Intra-annual double disturbances (sequential disturbances occurring within a single year, such as slash-and-burn) in tropical regions result in more significant ecosystem impacts than single disturbances. However, optical data-based monitoring is limited by insufficient continuous spatio-temporal observations. Approaches using higher-frequency Sentinel-1 (S1) data, relying solely on temporal or spatial features, struggle to distinguish double disturbances due to noise and temporal heterogeneity. To address this issue, we propose an approach that employs a three-dimensional deep learning architecture to simultaneously extract multi-level embedded spatial context and temporal change characteristics—such as magnitude and duration features from time series S1 data to differentiate single and double disturbances. Additionally, we tested the effectiveness of incorporating an attention mechanism into the model to highlight key spatio-temporal features and suppress irrelevant information. The optimal model was applied to map double disturbances in three deforestation hotspot states in the Brazilian Amazon for 2019, identifying the first disturbance time at a monthly scale and verifying slash-and-burn by overlaying burned area maps. Results show that (1) the 3D U-Net model effectively extracted the spatio-temporal information of doubles disturbance in S1 data (F1 score of 0.912), and the integration of the attention mechanism highlighted distinctive temporal features, improving the balanced accuracy by 2.7%; (2) The optimal model (i.e., 3D U-Net with attention mechanism) exhibited potential generalization capability when mapping double disturbances in three states for 2019, with user’s accuracy and producer’s accuracy exceeding 0.85, and the estimated double disturbances area in three states in 2019 were 2685 ± 189 km 2 ; (3) 53.57% of mapped forest disturbances were double disturbances, 94.4% of which occurred outside the protected areas, highlighting the critical role of protected areas in ecological preservation. This study presents a novel S1-based framework for accurately characterizing intra-annual slash-and-burn events on a finer-grained spatial and temporal scale, providing useful information for carbon emission estimation and the development of effective conservation interventions.
Sun et al. (Mon,) studied this question.