Sentinel-2 Time Series Research Articles

Automatic land cover classification with SAR imagery and Machine learning using Google Earth Engine

Land cover is the most critical information required for land management and planning because human interference on land can be easily detected through it. However, mapping land cover utilizing optical remote sensing is not easy due to the acute shortage of cloud-free images. Google Earth Engine (GEE) is an efficient and effective tool for huge land cover analysis by providing access to large volumes of imagery available within a few days after acquisition in one consolidated system. This article demonstrates the use of Sentinel-1 datasets to create a land cover map of Pusad, Maharashtra using the GEE platform. Sentinel-1 provides Synthetic Aperture Radar (SAR) datasets that have a temporally dense and high spatial resolution, which is renowned for its cloud penetration characteristics and round-the-year observations irrespective of the weather. VV and VH polarization sentinel-1 time series data were automatically classified using a support vector machine (SVM) and Random Forest (RF) machine learning algorithms. Overall accuracies (OA), ranging from 82.3% to 90%, were obtained depending on polarization and methodology used. RF algorithm with VV polarization dataset stands better in comparison to SVM achieving OA of 90% and Kappa coefficient of 0.86. The highest user accuracy was obtained for the water class with both classifiers.

Contribution of time-series data cubes to classify urban vegetation types by remote sensing

Mapping urban vegetation types is important for urban planning and assessing environmental justice. Nowadays, despite data cubes projects are providing Analysis Ready Data to facilitate time-series analysis, we did not found studies employing these data for improving urban vegetation mapping. By relying solely on open data and software, this work proposes and evaluates the integration of time-series data cubes in a hybrid image classification method to map the intra-urban space, differentiating Tree cover and Herb-shrub. The urban area of Goiânia, Goiás, Brazil, is the study area. The hybrid method combined object-based classification of a pan-sharpened CBERS-4A WPM image (spatial resolution of 2 m) with the pixel-based classification of Sentinel-2 MSI time-series data cubes (10 m). Both approaches used the Random Forest algorithm. Objects from the CBERS-4A segmentation composed the spatial unit of analysis and the class assignment depended on the Sentinel-2 time-series urban land cover probabilities. Based on both Maps probabilities, Shannon entropy was calculated to attribute the final urban land cover to the objects. Urban land cover probabilities presented similar spatial distribution patterns for both classification approaches. Regarding the thematic maps, the Herb-shrub cover area was 35% higher in Sentinel-2 time-series classification than in GEOBIA classification, but Tree cover was 21% lower. In general, 75% of the study area was equally classified by the initial approaches. However, for 9% of the remaining area, the hybrid classification improved vegetation classes accuracies by 35%, contributing to the vegetation covers identification. Thus, this study contributes to methodological procedures for urban land cover study and demonstrates that hybrid maps based on open data are effective to reduce classification mistakes, allowing more accurate monitoring, planning, and designing of different urban vegetation types. Future research efforts should focus on scale compatibility between data of different spatial resolutions and expand the use of data cubes to integrate time-series information into the GEOBIA classification.

Estimating the temporal heterogeneity of mowing events on grassland for haymilk-production using Sentinel-2 and greenness-index

The temporal and spatial heterogeneity of mowing on intensively used grasslands can be an important factor influencing the survival of insects or small animals after harvest. High heterogeneity contributes to the preservation of biodiversity on these grasslands. In our study we use satellite image time-series to analyze the temporal distribution of mowing events on grassland patches of Austrian farms. Our approach compares the standard deviation and the mean values of a greenness-index in Sentinel-2 time-series, and does not require to determine the date of mowing. A high standard deviation of the greenness-index demonstrates that the area of interest (AOI) includes grassland patches with tall and short (un-mowed and mowed) grass at a certain timepoint. If large parts of the AOI are mown at once, this results in a low standard deviation and a low mean value of the greenness-index. We were able to show that the haymilk farms in our study applied a more heterogeneous temporal mowing schedule on their grasslands compared to conventional farms. The data and information were obtained from the Austrian Sentinel-2 semantic EO data cube (Sen2Cube.at), which contains multi-spectral imagery data from Sentinel-2 satellites as well as derived semantic (categorical) and greenness-index layers since 2015.

