 36 days). An advanced change detection (ACD) method is proposed in this study for an improved soil moisture retrieval from Sentinel-1 data, including two main modifications with respect to the existing STCD methods: i) approximating the effect of temporal varying vegetation on the Sentinel-1 backscatter as a variation in the two-way attenuation, and ii) a temporal soil moisture constraint based on the coarse Soil Moisture Active Passive (SMAP) soil moisture product to partly remove the uncertainty caused by vegetation and/or roughness changes. The evaluation, based on time-series observations from 34 OzNet stations and ground samples collected during the Fifth Soil Moisture Active and Passive Experiment (SMAPEx-5) showed that the ACD improved the correlation coefficient (R), root mean square error (RMSE) and un biased RMSE (ubRMSE), achieving 0.66, 0.071 m3/m3 and 0.071 m3/m3 at the point scale, 0.77, 0.063 m3/m3 and 0.051 m3/m3 at 1-km scale, 0.80 , 0.055 m3/m3 and 0.050 m3/m3 at 3-km scale. The contribution of the two modifications was further investigated using 559 stations from 22 networks across the world, showing that: i) the two modifications can increase R by 0.08 - 0.13 and reduce the retrieval RMSE by 0.009 - 0.013 m3/m3 (10% - 15% relative), and ii) the retrieval over densely vegetated areas or areas with large temporal vegetation variation can benefit more from the proposed modifications. The ACD achieved stable performance for various Sentinel-1 orbits/passes and maintained a stable performance for retrieval windows up to 30 Sentinel-1 acquisitions, providing a promising alternative for achieving consistent soil moisture retrievals from Sentinel-1.
Article
Full-text available
The aim of this study is to estimate surface soil moisture at a spatial resolution of 500 m and a temporal resolution of at least 6 days, by combining remote sensing data from Sentinel-1 and optical data from Sentinel-2 and MODIS (Moderate-Resolution Imaging Spectroradiometer). The proposed methodology is based on the change detection technique, applied to a series of measurements over a three-year period (2015 to 2018). The algorithm described here as "Soil Moisture Estimations from the Synergy of Sentinel-1 and optical sensors (SMES)" proposes different options, allowing information from vegetation densities and seasonal conditions to be taken into account. The output from this algorithm is a moisture index ranging between 0 and 1, with 0 corresponding to the driest soils and 1 to the wettest soils. This methodology has been tested at different test sites (South of France, Central Tunisia, Western Benin and Southwestern Niger), characterized by a wide range of different climatic conditions. The resulting surface soil moisture estimations are compared with in situ measurements and already existing satellite-derived soil moisture ASCAT (Advanced SCATterometer) products. They are found to be well correlated, for the African regions in particular (RMSE below 6 vol.%). This outcome indicates that the proposed algorithm can be used with confidence to estimate the surface soil moisture of a wide range of climatically different sites.
Article
Full-text available
The ability of Synthetic Aperture Radar (SAR) Sentinel-1 data to detect the main wheat phenological phases was investigated in the Bekaa plain of Lebanon. Accordingly, the temporal variation of Sentinel-1 (S1) signal was analyzed as a function of the phenological phases’ dates observed in situ (germination; heading and soft dough), and harvesting. Results showed that S1 data, unlike the Normalized Difference Vegetation Index (NDVI) data, were able to estimate the dates of theses phenological phases due to significant variations in S1 temporal series at the dates of germination, heading, soft dough, and harvesting. Particularly, the ratio VV/VH at low incidence angle (32–34°) was able to detect the germination and harvesting dates. VV polarization at low incidence angle (32–34°) was able to detect the heading phase, while VH polarization at high incidence angle (43–45°) was better than that at low incidence angle (32–34°), in detecting the soft dough phase. An automated approach for main wheat phenological phases’ determination was then developed on the western part of the Bekaa plain. This approach modelled the S1 SAR temporal series by smoothing and fitting the temporal series with Gaussian functions (up to three Gaussians) allowing thus to automatically detect the main wheat phenological phases from the sum of these Gaussians. To test its robustness, the automated method was applied on the northern part of the Bekaa plain, in which winter wheat is harvested usually earlier because of the different weather conditions. The Root Mean Square Error (RMSE) of the estimation of the phenological phases’ dates was 2.9 days for germination, 5.5 days for heading, 5.1 days soft dough, 3.0 days for West Bekaa’s harvesting, and 4.5 days for North Bekaa’s harvesting. In addition, a slight underestimation was observed for germination and heading of West Bekaa (−0.2 and −1.1 days, respectively) while an overestimation was observed for soft dough of West Bekaa and harvesting for both West and North Bekaa (3.1, 0.6, and 3.6 days, respectively). These results are encouraging, and thus prove that S1 data are powerful as a tool for crop monitoring, to serve enhanced crop management and production handling.
