Satellite Image Time Series Research Articles

Improving crop type mapping by integrating LSTM with temporal random masking and pixel-set spatial information

Accurate and timely crop type classification is essential for effective agricultural monitoring, cropland management, and yield estimation. Unfortunately, the complicated temporal patterns of different crops, combined with gaps and noise in satellite observations caused by clouds and rain, restrict crop classification accuracy, particularly during early seasons with limited temporal information. Although deep learning-based methods have exhibited great potential for improving crop type mapping, insufficient and noisy training data may lead them to overlook more generalizable features and derive inferior classification performance. To address these challenges, we developed a Mask Pixel-set SpatioTemporal Integration Network (Mask-PSTIN), which integrates a temporal random masking technique and a novel PSTIN model. Temporal random masking augments the training data by selectively removing certain temporal information to improve data variability, enforcing the model to learn more generalized features. The PSTIN, comprising a pixel-set aggregation encoder (PSAE) and long short-term memory (LSTM) module, effectively captures comprehensive spatiotemporal features from time-series satellite images. The effectiveness of Mask-PSTIN was evaluated across three regions with different landscapes and cropping systems. Results demonstrated that the addition of PSAE in PSTIN significantly improved crop classification accuracy compared to a basic LSTM, with average overall accuracy (OA) increasing from 80.9% to 83.9%, and the mean F1-Score (mF1) rising from 0.781 to 0.818. Incorporating temporal random masking in training led to further improvements, increasing average OA and mF1 to 87.4% and 0.865, respectively. The Mask-PSTIN significantly outperformed traditional machine learning and deep learning methods (i.e., RF, SVM, Transformer, and CNN-LSTM) in crop type mapping across all three regions. Furthermore, Mask-PSTIN enabled earlier and more accurate crop type identification before or during their developing stages compared to machine learning models. Feature importance analysis based on the gradient backpropagation algorithm revealed that Mask-PSTIN effectively leveraged multi-temporal features, exhibiting broader attention across various time steps and capturing critical crop phenological characteristics. These results suggest that Mask-PSTIN is a promising approach for improving both post-harvest and early-season crop type classification, with potential applications in agricultural management and monitoring.

Morphological changes of river Bank Erosion and channel shifting assessment on Arial Khan River of Bangladesh using Landsat satellite time series images

The continuous processes of erosion and deposition in rivers are intricately linked to changes in land use and land cover (LULC), reflecting dynamic landscape transformations. This study investigates the rate of riverbank erosion, channel shifting, and the impact of erosion on LULC dynamics along the Arial Khan River. Utilizing GIS and remote sensing techniques, we analyzed time-series satellite images from 1993 to 2024 within a 2 km buffer zone of the floodplain. Supervised classification methods were applied for LULC analysis, with results validated using confusion matrices to assess classification accuracy. ArcGIS 10.8 was employed to analyze channel centerline migration, while NDVI analysis assessed vegetation health. The study reveals that between 1993 and 2024, 5142.60 ha were lost to erosion, while deposition affected 4756.52 ha, with erosion surpassing deposition overall. The river shifted an average of 0.66 km during this period, with higher migration rates in Muladi Upazila compared to Babuganj Upazila. Settlement areas expanded significantly from 137.82 ha in 1993 to 2768.68 ha in 2024, while agricultural land declined substantially. These findings provide valuable insights for policymakers to address erosion severity in confluence zones and lay the groundwork for sustainable land management practices.

