Abstract
Selective logging often marks the beginning of forest degradation, but remote sensing alert systems typically detect it long after it starts. This type of disturbance produces a subtle canopy signal not easily recognised in the early stages, causing alerts in some areas of the Brazilian Legal Amazon to lag behind the first visible signs by several months. This study explores whether a geospatial foundation model can help reduce this time gap. For each documented disturbance site, we compile Harmonized Landsat and Sentinel-2 (HLS) image time series spanning October 2024 to May 2026 into a datacube and use it to pretrain, without labels, a compact spatiotemporal masked autoencoder (ST-MAE). This model learns to embed the dynamics of forest degradation into latent-space representations, the so-called "embeddings". Following this, a lightweight downstream model — trained with few labelled samples — interprets ST-MAE's embeddings to evaluate every new HLS observation for recent logging activity. To effectively translate the evaluation scores into production-ready alerts, we apply a simple change-detection approach: an edge filter for sudden logging-level changes, paired with a persistence filter for sustained detections. We conducted a leave-one-out cross-validation assessment across 20 well-documented logging sites: yielding no false alarms prior to the first disturbance evidence, the change detector confirmed the onset of the 20 logging activities with a median delay of 15 days after the first post-disturbance image. The downstream model's scores remain low during the pre-disturbance period but increase once logging begins, confirming that it responds to logging-related events rather than to other landscape dynamics or calendar artefacts. This label-free pretraining enables effective performance with minimal annotations, providing a practical solution to reduce the temporal gap in monitoring forest degradation.
Showing the abstract — retrieve the full paper via the Exa API.