Sentinel-2 Time Series Research Articles

Spatio-temporal dynamics of tree cover change in semi-arid Ghana: evidence from Sentinel-2 time series data

ABSTRACT Trees provide vital ecosystem services, yet anthropogenic activities have continued to reduce tree cover (TC) globally, leading to disparities in the distribution of ecosystem services at various scales. Similarly, empirical analysis of indicators that measure variations in TC at various temporal and spatial scales is limited in the semi-arid regions where global environmental challenges such as climate change are already affecting TC distribution. This paper quantified recent (2017 - 2019) TC changes in the semi-arid region of Ghana using Sentinel-2 times series data and the application of landscape metrics. The results showed that TC declined by 30% within two years. Urbanizing areas showed more rapid TC depletion rates, while some rural areas experienced a slight increase in TC. The findings suggest the need for local environmental governance by consolidating actions towards conserving semi-arid TC. Sustainable charcoal and fuelwood production are imperative for a long-term sustainable supply of ecosystem services.

Extending polarimetric optimization of multi-temporal InSAR techniques on dual polarized Sentinel-1 data

The rationale of combining polarimetric SAR data with multi-temporal InSAR techniques, which is based on polarimetric optimization, is to improve their performance in identifying coherent pixels and increasing the accuracy of deformation monitoring. This paper extended the polarimetric optimization of Persistent Scatterer Interferometry (PSI) and Small Baseline Subset (SBAS) techniques on dual-polarized Sentinel-1 data. For this purpose, 20 dual-polarization (VV/VH) SAR images acquired by the Sentinel-1A sensor were used. The results indicated that by applying polarimetric optimization on dual-polarized data in the SBAS technique, the number of candidate and final coherent pixels increased 2.95 and 1.42 times, respectively. Similarly, the number of candidate and final coherent pixels (CP) increased 3.12 and 1.45 times in the PSI technique, respectively. Also, the effect of the polarimetric optimization on increasing the density of coherent points in different regions was investigated. In this framework, Sentinel-2 time-series images were classified into four classes (bare soil, vegetation, build-up, and others) using Google Earth Engine by applying remote sensing indices. This demonstrated that polarimetric optimization of the SBAS technique was more effective in bare soil and vegetated regions than in urban areas. However, polarimetric optimization of the PSI technique was more successful in non-urban regions than in urban areas. An external comparative analysis was also performed between the deformation results derived by PSI and SBAS techniques and GPS observation. The evaluation demonstrated that polarimetric optimization coupled with multi-temporal InSAR analysis yielded more reliable deformation results than conventional single-polarimetric data.

Continental-scale wetland mapping: A novel algorithm for detailed wetland types classification based on time series Sentinel-1/2 images

Wetlands have suffered from considerable degradation due to anthropogenic and natural disturbances in recent decades. Although some advancements have been made, effective methods that can produce large-scale wetland maps with detailed categories are still lacking due to the diversity and complexity of wetland ecosystems. To address this issue, we developed a novel algorithm for detailed wetland types classification integrating k-fold random forest and hierarchical decision tree, and so named two-step classification algorithm. Firstly, the phenology-based features were composited based on time series Sentinel-1/2 images, and the k-fold random forest was used to extract five rough wetland types in Google Earth Engine platform. Secondly, the hierarchical decision tree designed based on geometric features was used to separate the rough wetland types into fourteen detailed types. Application of the two-step classification method in Northern, Central and Southern Asia (NCSA) resulted in a continental-scale wetland map with an overall accuracy of 90.0 ± 0.5%. Wetland types, including inland marsh, lake, river, coastal swamp, estuarine water, lagoon, shallow marine water, reservoir, canal/channel and agricultural pond, had good accuracy with both UA and PA over 77%. The remaining wetland types had moderate accuracy, with both UA and PA over 58%. As we calculated, total wetland areas of NCSA were 1,375,489.27 km2. Among the fourteen wetland categories, the inland marsh had the largest area (544,584.38 km2) and was primarily distributed in subarctic and humid continental climates, while the canal/channel had the smallest area (1,651.57 km2) and was primarily scattered in desert, semiarid and humid subtropical climates. The lake and floodplain shared generally large areas with value of 392,413.55 km2 and 173,255.71 km2 respectively, which were typically distributed across desert and semiarid climates. This study successfully mapped continental-scale wetlands with detailed categories at a 10-m spatial resolution, which can provide valuable information for the management of wetland ecosystems and facilitate the implementation of wetland-related sustainable development goals.