In-season and dynamic crop mapping using 3D convolution neural networks and sentinel-2 time series

An accurate, frequently updated, automatic and reproducible mapping procedure to identify seasonal cultivated crops is a prerequisite for many crop monitoring activities. Deep learning was demonstrated to be an effective mapping approach already successfully applied to decametric resolution satellite images (like Sentinel-2 data) to produce yearly crop maps. In this framework, algorithm training is performed with ground truth typically consisting of spatially explicit information available after the end of the season (e.g. yearly crop maps and/or farmer declaration for subsidies at parcel level); however, such data (i) does not allow performing in-season prediction, and (ii) does not provide temporal details fundamental to describe a dynamic crop succession and/or to understand crop management (i.e. planting and harvesting). In this paper we present a Deep Neural Network-based approach capable of generating (i) a crop map of the current season at a specific point in time (“In season mapping” conventionally at the end of the current year), along with (ii) all intermediate maps during the season able to describe in near real-time the evolution of crop presence (“Dynamic-mapping” at the temporal granularity of satellite imagery revisiting, e.g., 5 days for Sentinel-2 data). This approach adopts a smart training procedure of a Deep Neural model by exploiting historical satellite data and ground truth. We introduce a method to automatically generate “short-term” ground truth maps (i.e. 5 days reference) starting from the “long-term” ones (i.e. available yearly static reference) and characterizing temporally the different crop presence by performing a phenological analysis of historical time series. The model was trained and validated in Lombardy (North of Italy) exploiting multi-annual authoritative crop maps from 2016 to 2019. Validation was performed both in time (same areas used for training in a different year) and space (different location) for the year 2019. The quantitative error metrics calculation and Spatio-temporal analysis clearly demonstrate that the model can predict in-season crop presence with a generalization capacity over the long-term (yearly maps: OA > 70% and Kappa > 0.64%) and that the short-term predictions (5 days maps) are coherent with the reference information from expert knowledge (local crop calendars). The model can produce dynamically along the season short-term maps with a medium-high crop-specific User Accuracy at the maximum green-up phase (UA > 53% up to 95%). These products are of extreme interest for final users providing information at the peak of plant development that dynamically changes according to the considered crop, the specific location and the investigated season. These results demonstrate that it is possible to produce a crop map early in the season and extract useful additional information such as crop intensity (e.g. double crops presence) and crop dynamics related to different sowing dates.

Exploring the potential of multi-source unsupervised domain adaptation in crop mapping using Sentinel-2 images

ABSTRACT Accurate crop mapping is critical for agricultural applications. Although studies have combined deep learning methods and time-series satellite images to crop classification with satisfactory results, most of them focused on supervised methods, which are usually applicable to a specific domain and lose their validity in new domains. Unsupervised domain adaptation (UDA) was proposed to solve this limitation by transferring knowledge from source domains with labeled samples to target domains with unlabeled samples. Particularly, multi-source UDA (MUDA) is a powerful extension that leverages knowledge from multiple source domains and can achieve better results in the target domain than single-source UDA (SUDA). However, few studies have explored the potential of MUDA for crop mapping. This study proposed a MUDA crop classification model (MUCCM) for unsupervised crop mapping. Specifically, 11 states in the U.S. were selected as the multi-source domains, and three provinces in Northeast China were selected as individual target domains. Ten spectral bands and five vegetation indexes were collected at a 10-day interval from time-series Sentinel-2 images to build the MUCCM. Subsequently, a SUDA model Domain Adversarial Neural Network (DANN) and two direct transfer methods, namely, the deep neural network and random forest, were constructed and compared with the MUCCM. The results indicated that the UDA models outperformed the direct transfer models significantly, and the MUCCM was superior to the DANN, achieving the highest classification accuracy (OA>85%) in each target domain. In addition, the MUCCM also performed best in in-season forecasting and crop mapping. This study is the first to apply a MUDA to crop classification and demonstrate a novel, effective solution for high-performance crop mapping in regions without labeled samples.

