International audience

Recent advances in deep learning architectures, driven by Artificial Intelligence (AI), have significantly enhanced time-series forecasting accuracy. However, the inherent lack of interpretability and causal reasoning in black-box models limits their adoption in high-impact applications such as traffic management and urban mobility. To address this challenge, we propose$\mathrm{C}^{2}$-STORM (Causal Counterfactual Spatio-Temporal TransfORmer Model), a novel framework that integrates deep learning to achieve both accurate and interpretable predictions. Our approach uses frontdoor-adjusted causal extraction to reduce confounding bias, grounded in structural causal principles. This is coupled with a Dynamic Causality Generator that learns context-aware Spatio-Temporal (ST) dependencies via graphstructured attention, enhancing node-specific feature prioritization. Furthermore, we develop a transformer-based counterfactual reasoning module that simulates actionable “what-if” scenarios, aligning interventions with domain-specific logic. Experiments conducted on real-world datasets (PeMSD7, LondonBike) demonstrate the effectiveness of$\mathrm{C}^{2}$-STORM, achieving state-of-the-art performance with improvements of 4.2- 6.6% in MAE/MAPE over baseline methods. Ablation studies confirm the significance of each component, while visual explanations, including causal graphs and feature importance maps, enhance the interpretability and transparency of predictions. By bridging causal inference with ST learning, C²-STORM represents a significant advancement in trustworthy forecasting for complex systems.

Effectively characterize the current land cover status is crucial for assessing agricultural production, monitoring natural resources, and making informed land management decisions. Satellite Image Time Series (SITS) data, capturing spatio-temporal information, are the primary source of information to support the general task of Land Use Land Cover (LULC) mapping. With the aim to effectively exploit the spatio-temporal information carried out by SITS data for the underlying task of land cover mapping, here we introduce a multi-scale classification framework that combines together pixel-level and object-level multivariate SITS information with the objective to ameliorate the pixel-level time series analysis. To assess the behaviour of the proposed framework, we provide a comparative analysis with several SITS-based land cover mapping strategies on a challenging study area, namely Koumbia, located in the Burkina Faso. The obtained results reveal that the joint use of pixel-level and object-level information clearly ameliorates the classification performances.

This study aims to assess the severity of desertification in Tunisia by using machine learning and deep learning models to analyze remote sensing data. By employing MODIS ecological indicators such as Modified Soil-Adjusted Vegetation Index (MSAVI), Normalized Difference Vegetation Index (NDVI), Albedo and Topsoil Grain Size Index (TGSI), different models including eXtreme Gradient Boosting (XGBoost), Random Forest (RF), Decision Tree (DT), LGBM, Convolutional Neural Networks (CNN), Variational Autoencoder (VAE) and Recurrent Neural Networks (RNN) will be trained and compared to effectively monitor desertification trends in Tunisia from 2016 to 2022. The evaluation using metrics such as accuracy, F1 score, recall, and precision showed that XG-Boost performed best with an accuracy of 92.35%. The study also identified a negative trend of desertification affecting the entire country. This research highlights the importance of using computer vision to monitor and address the challenges of desertification and suggests potential applications for environmental management.

Accurate forecasting of climate and vegetation changes through SDG 13 initiatives is critical for adapting agricultural practices, as it allows farmers to anticipate changes in growing conditions and vegetation patterns, ensuring long-term food security. In this paper, a data science hybrid pipeline is proposed to analyse and forecast the impact of climate change on vegetation using remote sensing data. The proposed pipeline follows a three-step process. In the first step, different data (i.e. normalized difference vegetation index (NDVI), standardized precipitation index and land cover) are collected and pre-processed. Subsequently, the second step involves feature extraction by analysing NDVI and SPI data based on the wavelet transform method and analysing the change detection using land cover information. Finally, the third step completes the process by exploring the extracted feature based on the wavelet transform to forecast the vegetation change based on the temporal convolution network (TCN) model. This pipeline performs and validates various analyses related to vegetation health, including trend component extraction, the timing of maximum vegetation index, response of vegetation changes to meteorological variables and forecasting vegetation change based on climatic data. These results, obtained for Tunisia, Turkey, and Spain, will empower decision-makers to improve agricultural practices.

The United Nations (UN) 2030 Agenda involves 17 major Sustainable Development Goals (SDGs). These SDGs have great implications for country-wide development and making plans in both developed and developing nations in the post-2015 period to 2030. The SDGs are a set of 17 interlinked global goals designed to be a plan to attain a better sustainable future by the combination of earth observation (EO) and artificial intelligence architecture. To attain the goal of global sustainable protection and utilization of terrestrial ecosystems, it is important to quantitatively determine the implementation of Sustainable development goal 15 (SDG-15) and goal 13 (SDG-13). In this paper, we focus on the integration of these SDGs in Tunisia as a new regional development plan. Thus, we present a complete benchmark that aims to solve the complicated analytical problems related to the sophisticated data type using Deep learning (DL) architecture.

Desertification detection is a challenging problem because of the dynamic climate change and human activity. This paper proposes a methodology to detect regions with desertification risk based on Landsat imagery (RGB bands and NDVI) and variational auto-Encoder (VAE). The considered features (RGB and NDVI) are extracted from multitemporal Landsat optical images taken from the freely available Landsat program from 2001 to 2020. The arid region around Jeffara in Medenine (Tunisia) is selected as a study area. The VAE model was evaluated and compared with two deep learning models: convolutional neural network (CNN), convolutional recurrent neural network (CNN_RNN), and VAE without NDVI. The comparative results showed that the VAE outperformed the other models for desertification detection, with an accuracy of over 98 %.