UAV-based reference data for the prediction of fractional cover of standing deadwood from Sentinel time series

Increasing tree mortality due to climate change has been observed globally. Remote sensing is a suitable means for detecting tree mortality and has been proven effective for the assessment of abrupt and large-scale stand-replacing disturbances, such as those caused by windthrow, clear-cut harvesting, or wildfire. Non-stand replacing tree mortality events (e.g., due to drought) are more difficult to detect with satellite data – especially across regions and forest types. A common limitation for this is the availability of spatially explicit reference data. To address this issue, we propose an automated generation of reference data using uncrewed aerial vehicles (UAV) and deep learning-based pattern recognition. In this study, we used convolutional neural networks (CNN) to semantically segment crowns of standing dead trees from 176 UAV-based very high-resolution (<4 cm) RGB-orthomosaics that we acquired over six regions in Germany and Finland between 2017 and 2021. The local-level CNN-predictions were then extrapolated to landscape-level using Sentinel-1 (i.e., backscatter and interferometric coherence), Sentinel-2 time series, and long short term memory networks (LSTM) to predict the cover fraction of standing deadwood per Sentinel-pixel. The CNN-based segmentation of standing deadwood from UAV imagery was accurate (F1-score = 0.85) and consistent across the different study sites and years. Best results for the LSTM-based extrapolation of fractional cover of standing deadwood using Sentinel-1 and -2 time series were achieved using all available Sentinel-1 and --2 bands, kernel normalized difference vegetation index (kNDVI), and normalized difference water index (NDWI) (Pearson’s r = 0.66, total least squares regression slope = 1.58). The landscape-level predictions showed high spatial detail and were transferable across regions and years. Our results highlight the effectiveness of deep learning-based algorithms for an automated and rapid generation of reference data for large areas using UAV imagery. Potential for improving the presented upscaling approach was found particularly in ensuring the spatial and temporal consistency of the two data sources (e.g., co-registration of very high-resolution UAV data and medium resolution satellite data). The increasing availability of publicly available UAV imagery on sharing platforms combined with automated and transferable deep learning-based mapping algorithms will further increase the potential of such multi-scale approaches.

Mapping Maize Tillage Practices over the Songnen Plain in Northeast China Using GEE Cloud Platform

As the population grows, the development of conservation tillage offers a means of promoting the sustainability of agricultural engineering. Remote sensing images with high spatial and temporal resolutions enable the accurate monitoring of conservation tillage on a broad spatial scale, further promoting conservation tillage research. This paper describes using streamlined time series Sentinel-2 images based on the Google Earth Engine (GEE) cloud platform for mapping maize tillage practices in the Songnen Plain region of Northeast China. Based on the correlation with the normalized difference tillage index (NDTI) and maize residue coverage (MRC) data, the optimal time series and streamlining functions in the GEE cloud platform are determined. Estimates of MRC and the mapping of tillage practices in the Songnen Plain for 2019–2022 are then determined using GEE and a previous model. Geostatistical analysis using ArcGIS is applied to analyze the spatial and temporal distribution characteristics of MRC and conservation tillage over the Songnen Plain. The results show that time series images from 20–30 May achieve an r value of 0.902 and an R2 value of 0.8136 when using the median streamlining function. The mean MRC for the study area in 2022 is 2.3%, and an overall upward trend in conservation tillage is observed (from 0.08% in 2019 to 0.25% in 2022). Our analysis shows that MRC monitoring and conservation tillage mapping can be performed over a broad spatial scale using remote sensing technology based on the GEE cloud platform. Spatial and temporal information on farm practices provides a theoretical basis for agricultural development planning efforts, which can promote sustainable agricultural development.