A Sub-Seasonal Crop Information Identification Framework for Crop Rotation Mapping in Smallholder Farming Areas with Time Series Sentinel-2 Imagery

Accurate crop rotation information is essential for understanding food supply, cropland management, and resource allocation, especially in the context of China’s basic situation of “small farmers in a big country”. However, crop rotation mapping for smallholder agriculture systems remains challenging due to the diversity of crop types, complex cropping practices, and fragmented cropland. This research established a sub-seasonal crop information identification framework for crop rotation mapping based on time series Sentinel-2 imagery. The framework designed separate identification models based on the different growth seasons of crops to reduce interclass similarity caused by the same crops in a certain growing season. Features were selected separately according to crops characteristics, and finally explored rotations between them to generate the crop rotation map. This framework was evaluated in the study area of Shandong Province, China, a mix of single-cropping and double-cropping smallholder area. The accuracy assessment showed that the two crop maps achieved an overall accuracy of 0.93 and 0.85 with a Kappa coefficient of 0.86 and 0.80, respectively. The results showed that crop rotation practice mainly occurred in the plains of Shandong, and the predominant crop rotation pattern was wheat and maize. In addition, Land Surface Water Index (LSWI), Soil-Adjusted Vegetation Index (SAVI), Green Chlorophyll Vegetation Index (GCVI), red-edge, and other spectral bands during the peak growing season enabled better performance in crop mapping. This research demonstrated the capability of the framework to identify crop rotation patterns and the potential of the multi-temporal Sentinel-2 for crop rotation mapping under smallholder agriculture system.

Vietnam wetland cover map: using hydro-periods Sentinel-2 images and Google Earth Engine to explore the mapping method of tropical wetland

Tropical wetland resources have high ecological value to flood storage, greenhouse gas emissions, and biodiversity, with diverse types and widespread distribution. However, cloud pollution, inundation fluctuations, and spectral similarity pose unique challenges in the monitoring of tropical wetland using remote sensing. In this study, we proposed a comprehensive method for mapping the various types of tropical wetland using time series Sentinel-2 images based on the Google Earth Engine (GEE) platform integrated inundation dynamic, phenological, and geography features for multi-class topical wetland mapping (IPG-MTWM). Obtained a precise Vietnam wetland cover map (VWeC), which covers ten types of wetlands within a spatial resolution of 10 m. The overall accuracy (OA) of the VWeC was 82.07 %. compared the VWeC results with other wetland products showing the advantages of the IPG-MTWM in mapping complete wetland cover types with high spatial resolution and accuracy at large scales. The VWeC is the first thematic mapping on wetland cover types of Vietnam, it indicates that Vietnam has approximately 1,367,502.16 ha wetlands, of which coastal wetland accounts for approximately 44.27 % (605,318.3 ha). Vietnam’s wetlands are mainly distributed in South Vietnam (SV). The integrated inundation dynamic, phenological, and geography features method, highlights the potential of precise and fast multi-class topical wetland mapping at a large-scale using hydro-periods Sentinel-2 images and GEE platform. This study provides an important technique and dataset for tropical wetland monitoring, protection, and management.

Global mapping of the landside clustering of aquaculture ponds from dense time-series 10 m Sentinel-2 images on Google Earth Engine

• Spatial and dense temporal features for extracting global land-side aquaculture. • Global aquaculture pond distribution information was obtained for the first time. • Asia accounted for 89.12% of global land-side aquaculture. • A 30-km buffer of coastline inland side accounted for 75.57% • The result is highly consistent with FAO aquaculture statistics. Spatial distribution information offers a valuable resource for the growing research on evaluating and managing aquaculture and the mechanism of its interaction with the environment, especially concerning landside clustering aquaculture ponds (LCAP). However, studies aimed at obtaining data on global LCAP distribution remains limited. This study combined edge detection and the morphology implemented on Google Earth Engine (GEE) platform to produce the first global spatial distribution data of LCAP using 4,015,054 tiles of the 10-m Sentinel-2 time-series images collected throughout 2020. The study findings showed a total area of global LCAP of 55,337.03 km 2 . Asia, led by China, had the highest distribution area of LCAP, accounting for 89.12 % of global LCAP. The main LCAP distribution area was the landside 30-km buffer region of coastline, accounting for 75.57 % of global LCAP. Meanwhile, the greatest proportion of LCAP (69.63 %) was located in China. Accuracy verification revealed that the precision and recall error of the results were 83.91 % and 92.49 % respectively, and F 1 score was 0.88. A comparison of China, Vietnam, and India, based on data drawn from most of the existing local studies that used remote sensing methods, showed that these countries’ aquaculture area differs by less than 10 %. Despite some variation from FAO statistics or official national statistics, the results of our remote sensing-based method show promise in such aspects as global coverage and temporal coherency when compared with the classical statistical method.