Article
Full-text available
This paper presents a comparison between the Sentinel-1/Sentinel-2-derived soil moisture product at plot scale (S2MP) and the new Copernicus surface soil moisture (C-SSM) product at 1-km scale over a wide region in southern France. In this study, both products were first evaluated using in situ measurements obtained by the calibrated time delay reflectometer in field campaigns. The accuracy against the in situ measurements was defined by the correlation coefficient R, the root mean square difference (RMSD), and the bias and the unbiased root mean square difference (ubRMSD). Then, the soil moisture estimations from both SSM products were intercompared over one year (October 2016–October 2017). Both products show generally good agreement with in situ measurements. The results show that using in situ measurements collected over agricultural areas and grasslands, the accuracy of the C-SSM is good (RMSD = 6.0 vol%, ubRMSD = 6.0 vol%, and R = 0.48) but less accurate than the S2MP (RMSD = 4.0 vol%, ubRMSD = 3.9 vol%, and R = 0.77). The intercomparison between the two SSM products over one year shows that both products are highly correlated over agricultural areas that are mainly used for cereals (R value between 0.5 and 0.9 and RMSE between 4 and 6 vol%). Over areas containing forests and vineyards, the C-SSM values tend to overestimate the S2MP values (bias > 5 vol%). In the case of well-developed vegetation cover, the S2MP does not provide SSM estimations while C-SSM sometimes provides underestimated SSM values.
Article
Full-text available
Mapping irrigated plots is essential for better water resource management. Today, the free and open access Sentinel-1 (S1) and Sentinel-2 (S2) data with high revisit time offers a powerful tool for irrigation mapping at plot scale. Up to date, few studies have used S1 and S2 data to provide approaches for mapping irrigated plots. This study proposes a method to map irrigated plots using S1 SAR (synthetic aperture radar) time series. First, a dense temporal series of S1 backscattering coefficients were obtained at plot scale in VV (Vertical-Vertical) and VH (Vertical-Horizontal) polarizations over a study site located in Catalonia, Spain. In order to remove the ambiguity between rainfall and irrigation events, the S1 signal obtained at plot scale was used conjointly to S1 signal obtained at a grid scale (10 km × 10 km). Later, two mathematical transformations, including the principal component analysis (PCA) and the wavelet transformation (WT), were applied to the several SAR temporal series obtained in both VV and VH polarization. Irrigated areas were then classified using the principal component (PC) dimensions and the WT coefficients in two different random forest (RF) classifiers. Another classification approach using one dimensional convolutional neural network (CNN) was also performed on the obtained S1 temporal series. The results derived from the RF classifiers with S1 data show high overall accuracy using the PC values (90.7%) and the WT coefficients (89.1%). By applying the CNN approach on SAR data, a significant overall accuracy of 94.1% was obtained. The potential of optical images to map irrigated areas by the mean of a normalized differential vegetation index (NDVI) temporal series was also tested in this study in both the RF and the CNN approaches. The overall accuracy obtained using the NDVI in RF classifier reached 89.5% while that in the CNN reached 91.6%. The combined use of optical and radar data slightly enhanced the classification in the RF classifier but did not significantly change the accuracy obtained in the CNN approach using S1 data.Remote Sensing
Article
Full-text available
Coastal land cover classification is a significant yet challenging task in remote sensing because of the complex and fragmented nature of coastal landscapes. However, availability of multitemporal and multisensor remote sensing data provides opportunities to improve classification accuracy. Meanwhile, rapid development of deep learning has achieved astonishing results in computer vision tasks and has also been a popular topic in the field of remote sensing. Nevertheless, designing an effective and concise deep learning model for coastal land cover classification remains problematic. To tackle this issue, we propose a multibranch convolutional neural network (MBCNN) for the fusion of multitemporal and multisensor Sentinel data to improve coastal land cover classification accuracy. The proposed model leverages a series of deformable convolutional neural networks to extract representative features from a single-source dataset. Extracted features are aggregated through an adaptive feature fusion module to predict final land cover categories. Experimental results indicate that the proposed MBCNN shows good performance, with an overall accuracy of 93.78% and a Kappa coefficient of 0.9297. Inclusion of multitemporal data improves accuracy by an average of 6.85%, while multisensor data contributes to 3.24% of accuracy increase. Additionally, the featured fusion module in this study also increases accuracy by about 2% when compared with the feature-stacking method. Results demonstrate that the proposed method can effectively mine and fuse multitemporal and multisource Sentinel data, which improves coastal land cover classification accuracy.