Natura 2000 Grassland Habitats Mapping Based on Spectro-Temporal Dimension of Sentinel-2 Images with Machine Learning

Habitat mapping is essential for the management and monitoring of Natura 2000 sites. Time-consuming field surveys are still the most frequently used solution for the implementation of the European Habitats Directive, but the use of remote sensing tools for this is becoming more common. The high temporal resolution of Sentinel-2 data, registering the visible, near, and shortwave infrared ranges of the electromagnetic spectrum, makes them valuable material in this context. In this study, we aimed to use multitemporal Sentinel-2 data for mapping three grassland Natura 2000 habitats in Poland. We performed the classification based on spectro-temporal features extracted from data collected from eight different terms within the year 2017 using Convolutional Neural Networks (CNNs), and we also tested other widely used machine learning algorithms for comparison, such as Random Forests (RFs) and Support Vector Machines (SVMs). Based on ground truth data, we randomly selected training and validation polygons and then performed the evaluation iteratively (100 times). The best resulting median F1 accuracies that we obtained for habitats were as follows: 6210, 0.85; 6410, 0.80; and 6510, 0.84 (with SVM). Finally, we concluded that the accuracy of the results was comparable, but we obtained the best results using SVM (median OA = 88%, with 86% for RF and 84% for CNNs). In this work, we confirmed the usefulness of the spectral dimension of Sentinel-2 time series data for mapping grassland habitats, and researchers of future work can further develop the use of CNNs for this purpose.

Sentinel-2 time series analysis for monitoring multi-taxon biodiversity in mountain beech forests

Biodiversity monitoring represents a major challenge to supporting proper forest ecosystem management and biodiversity conservation. The latter is indeed shifting in recent years from single-species to multi-taxon approaches. However, multi-taxonomic studies are quite rare due to the effort required for performing field surveys. In this context, remote sensing is a powerful tool, continuously providing consistent and open access data at a different range of spatial and temporal scales. In particular, the Sentinel-2 (S2) mission has great potential to produce reliable proxies for biological diversity. In beech forests of two Italian National Parks, we sampled the beetle fauna, breeding birds, and epiphytic lichens. First, we calculated Shannon’s entropy and Simpson’s diversity. Then, to produce variables for biodiversity assessment, we exploited S2 data acquired in the 4 years 2017–2021. S2 images were used to construct spectral bands and photosynthetic indices time series, from which 91 harmonic metrics were derived. For each taxon and multi-taxon community, we assessed the correlation with S2 harmonic metrics, biodiversity indices, and forest structural variables. Then, to assess the potential of the harmonic metrics in predicting species diversity in terms of Shannon’s and Simpson’s biodiversity indices, we also fit a random forests model between each diversity index and the best 10 harmonic metrics (in terms of absolute correlation, that is, the magnitude of the correlation) for each taxon. The models’ performance was evaluated via the relative root mean squared error (RMSE%). Overall, 241 beetle, 27 bird, and 59 lichen species were recorded. The diversity indices were higher for the multi-taxon community than for the single taxa. They were generally higher in the CVDA site than in GSML, except for the bird community. The highest correlation values between S2 data and biodiversity indices were recorded in CVDA for multi-taxon and beetle communities (| r| = 0.52 and 0.38, respectively), and in GSML for lichen and beetle communities (| r| = 0.34 and 0.26, respectively). RMSE% ranged between 2.53 and 9.99, and between 8.1 and 16.8 for the Simpson and Shannon index, respectively. The most important variables are phase and RMSE of red-Edge bands for bird and lichen communities, while RMSE and time of tassel cap and from EVI indices for beetles and multi-taxon diversity. Our results demonstrate that S2 data can be used for identifying potential biodiversity hotspots, showing that the herein presented harmonic metrics are informative for several taxa inhabiting wood, giving concrete support to cost-effective biodiversity monitoring and nature-based forest management in complex mountain systems.