Crop classification by using dual-pol SAR vegetation indices derived from Sentinel-1 SAR-C data

In the following study, an attempt is made for crop classification of rainy season through analyzing time-series Sentinel-1 SAR data of May 2020 to September 2020. The SVIDP index derived from dual-pol (VV and VH) bands consisting of NRPB ([Formula: see text]), DPDD [Formula: see text]), IDPDD ([Formula: see text]), and VDDPI [Formula: see text] ratios are utilized for discriminating inter-vegetative boundaries of crop pixels. This study was conducted near Karnal city region, Karnal district, Haryana, India. The Sentinel-1 data has the capability to penetrate thick cloud cover and provide high revisit frequency data for rain-fed crops. Obtained classification achieved higher accuracy in both RF (93.77%) and SVM (93.50%) classifiers. Obtained linear regression statistics of mean raster imagery reveals that IDPDD index is much sensitive to other crop which has highest standard deviations in σvh° and σvv° bands throughout the period, and high R2 with σvh° (0.70), VV (0.58), NRPB (0.693), and DPDD (0.697) indices. In contrast to this, IDPDD index has least correlation (< 0.289) with σvh°, σvv°, EVI 2, NRPB, and DPDD indices for water body which has smooth surface and lowest SAR backscattering with minimum standard deviations in the same period.

Wheat Yield Estimation Using Remote Sensing Indices Derived from Sentinel-2 Time Series and Google Earth Engine in a Highly Fragmented and Heterogeneous Agricultural Region

In Morocco, monitoring and estimation of wheat yield at the regional and national scales are critical issues for national food security. The recent Sentinel-2 imagery offers potential for managing grain production systems on a field and regional level. The present study was planned based on a time series of six remote sensing indices and Multiple Linear Regression (MLR) methods for real-time estimation of wheat yield using the Google Earth Engine (GEE) platform in a highly heterogeneous and fragmented agricultural region, such as the Tadla Irrigated Perimeter (TIP). First, the spatial distribution of wheat in the TIP region was mapped by performing Random Forest (RF) classification of Sentinel 2 images. Following that, using MLR models, the wheat yield of nine sampled fields was estimated for the different phenological stages of wheat. The yield measured in-situ was the independent variable of the regressions. The dependent variables included the remote sensing indices derived from Sentinel-2. The remote sensing index and the phenological period of the greatest model were investigated to estimate and map the wheat yield in the entire study area. The RF generated the wheat mapping of the study area with an overall accuracy (OA) of 93.82%. Furthermore, the coefficient of determination (R2) of the tested MLR was from 0.53 to 0.89, while the Root Mean Square Error (RMSE) varied from 4.29 to 7.78 q ha−1. The best model was the one that uses the Green Normalized Difference Vegetation Index (GNDVI) in the tillering and maturity stages.

Deep learning models to map an agricultural expansion area with MODIS and Sentinel-2 time series images