Mapping Crop Types for Beekeepers Using Sentinel-2 Satellite Image Time Series: Five Essential Crops in the Pollination Services

Driven by the widespread adoption of deep learning (DL) in crop mapping with satellite image time series (SITS), this study was motivated by the recent success of temporal attention-based approaches in crop mapping. To meet the needs of beekeepers, this study aimed to develop DL-based classification models for mapping five essential crops in pollination services in Quebec province, Canada, by using Sentinel-2 SITS. Due to the challenging task of crop mapping using SITS, this study employed three DL-based models, namely one-dimensional temporal convolutional neural networks (CNNs) (1DTempCNNs), one-dimensional spectral CNNs (1DSpecCNNs), and long short-term memory (LSTM). Accordingly, this study aimed to capture expert-free temporal and spectral features, specifically targeting temporal features using 1DTempCNN and LSTM models, and spectral features using the 1DSpecCNN model. Our findings indicated that the LSTM model (macro-averaged recall of 0.80, precision of 0.80, F1-score of 0.80, and ROC of 0.89) outperformed both 1DTempCNNs (macro-averaged recall of 0.73, precision of 0.74, F1-score of 0.73, and ROC of 0.85) and 1DSpecCNNs (macro-averaged recall of 0.78, precision of 0.77, F1-score of 0.77, and ROC of 0.88) models, underscoring its effectiveness in capturing temporal features and highlighting its suitability for crop mapping using Sentinel-2 SITS. Furthermore, applying one-dimensional convolution (Conv1D) across the spectral domain demonstrated greater potential in distinguishing land covers and crop types than applying it across the temporal domain. This study contributes to providing insights into the capabilities and limitations of various DL-based classification models for crop mapping using Sentinel-2 SITS.

Assessment of land restoration from optical satellite image time-series

Abstract. Ecosystem conservation and restoration is recognized by the international community as a key strategy for human well-being as well as our planet’s health. Assessing the efficiency of actions implemented in this context is essential, as their application on the ground may prove challenging. In the present paper, we present a methodology for characterising vegetation regeneration. This methodology has been developed and tested at the scale of the Sao Paulo state. It is based on two successive steps. Firstly, annual land cover maps are produced with breakpoints detection and Random Forest classification. This process is performed on biophysical variables derived from more than 10 000 optical images (2016 – 2021 included) and 4774 reference data. Secondly, we applied an expert-based rules algorithm to derive land regeneration map from annual land cover time series. The distinction is made between natural and anthropogenic regeneration. The accuracy assessment shows an overall accuracy of more than 80% for both the annual land cover maps and the regeneration map.

Temporal ViT-U-Net Tandem Model: Enhancing Multi-Sensor Land Cover Classification Through Transformer-Based Utilization of Satellite Image Time Series

Abstract. Semantic segmentation is essential in the field of remote sensing because it is used for various applications such as environmental monitoring and land cover classification. Recent advancements aim to collectively classify data from diverse sensors and epochs to improve predictive accuracy. With the availability of vast Satellite Image Time Series (SITS) data, supervised deep learning methods, such as Transformer models, become viable options. This paper introduces the Temporal Vision Transformer(ViT), designed to extract features from SITS. These features, capturing the temporal patterns of land cover classes, are integrated with features derived from aerial imagery to improve land cover classification. Drawing inspiration from the success of transformers in Natural language processing (NLP), Temporal ViT concurrently extracts spatial and temporal information from SITS data using tailored positional encoding strategies. The proposed approach fosters comprehensive feature learning across both domains, facilitating seamless integration of encoded data from SITS into aerial images. Furthermore, a training strategy is proposed that supports the Temporal ViT to focus on classes with a changing appearance over the year. Extensive experiments carried out in this work indicate the enhanced classification performance of Temporal ViT compared to existing state-of-the-art techniques for multi-modal land cover classification. Our model achieves a 3.8% increase in the mean IoU compared to the network solely relying on aerial images.