Crop type classification with combined spectral, texture, and radar features of time-series Sentinel-1 and Sentinel-2 data

ABSTRACT Crop type mapping visualizes the spatial distribution pattern and proportion of planting areas of different crop types, which is the basis for subsequent agricultural applications. Although optical remote sensing has been widely used to monitor crop dynamics, data are not always available due to cloud and other atmospheric effects on optical sensors. Satellite microwave systems such as Synthetic Aperture Radar (SAR) have all-time and all-weather advantages in monitoring ground and crop conditions, combining optical imagery and SAR imagery for crop type classification is of great significance. Our study mainly proposes seven feature combination schemes based on the combination of multi-temporal spectral features and texture features of Sentinel-2 (S2), and radar backscattering features of Sentinel-1 (S1) evaluate the influence of different data sources and different features on classification accuracy, obtains the optimal classification strategy and analyses the contribution of different features to classification result, in the aim of providing a new technical approach for the fine identification of crops from multi-source remote-sensing data. Results show that the crop classification accuracy of combined multi-time series spectral, texture, and radar features is higher than that of combining two types of features. The features subset selected by multi-period spectral, texture, and radar features have the best classification result, the overall accuracy (OA) and kappa coefficients reach 96.40% and 0.93, respectively. The study provides a method reference for future research on larger-scale remote-sensing crop precise extraction.

High-frequency time series comparison of Sentinel-1 and Sentinel-2 satellites for mapping open and vegetated water across the United States (2017–2021)

Frequent observations of surface water at fine spatial scales will provide critical data to support the management of aquatic habitat, flood risk and water quality. Sentinel-1 and Sentinel-2 satellites can provide such observations, but algorithms are still needed that perform well across diverse climate and vegetation conditions. We developed surface inundation algorithms for Sentinel-1 and Sentinel-2, respectively, at 12 sites across the conterminous United States (CONUS), covering a total of >536,000 km2 and representing diverse hydrologic and vegetation landscapes. Each scene in the 5-year (2017–2021) time series was classified into open water, vegetated water, and non-water at 20 m resolution using variables from Sentinel-1 and Sentinel-2, as well as variables derived from topographic and weather datasets. The Sentinel-1 algorithm was developed distinct from the Sentinel-2 model to explore if and where the two time series could potentially be integrated into a single high-frequency time series. Within each model, open water and vegetated water (vegetated palustrine, lacustrine, and riverine wetlands) classes were mapped. The models were validated using imagery from WorldView and PlanetScope. Classification accuracy for open water was high across the 5-year period, with an omission and commission error of only 3.1% and 0.9% for the Sentinel-1 algorithm and 3.1% and 0.5% for the Sentinel-2 algorithm, respectively. Vegetated water accuracy was lower, as expected given that the class represents mixed pixels. The Sentinel-2 algorithm showed higher accuracy (10.7% omission and 7.9% commission error) relative to the Sentinel-1 algorithm (28.4% omission and 16.0% commission error). Patterns over time in the proportion of area mapped as open or vegetated water by the Sentinel-1 and Sentinel-2 algorithms were charted and correlated for a subset of all 12 sites. Our results showed that the Sentinel-1 and Sentinel-2 algorithm open water time series can be integrated at all 12 sites to improve the temporal resolution, but sensor-specific differences, such as sensitivity to vegetation structure versus pixel color, complicate the data integration for mixed-pixel, vegetated water. The methods developed here provide inundation at 5-day (Sentinel-2 algorithm) and 12-day (Sentinel-1 algorithm) time steps to improve our understanding of the short- and long-term response of surface water to climate and land use drivers in different ecoregions.

Modelling of Acid Mine Drainage in Open Pit Lakes Using Sentinel-2 Time-Series: A Case Study from Lusatia, Germany

