Accurate time series forecasting is fundamental to geohazard early warning, yet remains a major challenge. Conventional in situ geotechnical monitoring remains costly and spatially constrained, whereas deep learning applied to remote sensing data has become increasingly prevalent but often suffers from opacity of model decision-making. To address this issue, we propose a Transformer-based forecasting framework, namely PatchTST-Fusion, adapted for multivariate Multi-Temporal Interferometric Synthetic Aperture Radar (MT-InSAR) time series. The framework integrates model interpretability analysis through TimeSHAP, providing temporal and feature-level attributions across the input sequence. Landslide deformation time series are first derived from Copernicus Sentinel-1 SAR data. Variational Mode Decomposition is then applied to decompose the non-linear signals into trend, seasonal, and noise components. The denoised displacement series are modeled and forecast using the proposed PatchTST-Fusion, which incorporates rainfall and reservoir water level fluctuations as feature-level drivers. Application to the Daping landslide cluster in the Three Gorges Reservoir Area in China demonstrates that our method captures both the long-term and transient non-linear coupling between deformation and its triggers, surpassing state-of-the-art models including CNN-BiGRU-Attention, Informer and original PatchTST with 7–55% improvements in MAE and 10–52% improvements in RMSE. Beyond predictive gains, feature attribution of environmental triggers via TimeSHAP reveals that rainfall and reservoir regulation exert temporally distinct influences on slope kinematics, with high relative importance concentrated in specific periods and characteristic lagged responses. This interpretable framework provides both enhanced forecasting accuracy and process-based insights, offering a broadly applicable tool for landslide early warning in reservoir regions.

Forests are home to an estimated 80% of the world’s terrestrial biodiversity. They play a crucial role in mitigating climate change by absorbing and sequestering carbon dioxide. Given their importance, it is essential to monitor and understand their structure on a global scale. Synthetic Aperture Radar (SAR) Tomography (TomoSAR) allows to create a 3D image of the forest vertical profile. Compressive Sensing (CS)-based algorithms are considered as the state of the art in TomoSAR reconstruction, thanks to their super-resolution and extra robustness capabilities [1]. However, they also come with a great computational burden as the sparse reconstruction is achieved through iterative solvers. Recently, a deep neural network γ-net [2], was proposed, using the Learned Iterative Shrinkage Thresholding Algorithm (LISTA) as introduced in [3]. LISTA unrolls the original ISTA as a RNN structure and was further extended to the complex domain (CV-LISTA), to solve TomoSAR inversion. γ-net was evaluated on urban areas and showed no significant degradation in super-resolving power compared to state-of-the-art solvers, while reducing the computational burden by 2 or 3 orders of magnitude In this work, we adapt the γ-net approach to TomoSAR of forest areas. Since the backscattering profile of forest in SAR data shows a strong volumetric component making it non-sparse, it is unsuitable for direct CS reconstruction. Therefore, this work will study different transformations to represent the continuous backscattering profile of forest with a set of sparse coefficients after projecting to a certain orthogonal basis. We will experiment Wavelet transforms which have been employed in previous model-based CS TomoSAR reconstruction [4]. Experiments on learning the sparse transformation directly from data will also be conducted and compared to Wavelet transforms. The overall performance of the proposed algorithm will be evaluated based on its super-resolution power and its efficiency using simulated data. Experiments on real data in tropical and boreal forests will also be performed, in order to compare our solution with traditional baseline methods, such as Capon Beamforming, as well as state-of-the-art model-based methods. References [1] X. X. Zhu and R. Bamler, “Tomographic SAR Inversion by L1-Norm Regularization—The Compressive Sensing Approach”, IEEE Transactions on Geoscience and Remote Sensing, vol. 48, no. 10, pp. 3839–3846, 2010. [2] K. Qian, Y. Wang, Y. Shi, and X. X. Zhu, “Y-Net: Superresolving SAR Tomographic Inversion via Deep Learning”, IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–16, 2022. [3] K. Gregor and Y. LeCun, “Learning Fast Approximations of Sparse Coding”, in Proceedings of the 27th International Conference on International Conference on Machine Learning, ICML’10, (Madison, WI, USA), p. 399–406, Omnipress, 2010. [4] E. Aguilera, M. Nannini, and A. Reigber, “Wavelet-Based Compressed Sensing for SAR Tomography of Forested Areas”, IEEE Transactions on Geoscience and Remote Sensing, vol. 51, no. 12, pp. 5283–5295, 2013.

To achieve super-resolution synthetic aperture radar (SAR) tomography (TomoSAR), compressive sensing (CS)-based algorithms are usually employed, which are, however, computationally expensive, and thus is not often applied in large-scale processing. Recently, deep unfolding techniques have provided a good combination of physical model-based algorithms and the ability of neural networks to learn from data. In this vein, iterative CS-based algorithms can usually be un-rolled as neural networks with only 10 to 20 layers. When trained, it shows great computational efficiency for further TomoSAR processing. However, the learning architecture of neural networks built in this approach tends to result in error propagation and information loss, thus degrading the performance. In this paper, we propose to employ complex-valued sparse long short-term memory (CV-SLSTM) units to tackle this problem by incorporating historically updating information into the optimization procedure and preserving full information. Simulations are carried out to validate the performance of the proposed algorithm.