Time Series approach to map areas of Agricultural Plastic Waste generation

Abstract. The escalating presence of plastics in agriculture has raised concerns regarding Agricultural Plastic Waste (APW). Hitherto, a lack of comprehensive plasticulture data impedes effective waste management strategies, potentially resulting in plastic pollution and contributing to microplastic formation. APW locations and quantities are pivotal for territorial planning and the formulation of public policies on waste management and land use. Remote detection of agri-plastics has garnered increased consideration, particularly in mapping plastic-mulched farmlands (PMFs) dispersed in extensive regions. This study investigates whether the use of satellite image time series can map PMFs accurately. We assessed pixel-based classification using a 16-day composite time series of Sentinel-2 imagery (S2-16D) obtained from the Brazil Data Cube project. The Plastic-Mulched Landcover Index (PMLI) was joined into S2-16D bands because it is an index focused on PMF detection. Four classifiers (RF, MLP, L-TAE, and TCNN) were compared through agreed classification metrics. The most promising outcome showed an overall accuracy of 100% employing L-TAE but with visible noise in the map. The time series enhanced accuracy while minimizing background confusion, offering a viable solution for PMF detection. The PMF map presented herein represents an initial stride toward fostering circularity in plasticulture throughout South America.

Spatio-temporal Analysis of Agroforestry Systems in Hotan Using Multi-source Remote Sensing and Deep Learning

In the Hotan region of Xinjiang, where arable land is scarce, an agroforestry system integrating walnut trees with crops has been implemented to maximize land-use efficiency. While this dense planting enhances land use, it also limits the availability of light to the understory crops, potentially impacting their yield and quality. To address this issue and enhance the system's sustainability and productivity, precise delineation of the planting structure is critical. This study proposes a novel framework that leverages multi-source remote sensing data combined with advanced deep learning techniques to analyze the agroforestry planting structure. The methodological approach consists of three key phases. First, an instance segmentation model was employed to extract farmland parcels from high-resolution imagery, providing a basis for vegetation classification. Next, a time series model using irregular satellite image time series (irSITS) tracked the growth dynamics of the vegetation. Finally, the spatial planting structure of the walnut trees was quantified using the D-LinkNet model, integrated with a template filling algorithm.The results demonstrated a classification accuracy of 97.85% in extracting parcel-level planting structures, identifying 42,955 farmland parcels, including 21,153 intercropped parcels. The temporal and spatial characteristics of the agroforestry system were then analyzed, leading to a grading of canopy cover and walnut tree density within the intercropped areas. This comprehensive spatiotemporal planting structure offers a valuable foundation for informed local agricultural policy adjustments. In conclusion, this approach advances the understanding of complex agroforestry systems and provides a robust scientific basis for optimizing intercropping practices, contributing to sustainable agricultural development in arid regions.

Improving cross-regional model transfer performance in crop classification by crop time series correction

ABSTRACT Cross-Regional Model Transfer (CRMT) provides a solution to crop classification challenges in target regions with limited labeled samples. However, when the source region (source domain) and the target region (target domain) are spatially distant, a shift in crop Satellite Image Time Series (SITS) occurs, resulting in suboptimal performance of SITS-based classification models in the target domain. To address this issue, we propose a novel domain adaptation method, named SITS Correction. This method mitigates the SITS shift by introducing correction factors to improve the classification performance of CRMT. The correction factors consist of source domain correction units and target domain correction units, which are from variables associated with the SITS shift. In cases where labeled samples are available in the target domain, the SITS have calculated Barycenter (i.e. SitsB) as the correction units, which then combine the source domain and the target domain’s SitsB to form a SITS-based correction factor. When no labeled samples are available, the correction units are substituted by Barycenter of the temperature variables time series (i.e. temperature indicator) influencing crop growth, and the correction factors are composed of temperature indicators from both the source and target domains. We validate SITS Correction through transfer experiments conducted in four regions, encompassing various years, crops, and classification methods. Experimental results demonstrate a significant improvement in crop classification performance with SITS Correction in both scenarios – target domain with labeled samples (accuracy increased by 39.1%) and target domain without labeled samples (accuracy increased by 14.8%). This research provides a practical option for crop classification in target regions with either limited or no labeled samples.