Soil moisture (SM) is an important parameter used to control a broad range of environmental applications. An increasing attention has been recently given to machine learning (ML) methods for SM retrieval that provide promising performance. Nevertheless, most of them are based on a supervised learning strategies that depend on the used labeled training samples. In fact, they are unaffordable or costly. In this letter, new teacher–student for SM estimation, called (TS-SME), relying on teacher–student (TS) framework using synthetic aperture radar (SAR) and optical data, was proposed to estimate SM. The main advantage of this framework is to enroll a large amount of unlabeled data together with a small amount of labeled data. Experiments were carried out on two arid areas in southern Tunisia. The input data include the backscatter coefficient in two-mode polarization ($\sigma ^{\circ }_{\textrm {VV}}$and$\sigma ^{\circ }_{\textrm {VH}}$) for Sentinel-1A, normalized difference vegetation index (NDVI) and normalized difference infrared index (NDII) for Sentinel-2A andin situmeasurements. Extensive experimental results demonstrated that TS-SME framework is capable of generating a well-performed student model, with the estimation accuracy is superior to all teacher models [artificial neural network (ANN), eXtreme gradient boosting (XGBoost), random forest regressor (RFR), and water cloud model (WCM)]. It was highly correlated with thein situmeasurements with high Pearson’s correlation coefficient$R$(${R}_{\textrm {RF}} =0.86$,${R}_{\textrm {ANN}} =0.75$,${R}_{\textrm {XGBoost}} =0.77$,${R}_{\textrm {WCM}} =0.77$,${R}_{{\,\,\textrm {TS-SME}}} =0.96$) and low root mean square error (RMSE) ($\textrm {RMSE}_{\textrm {RF}} =1.09$%,$\textrm {RMSE}_{\textrm {ANN}} =1.49$%,$\textrm {RMSE}_{\textrm {XGBoost}} =1.39$%,$\textrm {RMSE}_{\textrm {WCM}} =1.12$%,$\textrm {RMSE}_{\,\,\textrm {TS-SME{} }} =0.8$%), respectively.

Soil Moisture (SM) is an important parameter for many environmental applications. In recent years, different Machine Learning (ML) methods have been developed in order to estimate the SM. The aim of this paper is to evaluate three ML methods based methodology to estimate the SM, namely Artificial Neural Networks (ANN), Random Forest Regression (RFR) and extreme Gradient Boosting (XGBoost). The used input data for the three ML methods include the in situ measurements SM, the Backscatter coefficient in two modes polarizations (σ°yy and σ°YH) from Sentinel-1A and Normalized Difference Vegetation Index (NDVI) from Sentinel-2A data. The comparative results show that the SM obtained using the ANN, RF and XG Boost algorithms are highly correlated to the in-situ SM, with high Pearson's correlation coefficient R (RRF= 0.86, RANN= 0.75, R XGBoost= 0.77) respectively and low RMSE (RMSERF= 1.09%, RMSEANN= 1.49%, RMSExGBoost= 1.39%) respectively.

Soil Moisture (SM) estimation is of growing interest. In the recent years, different Machine Learning (ML) methods were developed in order to better understand the spatio-temporal variability of SM. Among the different ML methods, neural networks were the most used for SM estimation. The purpose of this paper is to propose a Long Short-Term Memory (LSTM) based methodology to estimate the SM. The input data used in the LSTM model include the Soil Moisture Active and Passive mission (SMAP) Brightness Temperature (TB), the Moderate Resolution Imaging Spectroradiometer Vegetation Water Content (MODIS-VWC) and the soil temperature. The target SM data used to train the LSTM model is provided by the Real-time In situ Soil Monitoring for Agriculture (RISMA) network installed by Agriculture and Agri-Food Canada (AAFC). LSTM shows good ability to estimate the SM values with good accuracy.

Nowadays, Normalized Difference Vegetation Index (NDVI) time-series has been successfully used in research regarding global environmental change. NDVI time series have proven to be a useful means of indicating drought-related vegetation conditions, due to their real-time coverage across the globe at relatively high spatial resolution. In this paper, we propose a method for detecting rare events in NDVI series. These events are particularly rare and infrequent which increases their complexity of detecting and analyzing them. The proposed method is based on the analysis of the random component by Jarque-Bera test to verify the presence of rare events and obtain their features (time and amplitude). For validation, we have used a database for regions in Northwestern of Tunisia. These data come from MODIS for a period from 18 February 2000 to 17 November 2013 at a spatial resolution of 250 m by 250m.

Nowadays, the process of change detection is regarded as an outstanding way for urban planning and design. The major concern of this paper is to investigate the non-stationary character of multi-temporal time series. To overcome this problem, we propose an adaptive multiplicative decomposition of non-stationary multi-temporal satellite image, which allows to decompose the series into three components: trend, seasonal and random, to properly model the evolution of land cover. We carried several experiments to validate our approach based on Landsat images covering the region of “Tres Cantos-Madrid” in Spain. The obtained results show the effectiveness of our proposed method comparing to some conventional methods.