Mapping changing land use and land cover (LULC) is important for land management and environment analysis. We tried to build deep learning models to classify LULC over time at an agricultural expansion area in Matopiba region, Brazil with MCD43A4 V006 moderate resolution imaging spectroradiometer (MODIS), and Sentinel-2 multispectral instrument (MSI) time series data. We collected time series MODIS data and Sentinel-2 A/B MSI data from 2015 to 2020 and prepared small patches with containing blue, green, red, near-infrared, and shortwave-infrared-1 bands as features. Then the both datasets were used to build and train the convolutional neural network (CNN) model, the CNN gate recurrent unit (CNN-GRU) model, and the CNN long short-term memory (CNN-LSTM) model, respectively. We evaluated these three trained models with ground truth data, and the CNN-LSTM model (overall accuracy: 91.29% from MODIS data and 89.47% from Sentinel-2 data) was better than the CNN-GRU model (overall accuracy: 89.19% from MODIS data and 88.61% from Sentinel-2 data) and the CNN model (overall accuracy: 89.17% from MODIS data and 86.02% from Sentinel-2 data). Our results also showed that the accuracy from cropland and savanna classes were higher than grassland and forest classes in all three models. These two classes generated from the CNN-LSTM model performed better than the other two deep learning models. The results from these two datasets indicated that the methods were reliable for both coarse and medium spatial resolution satellite images and time series remote sensing images worked better than single image for classification problems when considering LULC change over time. The results also provided an alternative way to prepare input data from satellite images for deep learning models. Furthermore, the classification results of the whole agricultural expansion area were reasonable and it can be used as an additional dataset for further environmental analysis at a regional scale.

Assessing the Added Value of Sentinel-1 PolSAR Data for Crop Classification

Crop classification is an important remote sensing task with many applications, e.g., food security monitoring, ecosystem service mapping, climate change impact assessment, etc. This work focuses on mapping 10 crop types at the field level in an agricultural region located in the Spanish province of Navarre. For this, multi-temporal Synthetic Aperture Radar Polarimetric (PolSAR) Sentinel-1 imagery and multi-spectral Sentinel-2 data were jointly used. We applied the Cloude–Pottier polarimetric decomposition on PolSAR data to compute 23 polarimetric indicators and extracted vegetation indices from Sentinel-2 time-series to generate a big feature space of 818 features. In order to assess the relevance of the different features for the crop mapping task, we run a number of scenarios using a Support Vector Machines (SVM) classifier. The model that was trained using only the polarimetric data demonstrates a very promising performance, achieving an overall accuracy over 82%. A genetic algorithm was also implemented as a feature selection method for deriving an optimal feature subset. To showcase the positive effect of using polarimetric data over areas suffering from cloud coverage, we contaminated the original Sentinel-2 time-series with simulated cloud masks. By incorporating the genetic algorithm, we derived a high informative feature subset of 120 optical and polarimetric features, as the corresponding classification model increased the overall accuracy by 5% compared to the model trained only with Sentinel-2 features. The feature importance analysis indicated that apart from the Sentinel-2 spectral bands and vegetation indices, several polarimetric parameters, such as Shannon entropy, second eigenvalue and normalised Shannon entropy are of high value in identifying crops. In summary, the findings of our study highlight the significant contribution of Sentinel-1 PolSAR data in crop classification in areas with frequent cloud coverage and the effectiveness of the genetic algorithm in discovering the most informative features.

Crop type mapping using time-series Sentinel-2 imagery and U-Net in early growth periods in the Hetao irrigation district in China

The information on spatial distribution pattern and area of different crops is particularly useful for monitoring and managing the sustainability of agricultural resources. However, there are still challenges in timely mapping of crop types and planting areas to support production during the present growing season. Here, we proposed to use time-series Sentinel-2 imagery and a deep learning method for major crop mapping (wheat, maize, sunflower, and squash) in the agricultural production areas in the Hetao Irrigation District (HID), Inner Mongolia Autonomous Region, China. A feature selection method that combines the global separability index and feature recursive elimination was proposed to optimize the band features of Sentinel-2 imagery. The Random Forest, Extreme Gradient Boosting, U-Net and deeplabv3+ algorithms were then used for image classification with all spectral bands and the selected band features, respectively. The selected bands and best method were used to determine the key time window for the early identification of four major crops. Integrate all image datasets in the early window period, and build mapping models suitable for images at different times in the window period by combining the selected bands and best method. The results showed that the proposed feature selection method effectively removed redundant features and improved the classification efficiency. The U-Net algorithm was more suitable for the classification of major crops in HID with a higher accuracy whether using all band features or the selected features. The key periods for early recognition of wheat, maize, sunflower, and squash in HID were mid-May, late June, early July, and late June, respectively. The constructed models for crop type mapping in the early period were tested in 2020 and 2021 in the HID, and the F1-score, the mean intersection-over-union and overall accuracy were 86.71 %, 84.66 %, and 98.89 %, respectively. The proposed feature selection method, U-Net classification algorithm, and screening of the early mapping window period in this study provide an efficient, accurate and timely crop mapping tool in HID. Future research can further verify the applicability of the early identification time window proposed in this study to support other agricultural management activities in the areas.