Enhanced Monitoring of Sub-Seasonal Land Use Dynamics in Vietnam’s Mekong Delta through Quantile Mapping and Harmonic Regression

In response to economic and environmental challenges like sea-level rise, salinity intrusion, groundwater extraction, sand mining, and sinking delta phenomena, the demand for solutions to adapt to changing conditions in riverine environments has increased significantly. High-quality analyses of land use and land cover (LULC) dynamics play a critical role in addressing these challenges. This study introduces a novel high-spatial resolution satellite-based approach to identify sub-seasonal LULC dynamics in the Mekong River Delta (MRD), employing a three-year (2021–2023) Sentinel-1 and Sentinel-2 satellite data time series. The primary obstacle is discerning detailed vegetation dynamics, particularly the seasonality of rice crops, answered through quantile mapping, harmonic regression with Fourier transform, and phenological metrics as inputs to a random forest machine learning classifier. Due to the substantial data volume, Google’s cloud computing platform Earth Engine was utilized for the analysis. Furthermore, the study evaluated the relative significance of various input features. The overall accuracy of the classification is 82.6% with a kappa statistic of 0.81, determined using comprehensive reference data collected in Vietnam. While the purely pixel-based approach has limitations, it proves to be a viable method for high-spatial resolution satellite image time series classification of the MRD.

Self-Supervised Learning across the Spectrum

Satellite image time series (SITS) segmentation is crucial for many applications, like environmental monitoring, land cover mapping, and agricultural crop type classification. However, training models for SITS segmentation remains a challenging task due to the lack of abundant training data, which requires fine-grained annotation. We propose S4, a new self-supervised pretraining approach that significantly reduces the requirement for labeled training data by utilizing two key insights of satellite imagery: (a) Satellites capture images in different parts of the spectrum, such as radio frequencies and visible frequencies. (b) Satellite imagery is geo-registered, allowing for fine-grained spatial alignment. We use these insights to formulate pretraining tasks in S4. To the best of our knowledge, S4 is the first multimodal and temporal approach for SITS segmentation. S4’s novelty stems from leveraging multiple properties required for SITS self-supervision: (1) multiple modalities, (2) temporal information, and (3) pixel-level feature extraction. We also curate m2s2-SITS, a large-scale dataset of unlabeled, spatially aligned, multimodal, and geographic-specific SITS that serves as representative pretraining data for S4. Finally, we evaluate S4 on multiple SITS segmentation datasets and demonstrate its efficacy against competing baselines while using limited labeled data. Through a series of extensive comparisons and ablation studies, we demonstrate S4’s ability as an effective feature extractor for downstream semantic segmentation.

Morphological and Functional Polycentric Urbanization in Colombo Metropolitan of Sri Lanka Using Time-Series Satellite Images from 1988–2022

Examining urban growth in metropolitan areas from both morphological and functional perspectives is paramount for sustainable development in developing and developed countries. As a collaborative and critical approach, polycentrism operates concurrently at local, regional, and national levels. However, the decentralized morphologies and functional attributes of polycentric urban spaces are yet to be adequately explored, despite their significant influence on sustainable urban planning and management. Colombo, the only metropolitan area in Sri Lanka, has experienced considerable yet ambiguously imbalanced spatial growth due to insufficient data. This study employed time-series satellite imagery from 1988 to 2022, in conjunction with a network model, to elucidate Colombo’s uneven urban patterns, while achieving an accuracy of over 80% throughout the period. The results revealed that the metropolis has evolved into a polynuclear city, complicating the identification of centers and their corresponding developmental stages. Nevertheless, the study successfully pinpointed the morphological and functional Grama Niladhari divisions, disclosing the morphological features of centralized dispersion and the functional traits of decentralized concentration. These insights provide significant implications for local government, guiding urban planning initiatives to achieve sustainable development goals.