Strong acid mine drainage (AMD) processes in the flooded, formerly open pits in the Lusatia area present an enormous environmental challenge for the rehabilitation of the post-mining landscape. Extensive and costly monitoring is required for optimal AMD management and remediation planning and control. Because of the large size of the area and the dimension of the problem, the regular sampling can only provide limited point data, which needs to be extrapolated to the entire area. Consequently, the search for effective approaches for extrapolating the point data to the area of all water bodies is essential for rehabilitation success monitoring and for understanding the dependencies between AMD and environmental factors such as land use, weather conditions, geology, and hydrogeology. The main aim of this study was to investigate the suitability of Sentinel-2 multispectral imagery and artificial neural networks (ANNs) for the quantitative mapping of acid mine drainage (AMD) constituents, such as dissolved iron, pH value, and sulfate in large water bodies, for an area of approximately 7220 km2 (the area of the pit lakes is about 185 km2). Correlations between different chemical water parameters were also investigated. An extensive water monitoring dataset was used to train artificial neural networks for the identification of dependencies between the multispectral remote sensing data and the water quality ground measurements. Respective relationships have been identified, especially for dissolved iron and pH. These trained ANNs have been used to produce water quality maps with high spatial (10 × 10 m) and temporal (any cloud-free period) resolution, which show the wide variability of water quality in the different parts of the mining region. Concrete sources of AMD can be identified using the water quality maps of single lakes, and the success of sanitation measures such as liming was visualized. The approach opens many doors for the optimization of both the monitoring program and sanitation technology.

An integrated approach to reconstructing snow cover under clouds and cloud shadows on Sentinel-2 Time-Series images in a mountainous area

Snow is an important type of surface cover and an important freshwater resource in the arid regions of northwest China, and optical satellites are an important means for monitoring variation in snow cover. However, the radiation values measured by satellite sensors are affected by cloud cover and cloud shadows; thus, the reconstruction of snow cover under clouds and cloud shadows becomes a bottleneck problem especially in mountains areas. Based on the digital elevation model (DEM), in this paper, an integrated method of snow cover reconstruction on Sentinel-2 images is created by combining the improved SNOWL (snow line) algorithm considering unstable snow cover areas, Fmask cloud detection algorithms, and the Sen2Cor clear sky snow cover detection tool with a geospatial processing platform, GEE (Google Earth Engine). Considering the Babao River Basin as the study area, 474 Sentinel-2 scenes (10 m) were selected to obtain the snow cover spatiotemporal variations of 158 days from November 2016 to March 2021. Compared with the high-resolution GF-2 satellite images, the integrated method can extract snow under clouds and cloud shadows well, and the improved SNOWL algorithm improves the overall accuracy from 66.05 % to 84.26 %. The experimental results show that approximately half of the Babao River Basin is unstable snow cover, and areas above 3977 m are covered with snow for at least half a year.

Early Identification of Crop Type for Smallholder Farming Systems Using Deep Learning on Time-Series Sentinel-2 Imagery.

Climate change and the COVID-19 pandemic have disrupted the food supply chain across the globe and adversely affected food security. Early estimation of staple crops can assist relevant government agencies to take timely actions for ensuring food security. Reliable crop type maps can play an essential role in monitoring crops, estimating yields, and maintaining smooth food supplies. However, these maps are not available for developing countries until crops have matured and are about to be harvested. The use of remote sensing for accurate crop-type mapping in the first few weeks of sowing remains challenging. Smallholder farming systems and diverse crop types further complicate the challenge. For this study, a ground-based survey is carried out to map fields by recording the coordinates and planted crops in respective fields. The time-series images of the mapped fields are acquired from the Sentinel-2 satellite. A deep learning-based long short-term memory network is used for the accurate mapping of crops at an early growth stage. Results show that staple crops, including rice, wheat, and sugarcane, are classified with 93.77% accuracy as early as the first four weeks of sowing. The proposed method can be applied on a large scale to effectively map crop types for smallholder farms at an early stage, allowing the authorities to plan a seamless availability of food.

Long short term memory deep net performance on fused Planet-Scope and Sentinel-2 imagery for detection of agricultural crop.