Early-Season Crop Identification in the Shiyang River Basin Using a Deep Learning Algorithm and Time-Series Sentinel-2 Data

Timely and accurate crop identification and mapping are of great significance for crop yield estimation, disaster warning, and food security. Early-season crop identification places higher demands on the quality and mining of time-series information than post-season mapping. In recent years, great strides have been made in the development of deep-learning algorithms, and the emergence of Sentinel-2 data with a higher temporal resolution has provided new opportunities for early-season crop identification. In this study, we aimed to fully exploit the potential of deep-learning algorithms and time-series Sentinel-2 data for early-season crop identification and early-season crop mapping. In this study, four classifiers, i.e., two deep-learning algorithms (one-dimensional convolutional networks and long and short-term memory networks) and two shallow machine-learning algorithms (a random forest algorithm and a support vector machine), were trained using early-season Sentinel-2 images and field samples collected in 2019. Then, these algorithms were applied to images and field samples for 2020 in the Shiyang River Basin. Twelve scenarios with different classifiers and time intervals were compared to determine the optimal combination for the earliest crop identification. The results show that: (1) the two deep-learning algorithms outperformed the two shallow machine-learning algorithms in early-season crop identification; (2) the combination of a one-dimensional convolutional network and 5-day interval time-series Sentinel-2 data outperformed the other schemes in obtaining the early-season crop identification time and achieving early mapping; and (3) the early-season crop identification mapping time in the Shiyang River Basin was identified as the end of July, and the overall classification accuracy reached 0.83. In addition, the early identification time for each crop was as follows: the wheat was in the flowering stage (mid-late June); the alfalfa was in the first harvest (mid-late June); the corn was in the early tassel stage (mid-July); the fennel and sunflower were in the flowering stage (late July); and the melons were in the fruiting stage (around late July). This study demonstrates the potential of using Sentinel-2 time-series data and deep-learning algorithms to achieve early-season crop identification, and this method is expected to provide new solutions and ideas for addressing early-season crop identification monitoring.

Improved LSTM Model for Boreal Forest Height Mapping Using Sentinel-1 Time Series

Time series of SAR imagery combined with reference ground data can be suitable for producing forest inventories. Copernicus Sentinel-1 imagery is particularly interesting for forest mapping because of its free availability to data users; however, temporal dependencies within SAR time series that can potentially improve mapping accuracy are rarely explored. In this study, we introduce a novel semi-supervised Long Short-Term Memory (LSTM) model, CrsHelix-LSTM, and demonstrate its utility for predicting forest tree height using time series of Sentinel-1 images. The model brings three important modifications to the conventional LSTM model. Firstly, it uses a Helix-Elapse (HE) projection to capture the relationship between forest temporal patterns and Sentinel-1 time series, when time intervals between datatakes are irregular. A skip-link based LSTM block is introduced and a novel backbone network, Helix-LSTM, is proposed to retrieve temporal features at different receptive scales. Finally, a novel semisupervised strategy, Cross-Pseudo Regression, is employed to achieve better model performance when reference training data are limited. CrsHelix-LSTM model is demonstrated over a representative boreal forest site located in Central Finland. A time series of 96 Sentinel-1 images are used in the study. The developed model is compared with basic LSTM model, attention-based bidirectional LSTM and several other established regression approaches used in forest variable mapping, demonstrating consistent improvement of forest height prediction accuracy. At best, the achieved accuracy of forest height mapping was 28.3% relative root mean squared error (rRMSE) for pixel-level predictions and 18.0% rRMSE on stand level. We expect that the developed model can also be used for modeling relationships between other forest variables and satellite image time series.

