Deep learning and automatic reference label harvesting for Sentinel-1 SAR-based rapid tropical dry forest disturbance mapping

Abstract
The advent of temporally dense radar data such as the Sentinel-1 SAR have opened the door for rapid forest disturbance detection in the humid tropics. Tropical dry forest disturbance detection, however, were challenged by seasonality and more open canopy characteristics. In this manuscript, we proposed a Sentinel-1 SAR and deep learning based rapid forest disturbance detection approach for tropical dry forests. We demonstrated a weakly supervised method for reference label harvesting based on medium resolution globally available forest and forest disturbance maps. We trained a deep neural network model to derive forest and forest disturbance probabilities from Sentinel-1 images in the first step. We then implemented a probabilistic disturbance detection and refinement method to map forest disturbances in near real-time in two test regions in Paraguay and Mozambique. We mapped new forest disturbances in an emulated near real-time scenario for 2020 and 2021 and evaluated the spatial accuracy of the disturbance alerts by generating area adjusted precision, recall and F-1 score. We also evaluated the improvement in timeliness of disturbance detection by estimating mean time difference of disturbance events detection with that of Landsat-based GLAD alerts. The generated alerts in Paraguay and Mozambique achieved a precision, recall and F-1 score of 0.99, 0.61, 0.75 and 0.97, 0.51, 0.66, respectively. The proposed method detected disturbances with a mean of 21 days ( 18 days) earlier in Paraguay and 18 days ( 18 days) earlier in Mozambique than the Landsat-based GLAD alerts. These results demonstrated the efficacy of the proposed method and its viability to be used in an operational setting to generate large area rapid near real-time disturbance alerts in the dry tropics.

Keywords: Dry forests; Machine learning; Sentinel-1 SAR