In view of the challenges faced by organizations and departments concerned with agricultural capacity observations, we collected In-Situ data consisting of diverse crops (More than 11 consumable vegetation types) in our pilot region of Harichand Charsadda, Khyber Pakhtunkhwa (KP), Pakistan. Our proposed Long Short-Term Memory based Deep Neural network model was trained for land cover land use statistics generation using the acquired ground truth data, for a synergy between Planet-Scope Dove and European Space Agency's Sentinel-2. Total of 4 bands from both sentinel-2 and planet scope including Red, Green, Near-Infrared (NIR) and Normalised Difference Vegetation Index (NDVI) were used for classification purpose. Using short temporal frame of Sentinel-2 comprising 5 date images, we propose an realistic and implementable procedure for generating accurate crop statistics using remote sensing. Our self collected data-set consists of a total number of 107,899 pixels which was further split into 70% and 30% for training and testing purpose of the model respectively. The collected data is in the shape of field parcels, which has been further split for training, validation and test sets, to avoid spatial auto-correlation. To ensure the quality and accuracy 15% of the training data was left out for validation purpose, and 15% for testing. Prediction was also performed on our trained model and visual analysis of the area from the image showed significant results. Further more a comparison between Sentinel-2 time series is performed separately from the fused Planet-Scope and Sentinel-2 time-series data sets. The results achieved shows a weighted average of 93% for Sentinel-2 time series and 97% for fused Planet-Scope and Sentinel-2 time series.

A satellite-based monitoring system for quantifying surface water and mesic vegetation dynamics in a semi-arid region

Semi-arid and arid systems cover one third of the earth’s land surface, and are becoming increasingly drier, but existing datasets do not capture all of the types of water resources that sustain these systems. In semi-arid environments, small surface water bodies and areas of mesic vegetation (wetlands, wet meadows, riparian zones) function as critical water resources. However, the most commonly-used maps of water resources are derived from the Landsat time series or single date aerial photographs, and are too coarse either spatially or temporally to effectively monitor water resource dynamics. In this study, we produced a Sentinel Fusion (SF) water resources product for a semi-arid mountainous region of the western United States, which includes monthly maps of both a) surface water and b) mesic vegetation at 10 m spatial resolution using freely available Earth observation data on an open access platform. We applied random forest classifiers to optical data from the Sentinel-2 time series, synthetic aperture radar (SAR) data from the Sentinel-1 time series, and topographic variables. We compared our SF product with three commonly used and publicly available datasets in the western U.S. We found that our surface water class contained fewer omission errors than a leading global surface water product (94 % producer’s accuracy (PA) vs 84 %) and comparable user’s accuracy (UA) (91 % vs 97 %) with commission errors occurring largely in mixed water pixels. Our mesic vegetation class had up to 43 % higher PAs compared to the National Wetlands Inventory (NWI) estimates and up to 78 % higher UAs over the Sage Grouse Initiative mesic resources maps during the most critical part of the water year. We found that while inclusion of SAR data from the C-band Sentinel-1 sensor consistently improved estimates of water resources in each of the last four months of the 2021 water year when compared to optical-only + topographic variables, only in September did those improvements lie outside of the 95 % confidence interval. With nine times finer spatial resolution and more frequent image collection, our SF maps characterize intra-annual dynamics of smaller water bodies (<30 m wide) and mesic vegetation integral to ecosystem functioning in semi-arid systems compared to leading Landsat-derived products. Further, our workflow is easily reproducible using freely available data on an open access platform, and can be adopted to help guide land use decisions related to water resources by farmers, ranchers, and conservationists in semi-arid environments.

Using dense Sentinel-2 time series to explore combined fire and drought impacts in eucalypt forests

Following one of the driest years on record, millions of hectares of forests in southeast Australia were burned in the 2019–2020 “Black Summer” wildfires. In addition to the areas burned, drought related canopy collapse, dieback and tree mortality was widely observed. In this paper, we present a method to map canopy damage due to drought and fire across a large area. Sentinel-2 satellite imagery was used in a monthly time series to highlight areas of forest where the Normalized Burn Ratio index was significantly below a pre-disturbance “stable” period. The stable period was defined as the 3 years prior to 2019 and the disturbance thresholds are based on bioregion specific standard deviations below pre-disturbance means. The novel methods enabled drought impacted forests to be identified, including those which were subsequently burned by wildfire. Across the 20 Mha of forests studied, 9.9 Mha (49%) fell below the disturbance threshold. Of that, 5.8 Mha was disturbed by fire and a further 4.1 Mha by drought outside of the fire extent. Within the fire extent, almost 0.9 Mha was identified as being significantly drought affected prior to being burned. An analysis of spectral recovery following substantial rainfall from February 2020 onward indicates that most of the areas impacted by both drought and fire have similar rates of recovery to those impacted only by fire. There are some areas, however, where the combined effects of the “double disturbance” appears to be hindering recovery. The methods presented here are easily transferrable and demonstrate an approach for monitoring forest disturbance at higher temporal and spatial scales than those typically used.