Predicting plant biomass and species richness in temperate grasslands across regions, time, and land management with remote sensing and deep learning

Spatial predictions of biomass production and biodiversity at regional scale in grasslands are critical to evaluate the effects of management practices across environmental gradients. New generations of remote sensing sensors and machine learning approaches can predict these grassland characteristics with varying accuracy. However, such studies frequently fail to cover a sufficiently broad range of environmental conditions, and their prediction models are often case-specific. To address this gap, we have modelled above-ground biomass and species richness in 150 spatially independent grassland plots of three geographical regions in Germany. These regions follow a North-South climate gradient and differ in soil types, topography, elevation, climatic conditions, historical contexts, and management intensities. The predictors tested in this study are Sentinel-1 backscatter, Sentinel-2 time series of surface reflectance along with derived vegetation indices and Rao's Q, and a set of topoedaphic variables. We compared the performance of a feed-forward deep neural network (DNN) with a random forest (RF) regression algorithm. The DNN achieved the best estimations of biomass (r2 = 0.45) when trained with Sentinel-2 surface reflectance only. Moreover, the DNN showed a higher generalizability than RF during spatial cross-validations (i.e., calibrating and validating in different regions, r2 = 0.38 vs. 0.26). Species richness predictions by both algorithms improved when the full time series of Sentinel-2 surface reflectance values were used (highest r2 = 0.42 achieved by the DNN), but both performed poorly during spatial cross-validations. Overall, the DNN-based models were more robust than RF models, showed a lower bias and lower systematic error, and required fewer inputs. Explainability analysis indicated that red-edge and near infrared information from May and October was the most relevant to predict species richness. This study presents an important step forward in generating robust spatially explicit predictions of grassland attributes and biodiversity variables across large areas, environmental gradients, and phenological stages.

Deep semantic segmentation of mangroves in Brazil combining spatial, temporal, and polarization data from Sentinel-1 time series

The automatic and accurate detection of mangroves from remote sensing data is essential to assist in conservation strategies and decision-making that minimize possible environmental damage, especially for the Brazilian coast with continental dimensions. In this context, deep learning segmentation techniques are powerful tools with successful applications in several fields. However, few studies used deep learning for mangrove areas, and none considered radar time series. The present research has the following objectives: (a) develop a deep learning mangrove dataset in the Southeast region of Brazil considering the spatial, temporal, and polarization dimensions; (b) evaluate the U-net architecture with three backbones (ResNet-101, VGG16, and Efficient-net-B7); (c) compare the VV and VH polarizations and their combination; (d) evaluate the time series composition with a varying number of images (29, 15, 8, 4 images); and compare the sliding windows approach using five stride values (8, 16, 32, 64, and 128) for large image classification. This research used the annual Sentinel-1 time series for the period 2017–2020. Image annotation used manual interpretation, and the developed dataset has 2886 images with spatial dimensions of 128 × 128 pixels (2136 for training, 450 for validation, and 300 for testing). We found that the best results considered: (a) both polarizations (VV&VH), (b) the maximum number of images in the time series images (29), (c) the U-net with the Efficient-net-B7 backbone (97.35% overall accuracy, 85.77% precision, 84.96% recall, 85.36% F-score, and 74.46% IoU), and (d) the smallest stride value (8) presented the best results for large image classification. The proposed method is suitable and effective for monitoring mangrove patterns over time, providing accurate maps of these ecosystems.

Application of Parallel Factor Analysis (PARAFAC) to the Regional Characterisation of Vineyard Blocks Using Remote Sensing Time Series