Improving rapid flood impact assessment: An enhanced multi-sensor approach including a new flood mapping method based on Sentinel-2 data

Rapid flood impact assessment methods need complete and accurate flood maps to provide reliable information for disaster risk management, in particular for emergency response and recovery and reconstruction plans. With the aim of improving the rapid assessment of flood impacts, this work presents a new impact assessment method characterized by an enhanced satellite multi-sensor approach for flood mapping, which improves the characterization of the hazard. This includes a novel flood mapping method based on the new multi-temporal Modified Normalized Difference Water Index (MNDWI) that uses multi-temporal statistics computed on time-series of Sentinel-2 multi-spectral satellite images. The multi-temporal aspect of the MNDWI improves characterization of land cover over time and enhances the temporary flooded areas, which can be extracted through a thresholding technique, allowing the delineation of more precise and complete flood maps. The methodology, if implemented in cloud-based environments such as Google Earth Engine (GEE), is computationally light and robust, allowing the derivation of flood maps in matters of minutes, also for large areas.The flood mapping and impact assessment method has been applied to the seasonal flood occurred in South Sudan in 2020, using Sentinel-1, Sentinel-2 and PlanetScope satellite imagery. Flood impacts were assessed considering damages to buildings, roads, and cropland. The multi-sensor approach estimated an impact of 57.4 million USD (considering a middle-bound scenario), higher than what estimated by using Sentinel-1 data only, and Sentinel-2 data only (respectively 24% and 78% of the estimation resulting from the multi-sensor approach). This work highlights the effectiveness and importance of considering multi-source satellite data for flood mapping in a context of disaster risk management, to better inform disaster response, recovery and reconstruction plans.

Towards Optimising the Derivation of Phenological Phases of Different Crop Types over Germany Using Satellite Image Time Series

Operational crop monitoring applications, including crop type mapping, condition monitoring, and yield estimation, would benefit from the ability to robustly detect and map crop phenology measures related to the crop calendar and management activities like emergence, stem elongation, and harvest timing. However, this has proven to be challenging due to two main issues: first, the lack of optimised approaches for accurate crop phenology retrievals, and second, the cloud cover during the crop growth period, which hampers the use of optical data. Hence, in the current study, we outline a novel calibration procedure that optimises the settings to produce high-quality NDVI time series as well as the thresholds for retrieving the start of the season (SOS) and end of the season (EOS) of different crops, making them more comparable and related to ground crop phenological measures. As a first step, we introduce a new method, termed UE-WS, to reconstruct high-quality NDVI time series data by integrating a robust upper envelope detection technique with the Whittaker smoothing filter. The experimental results demonstrate that the new method can achieve satisfactory performance in reducing noise in the original NDVI time series and producing high-quality NDVI profiles. As a second step, a threshold optimisation approach was carried out for each phenophase of three crops (winter wheat, corn, and sugarbeet) using an optimisation framework, primarily leveraging the state-of-the-art hyperparameter optimization method (Optuna) by first narrowing down the search space for the threshold parameter and then applying a grid search to pinpoint the optimal value within this refined range. This process focused on minimising the error between the satellite-derived and observed days of the year (DOY) based on data from the German Meteorological Service (DWD) covering two years (2019–2020) and three federal states in Germany. The results of the calculation of the median of the temporal difference between the DOY observations of DWD phenology held out from a separate year (2021) and those derived from satellite data reveal that it typically ranged within ±10 days for almost all phenological phases. The validation results of the detection of dates of phenological phases against separate field-based phenological observations resulted in an RMSE of less than 10 days and an R-squared value of approximately 0.9 or greater. The findings demonstrate how optimising the thresholds required for deriving crop-specific phenophases using high-quality NDVI time series data could produce timely and spatially explicit phenological information at the field and crop levels.