Article
Full-text available
This study proposes an effective method to map rice crops using the Sentinel-1 SAR (Synthetic Aperture Radar) time series over the Camargue region, Southern France. First, the temporal behavior of the SAR backscattering coefficient over 832 plots containing different crop types was analyzed. Through this analysis, the rice cultivation was identified using metrics derived from the Gaussian profile of the VV/VH time series (3 metrics), the variance of the VV/VH time series (one metric), and the slope of the linear regression of the VH time series (one metric). Using the derived metrics, rice plots were mapped through two different approaches: decision tree and Random Forest (RF). To validate the accuracy of each approach, the classified rice map was compared to the available national data. Similar high overall accuracy was obtained using both approaches. The overall accuracy obtained using a simple decision tree reached 96.3%, whereas an overall accuracy of 96.6% was obtained using the RF classifier. The approach, therefore, provides a simple yet precise and powerful tool to map paddy rice areas.
Article
Full-text available
Latest remote sensing sensors are capable of acquiring high spatial and spectral Satellite Image Time Series (SITS) of the world. These image series are a key component of classification systems that aim at obtaining up-to-date and accurate land cover maps of the Earth’s surfaces. More specifically, current SITS combine high temporal, spectral and spatial resolutions, which makes it possible to closely monitor vegetation dynamics. Although traditional classification algorithms, such as Random Forest (RF), have been successfully applied to create land cover maps from SITS, these algorithms do not make the most of the temporal domain. This paper proposes a comprehensive study of Temporal Convolutional Neural Networks (TempCNNs), a deep learning approach which applies convolutions in the temporal dimension in order to automatically learn temporal (and spectral) features. The goal of this paper is to quantitatively and qualitatively evaluate the contribution of TempCNNs for SITS classification, as compared to RF and Recurrent Neural Networks (RNNs) —a standard deep learning approach that is particularly suited to temporal data. We carry out experiments on Formosat-2 scene with 46 images and one million labelled time series. The experimental results show that TempCNNs are more accurate than the current state of the art for SITS classification. We provide some general guidelines on the network architecture, common regularization mechanisms, and hyper-parameter values such as batch size; we also draw out some differences with standard results in computer vision (e.g., about pooling layers). Finally, we assess the visual quality of the land cover maps produced by TempCNNs.
Article
Full-text available
The objective of this paper is to present an analysis of Sentinel-1 derived surface soil moisture maps (S1-SSM) produced with high spatial resolution (at plot scale) and a revisit time of six days for the Occitanie region located in the South of France as a function of precipitation data, in order to investigate the potential of S1-SSM maps for detecting heavy rainfalls. First, the correlation between S1-SSM maps and rainfall maps provided by the Global Precipitation Mission (GPM) was investigated. Then, we analyzed the effect of the S1-SSM temporal resolution on detecting heavy rainfall events and the impact of these events on S1-SSM values as a function of the number of days that separated the heavy rainfall and the S1 acquisition date (cumulative rainfall more than 60 mm in 24 hours or 80 mm in 48 hours). The results showed that the six-day temporal resolution of the S1-SSM map doesn’t always permit the detection of an extreme rainfall event, because confusion will appear between high S1-SSM values due to extreme rainfall events occurring six days before the acquisition of S1-SSM, and high S1-SSM values due to light rain a few hours before the acquisition of Sentinel-1 images. Moreover, the monitoring of extreme rain events using only soil moisture maps remains difficult, since many environmental parameters could affect the value of SSM, and synthetic aperture radar (SAR) doesn’t allow the estimation of very high soil moistures (higher than 35 vol.%).
Article
This letter proposes a deep learning model to deal with the spatial transfer challenge for the mapping of irrigated areas through the analysis of Sentinel-1 data. First, a convolutional neural network (CNN) model called "Teacher Model" is trained on a source geographical area characterized by a huge volume of samples. Then, this model is transferred from the source area to a target area characterized by a limited number of samples. The transfer learning framework is based on a distill and refine strategy in which the teacher model is firstly distilled into a student model and, successively, refined by data samples coming from the target geographical area. The proposed strategy is compared to different approaches including a random forest (RF) classifier trained on the target dataset, a CNN trained on the source dataset and directly applied on the target area as well as several CNN classifiers trained on the target dataset. The evaluation of the performed transfer strategy shows that the "distill and refine" framework obtains the best performance compared to other competing approaches. The obtained findings represent a first step towards the understanding of the spatial transferability of deep learning models in the Earth Observation domain.