Mapping selective logging with Sentinel-1 image time series and a novel transformer-based model

Abstract

Selective logging (SL) is a major driver of forest degradation, yet its small and short-lived canopy openings remain difficult to detect in persistently cloudy tropical regions. Optical approaches miss many events, while synthetic aperture radar (SAR) time series are challenging to model with architectures that either fuse multivariate features too early or smooth over transient signals. We present InterFormer, a Transformer-based architecture that jointly models feature interactions and temporal dynamics via a feature block with Multi-head Intercross Self Attention (MISA) and a temporal block with Multi-head Self Dissimilar Attention (MSDA). Texture features were further utilized as complementary input. Since precise SL labels are unavailable, the model was trained and internally tested using a weak-label benchmark, in which pixel labels were extracted within the official published Real-Time Deforestation Detection System (DETER) polygons. To complement this internal weak-label benchmark, we designed two independent practical experiments to evaluate the model performance in low-intensity Annual Production Units (UPAs) based on field-inventory data and at the municipality scale ($>$ 30,000 km²) in the Brazilian Amazon. In the independent UPA validation, InterFormer localized 92.65% of tree-falling gaps in UPA areas and outperformed standard Transformer variants and Planet-based alerts from the Brazilian Forest Service by approximately 20% at the tree level. At the municipality level, our predicted alert polygons achieved the highest Detection Rate against DETER across multiple reference-overlap thresholds in both Paragominas and Feliz Natal, while also producing fewer unmatched alerts than RAdar for Detecting Deforestation (RADD) and Tropical Moist Forest dataset (TMF). However, substantial over-expansion of predicted alert regions remained. These results suggest that InterFormer provides a useful framework for representing weak and short-lived SL signals in Sentinel-1 time series, and that openly available SAR data have the potential to support large-scale SL monitoring under weak supervision. Nevertheless, the current municipality-scale workflow still relies on optical-derived masking as a supporting step, and the development of a fully SAR-only large-area monitoring framework remains an important direction for future work.