Nrt: operational monitoring of satellite image time-series in Python

nrt: operational monitoring of satellite image time-series in Python

Monopile-induced turbulence and sediment redistribution form visible wakes in offshore wind farms

Offshore wind farms are becoming an increasingly common feature in the marine environment as a renewable energy source. There is a growing body of evidence on the effects of wind farms on the seabed and its organisms. However, an important and understudied aspect of site development is the interaction of turbine foundations on the surrounding marine environment. Structures exert significant disturbance on tides, waves and currents; these are visible as optically-distinct, elongate wakes at the sea surface with elevated suspended particulate matter. Despite this, there is uncertainty on the mechanisms that lead to the visible manifestation of wakes at turbine foundations, primarily due to a lack of direct measurements. Here, in situ measurements along with a 15-year time series of satellite images of the Thanet offshore wind farm, located within the Thames Estuary, were used to investigate the formation of visible monopile wakes, and the effects these have on the surrounding water column. We show the optically distinct wakes are near-constant at Thanet; visible in >90% of all satellite images, yet no regional change in sea surface turbidity could be attributed to wind farm construction or operation. Monopile wake in situ water samples and acoustic Doppler current profiler (ADCP) backscatter measurements demonstrated colour change related to elevated sea surface sediment concentration. However, averaged water column measurements of suspended sediment within wakes, and upstream of monopiles, remained consistent. These measurements demonstrate that sediment was redistributed towards surface waters, rather than additional sediment becoming suspended in the wake. ADCP velocity measurements supported a mechanism of sediment lofting towards the surface, with enhanced vertically upwards flow recorded in wakes.

Transformer models for Land Cover Classification with Satellite Image Time Series

In this paper we address the task of pixel-wise land cover (LC) classification using satellite image time series (SITS). For that purpose, we use a supervised deep learning model and focus on combining spatial and temporal features. Our method is based on the Swin Transformer and captures global temporal features by using self-attention and local spatial features by convolutions. We extend the architecture to receive multi-temporal input to generate one output label map for every input image. In our experiments we focus on the application of pixel-wise LC classification from Sentinel‑2 SITS over the whole area of Lower Saxony (Germany). The experiments with our new model show that by using convolutions for spatial feature extraction or a temporal weighting module in the skip connections the performance improves and is more stable. The combined usage of both adaptations results in the overall best performance although this improvement is only minimal. Compared to a fully convolutional neural network without any self-attention layers our model improves the results by 2.1% in the mean F1-Score on a corrected test dataset. Additionally, we investigate different types of temporal position encoding, which do not have a significant impact on the performance.

Enhanced calendar support for temporal datacube queries

AbstractA significant share of our data are timeseries. Time can occur as the single dimension, or further (such as spatial) dimensions can be present as well, leading to multidimensional data such as 3D x/y/t satellite image timeseries or 4D x/y/z/t atmospheric simulations like weather forecasts. However, no user would like to count “seconds since epoch,” rather we all prefer calendar notation. Libraries, SQL, etc. support calendar arithmetics, yet more advanced functionality is needed. Instead of demoting advanced calendar functionality to low‐level coding, it seems desirable to have semantic support for calendars at high‐level, declarative query level. We present a conceptual framework for temporal modeling which allows the expression of a wide range of temporal queries. It smoothly enhances existing calendar handling by adding two new data and query description parameters, period of validity, and granularity. While our work is embedded in spatio‐temporal geo services, it can be incorporated in a non‐breaking manner in any environment needing calendar addressing. The concepts are implemented in an Array DBMS operational on several Petabytes of Earth datacubes.

Application of Remote Sensing for Identifying Soil Erosion Processes on a Regional Scale: An Innovative Approach to Enhance the Erosion Potential Model