A super-ensemble approach to map land cover types with high resolution over data-sparse African savanna landscapes

Accurate and timely land cover products are critical inputs for landscape planning, and provide key information for biodiversity conservation and food security. However, poor mapping quality and low resolution are considerable issues in existing land cover maps over the African savanna, where land use is complex and changing rapidly, and necessary ground-truth data are sparse and hard to obtain. To overcome this problem, to make optimal use of existing maps, and to minimize manual training data collection, we developed a three-stage ensemble method to make land cover maps. In the first stage, we extracted the consensus of multiple existing land cover products to generate fragmented pixel-wise training labels. In the second stage, we translated pixel-wise training labels to image-wise labels using Random Forest (RF) as a “gap-filling model”, with temporal features extracted from Sentinel-1 time series, raw bands, and vegetation indices derived from PlanetScope basemaps. These image-wise labels were scored and edited by humans and the quality information was used in the next stage. For stage three, we trained a U-Net network based upon these image-wise labels, using Sentinel-1 time series and raw bands of PlanetScope basemaps as image features. Using the information on label quality, a quality-weighted loss function was used in the network to reduce the impact of noise in the training labels. Using Northern Tanzania as a case study, the results demonstrate that ensembles of existing land cover maps provide a useful source of data for developing improved land cover maps over hard-to-classify, data-sparse landscapes. The Random Forest “gap-filling model” had an overall accuracy of 80.26% on our independent test dataset with 7 classes. The final U-Net model had an overall accuracy of 83.57%. This approach can be readily applied to other regions and extents (e.g., regional, global) and other data sources (e.g., Sentinel-2).

City-wide, high-resolution mapping of evapotranspiration to guide climate-resilient planning

The impacts of global change, including extreme heat and water scarcity, are threatening an ever-growing urban world population. Evapotranspiration (ET) mitigates the urban heat island, reducing the effect of heat waves. It can also be used as a proxy for vegetation water use, making it a crucial tool to plan resilient green cities. To optimize the trade-off between urban greening and water security, reliable and up-to-date maps of ET for cities are urgently needed. Despite its importance, few studies have mapped urban ET accurately for an entire city in high spatial and temporal resolution. We mapped the ET of Berlin, Germany in high spatial (10-m) and temporal (hourly) resolution for the year of 2019. A novel machine learning (ML) approach combining Sentinel-2 time series, open geodata, and flux footprint modeling was applied. Two eddy flux towers with contrasting surrounding land cover provided the training and testing data. Flux footprint modeling allowed us to incorporate comprehensive land cover types in training the ML models. Open remote sensing and geodata used as model inputs included Normalized Difference Vegetation Index (NDVI) from Sentinel-2, building height, impervious surface fraction, vegetation fraction, and vegetation height. NDVI was used to indicate vegetation phenology and health, as plant transpiration contributes to the majority of terrestrial ET. Hourly reference ET (RET) was calculated and used as input to capture the temporal dynamics of the meteorological conditions. Predictions were carried out using Random Forest (RF) regression. Weighted averages extracted from hourly ET maps using flux footprints were compared to measured ET from the two flux towers. Validation showed that the approach is reliable for mapping urban ET, with a mean R2 of 0.76 and 0.56 and a mean RMSE of 0.0289 mm and 0.0171 mm at the more vegetated site and the city-center site, respectively. Lastly, the variation of ET between Local Climate Zones (LCZs) was analyzed to support urban planning. This study demonstrated the capacity to map urban ET at an unprecedented high spatial and temporal resolution with a novel methodology, which can be used to support the sustainable management of green infrastructure and water resources in an urbanizing world facing climate change.