Monitoring wine-growing regions and maximising the value of production based on their region/local specificities requires accurate spatial and temporal monitoring. The increasing amount and variability of information from remote sensing data is a potential tool to assess this challenge for the grape and wine industry. This article provides a first insight into the capacity of a multiway analysis method applied to Sentinel-2 time series to assess the value of simultaneously considering spectral and temporal information to highlight site-specific canopy evolution in relation to environmental factors and management practices, which present a large diversity at this regional scale. Parallel Factor Analysis (PARAFAC) was used as an unsupervised technique to recover pure spectra and temporal signatures from multi-way spectral imagery of vineyards in the Languedoc-Roussillon region in the south of France. The model was developed using a time series of Sentinel-2 satellite imagery collected over 4978 vineyard blocks between May 2019 and August 2020. From the Sentinel-2 (spectral and temporal) signal, the PARAFAC analysis allowed the identification of spectral and temporal profiles in the form of pure components, which corresponded to vegetation and soil. The PARAFAC analysis also identified that two of the pure spectra were strongly related to characteristics and dynamics of vineyard cultivation at a regional scale. A conceptual framework was proposed in order to simultaneously consider both vegetation and soil profiles and to summarise the mass of data accordingly. This methodology allowed the computation of a concentration index that characterised how close a field was to a vegetation or a soil profile over the season. The concentration indices were validated for the vegetation and the soil over two growing seasons (2019 and 2020) with geostatistical analysis. A non-random distribution of the concentration index at the regional scale was assumed to highlight a strongly spatially organised phenomenon related to spatially organised environmental factors (soil, climate, training system, etc.). In a second step, spatial patterns of indices were subjected to the expertise of a panel of advisors of the wine industry in order to validate them in relation to vine-growing conditions. Results showed that the introduction of the PARAFAC method opened up the possibility to identify relevant spectro-temporal profiles for vine monitoring purposes.

Crop Classification and Representative Crop Rotation Identifying Using Statistical Features of Time-Series Sentinel-1 GRD Data

Compared with a monoculture planting mode, the practice of crop rotations improves fertilizer efficiency and increases crop yield. Large-scale crop rotation monitoring relies on the results of crop classification using remote sensing technology. However, the limited crop classification accuracy cannot satisfy the accurate identification of crop rotation patterns. In this paper, a crop classification and rotation mapping scheme combining the random forest (RF) algorithm and new statistical features extracted from time-series ground range direction (GRD) Sentinel-1 images. First, the synthetic aperture radar (SAR) time-series stacks are established, including VH, VV, and VH/VV channels. Then, new statistical features named the objected generalized gamma distribution (OGΓD) features are introduced to compare with other object-based features for each polarization. The results showed that the OGΓD σVH achieved 96.66% of the overall accuracy (OA) and 95.34% of the Kappa, improving around 4% and 6% compared with the object-based backscatter in VH polarization, respectively. Finally, annual crop-type maps for five consecutive years (2017–2021) are generated using the OGΓD σVH and the RF. By analyzing the five-year crop sequences, the soybean-corn (corn-soybean) is the most representative rotation in the study region, and the soybean-corn-soybean-corn-soybean (together with corn-soybean-corn-soybean-corn) has the highest count with 100 occurrences (25.20% of the total area). This study offers new insights into crop rotation monitoring, giving the basic data for government food planning decision-making.

DeepForest: Novel Deep Learning Models for Land Use and Land Cover Classification Using Multi-Temporal and -Modal Sentinel Data of the Amazon Basin

Land use and land cover (LULC) mapping is a powerful tool for monitoring large areas. For the Amazon rainforest, automated mapping is of critical importance, as land cover is changing rapidly due to forest degradation and deforestation. Several research groups have addressed this challenge by conducting local surveys and producing maps using freely available remote sensing data. However, automating the process of large-scale land cover mapping remains one of the biggest challenges in the remote sensing community. One issue when using supervised learning is the scarcity of labeled training data. One way to address this problem is to make use of already available maps produced with (semi-) automated classifiers. This is also known as weakly supervised learning. The present study aims to develop novel methods for automated LULC classification in the cloud-prone Amazon basin (Brazil) based on the labels from the MapBiomas project, which include twelve classes. We investigate different fusion techniques for multi-spectral Sentinel-2 data and synthetic aperture radar Sentinel-1 time-series from 2018. The newly designed deep learning architectures—DeepForest-1 and DeepForest-2—utilize spatiotemporal characteristics, as well as multi-scale representations of the data. In several data scenarios, the models are compared to state-of-the-art (SotA) models, such as U-Net and DeepLab. The proposed networks reach an overall accuracy of up to 75.0%, similar to the SotA models. However, the novel approaches outperform the SotA models with respect to underrepresented classes. Forest, savanna and crop were mapped best, with F1 scores up to 85.0% when combining multi-modal data, compared to 81.6% reached by DeepLab. Furthermore, in a qualitative analysis, we highlight that the classifiers sometimes outperform the inaccurate labels.