Soil erosion represents a complex ecological issue that is present on a global level, with negative consequences for environmental quality, the conservation and availability of natural resources, population safety, and material security, both in rural and urban areas. To mitigate the harmful effects of soil erosion, a soil erosion map can be created. Broadly applied in the Balkan Peninsula region (Serbia, Bosnia and Herzegovina, Croatia, Slovenia, Montenegro, North Macedonia, Romania, Bulgaria, and Greece), the Erosion Potential Method (EPM) is an empirical erosion model that is widely applied in the process of creating soil erosion maps. In this study, an innovation in the process of the identification and mapping of erosion processes was made, creating a coefficient of the types and extent of erosion and slumps (φ), representing one of the most sensitive parameters in the EPM. The process of creating the coefficient (φ) consisted of applying remote sensing methods and satellite images from a Landsat mission. The research area for which the satellite images were obtained and thematic maps of erosion processes (coefficient φ) were created is the area of the Federation of Bosnia and Herzegovina and the Brčko District (situated in Bosnia and Herzegovina). The Google Earth Engine (GEE) platform was employed to process and retrieve Landsat 7 Enhanced Thematic Mapper plus (ETM+) and Landsat 8 Operational Land Imager and Thermal Infrared Sensor (OLI/TIRS) satellite imagery over a period of ten years (from 1 January 2010 to 31 December 2020). The mapping and identification of erosion processes were performed based on the Bare Soil Index (BSI) and by applying the equation for fractional bare soil cover. The spatial–temporal distribution of fractional bare soil cover enabled the definition of coefficient (φ) values in the field. An accuracy assessment was conducted based on 190 reference samples from the field using a confusion matrix, overall accuracy (OA), user accuracy (UA), producer accuracy (PA), and the Kappa statistic. Using the confusion matrix, an OA of 85.79% was obtained, while UA ranged from 33% to 100%, and PA ranged from 50% to 100%. Applying the Kappa statistic, an accuracy of 0.82 was obtained, indicating a high level of accuracy. The availability of a time series of multispectral satellite images for each month is a crucial element in monitoring the occurrence of erosion processes of various types (surface, mixed, and deep) in the field. Additionally, it contributes significantly to decision-making, strategies, and plans in the domain of erosion control work, the development of plans for identifying erosion-prone areas, plans for defense against torrential floods, and the creation of soil erosion maps at local, regional, and national levels.

Forecasting water resources from satellite image time series using a graph-based learning strategy

Abstract. In the context of climate change, it is important to monitor the dynamics of the Earth’s surface in order to prevent extreme weather phenomena such as floods and droughts. To this end, global meteorological forecasting is constantly being improved, with a recent breakthrough in deep learning methods. In this paper, we propose to adapt a recent weather forecasting architecture, called GraphCast, to a water resources forecasting task using high-resolution satellite image time series (SITS). Based on an intermediate mesh, the data geometry used within the network is adapted to match high spatial resolution data acquired in two-dimensional space. In particular, we introduce a predefined irregular mesh based on a segmentation map to guide the network’s predictions and bring more detail to specific areas. We conduct experiments to forecast water resources index two months ahead on lakes and rivers in Italy and Spain. We demonstrate that our adaptation of GraphCast outperforms the existing frameworks designed for SITS analysis. It also showed stable results for the main hyperparameter, i.e., the number of superpixels. We conclude that adapting global meteorological forecasting methods to SITS settings can be beneficial for high spatial resolution predictions.

Mapping drivers of tropical forest loss with satellite image time series and machine learning

The rates of tropical deforestation remain high, resulting in carbon emissions, biodiversity loss, and impacts on local communities. To design effective policies to tackle this, it is necessary to know what the drivers behind deforestation are. Since drivers vary in space and time, producing accurate spatially explicit maps with regular temporal updates is essential. Drivers can be recognized from satellite imagery but the scale of tropical deforestation makes it unfeasible to do so manually. Machine learning opens up possibilities for automating and scaling up this process. In this study, we developed and trained a deep learning model to classify the drivers of any forest loss—including deforestation—from satellite image time series. Our model architecture allows understanding of how the input time series is used to make a prediction, showing the model learns different patterns for recognizing each driver and highlighting the need for temporal data. We used our model to classify over 588′000 sites to produce a map detailing the drivers behind tropical forest loss. The results confirm that the majority of it is driven by agriculture, but also show significant regional differences. Such data is a crucial source of information to enable targeting specific drivers locally and can be updated in the future using free satellite data.