Early Crop Classification via Multi-Modal Satellite Data Fusion and Temporal Attention

In this article, we propose a deep learning-based algorithm for the classification of crop types from Sentinel-1 and Sentinel-2 time series data which is based on the celebrated transformer architecture. Crucially, we enable our algorithm to do early classification, i.e., predict crop types at arbitrary time points early in the year with a single trained model (progressive intra-season classification). Such early season predictions are of practical relevance for instance for yield forecasts or the modeling of agricultural water balances, therefore being important for the public as well as the private sector. Furthermore, we improve the mechanism of combining different data sources for the prediction task, allowing for both optical and radar data as inputs (multi-modal data fusion) without the need for temporal interpolation. We can demonstrate the effectiveness of our approach on an extensive data set from three federal states of Germany reaching an average F1 score of 0.92 using data of a complete growing season to predict the eight most important crop types and an F1 score above 0.8 when doing early classification at least one month before harvest time. In carefully chosen experiments, we can show that our model generalizes well in time and space.

Tracking changes in coastal land cover in the Yellow Sea, East Asia, using Sentinel-1 and Sentinel-2 time-series images and Google Earth Engine

Coastal zones are essential ecosystems due to their provision of invaluable ecosystem services. However, the geomorphologic characteristics of coastal zones are becoming more complex and changeable due to global warming, sea-level rise (SLR), and the intensification of anthropogenic activities. Therefore, accurate and timely knowledge of coastal land cover types (including tidal flats, coastal vegetation, and year-long water cover) is needed for coastal research and sustainable management. To date, land cover products for coastal areas are mainly derived from moderate resolution imaging spectroradiometer images, but few studies have used Sentinel-1 synthetic aperture radar (S1) and Sentinel-2 Multispectral Instrument (S2) images, which can provide more detailed maps. We developed a Rule-based Time Series Classification (RTSC) approach to map coastal land cover types at a 10 m resolution, combining S1/S2 time-series images (2015–2019) and Google Earth Engine (GEE). These products were developed for the coastal zone of the Yellow Sea (YS), East Asia, which is an essential ecosystem protecting a coastal population of 60 million people from storms and SLR effects. Accuracy assessment showed that the annual maps of coastal land cover had high overall accuracy. The coastal land cover types for the YS in 2019 comprised 3593.42 km2 of tidal flats, 28,506.98 km2 of coastal vegetation, and 5436.92 km2 of coastal year-long water. The interannual dynamics of the coastal land cover area in the YS during 2015–2019 were smaller. This study provides a promising method that combines S1/2 time series, a RTSC approach, and GEE to map coastal land cover areas at large scales. The 10 m resolution maps generated in this study are the most current dataset of coastal land cover types for the YS, and they potentially provide a basis for the sustainable management and conservation of this important coastal zone.

Evaluation of time-series Sentinel-2 images for early estimation of rice yields in south-west of Iran

PurposeIn the past three decades, remote sensing-based models for estimating crop yield have addressed critical problems of general food security, as the unavailability of grains such as rice creates serious worldwide food insecurity problems. The main purpose of this study was to compare the potential of time-series Landsat-8 and Sentinel-2 data to predict rice yield several weeks before harvest on a regional scale.Design/methodology/approachTo this end, the sum of normalized difference vegetation index (NDVI)-based models created the best agreement with actual yield data at the golden time window of six weeks before harvest when rice grains were in milky and mature growth stages. The application of nine other vegetation indicators was also investigated in the golden time window in comparison to NDVI.FindingsThe findings of this study demonstrate the viability of identifying locations with poor and superior performance in terms of production management approaches through a rapid and economical solution for early rice grain yield assessment. Results indicated that while some of those, such as enhanced vegetation index (EVI) and optimized soil adjusted vegetation index, were able to estimate rice yield with high accuracy, NDVI is still the best indicator to predict rice yield before harvest. However, experiments can be conducted in different regions in future studies to evaluate the generalizability of the approach.Originality/valueTo achieve this objective, the authors considered the following purposes: using Sentinel-2 time-series data, determining the appropriate growth stage for estimating rice yield and evaluating different vegetation indices for estimating rice yield.