Near-Real-Time African Tropical Forest Loss Monitoring with Sentinel-1 data: From Threshold-Based to Deep Learning Detection

Tropical forest monitoring and mapping are in the center of ecosystem preservation of the 21st century. Satellite remote sensing has become an essential tool to assess forest degradation in wide and remote areas at regional scale. Nevertheless, annual assessments based on optical imagery tend to show their limits in the African tropical forests, where the semi-persistent cloud cover prevents the proper detection of short-duration and small-scale events. The 10-year-old Sentinel-1 C-band SAR dataset provides now a proper archive to meet the challenge of continuous forest monitoring, with one acquisition every 12 days at 10-m ground sampling distance in all tropical regions. Current early-warning systems have proven the asset of SAR data to detect forest loss events. Most of the time, a loss detection is triggered at every time step, when a determined threshold identified as outlier is exceeded. Those inflexible thresholds are established empirically from trends observed on previous stability periods and forest loss samples, and the detection can require a mid-term period to be confirmed. We introduce a novel detection system using temporally standardized Sentinel-1 C-band SAR data to efficiently monitor and map forest loss in the Congo Basin using machine learning instead of systematic empirical thresholds. The system processes 3 successive acquisitions (covering a period of 24 days) and assesses forest degradation every 12 days in a temporal sliding window. Training and calibration are performed in Dekese, DRC (3°30' S), and the model is applied and validated in Likati, DRC (3°29' N). The distance of 800 km separating the two study areas ensures the potential of the model for generalization across the tropical moist forest of the Congo Basin without additional training. Three key steps have been implemented to leverage machine learning for SAR-based tropical forest monitoring and achieve automated, continuous deep learning detection: 1. Multitemporal training with machine learning using random forest (RF) algorithm: Training samples are created and weighted between classes (stable or loss), and a model is developed using multiple time periods. The trained model is then transferred to a distinct remote site, demonstrating its potential for cross-regional generalization. 2. Neural network implementation with a multi-layer perceptron (MLP): The Random Forest model is converted into an MLP, with calibration of layers to leverage the advantages of a neural network over traditional machine learning techniques. The performance improvement in detecting forest loss is assessed. 3. Refinement in deep learning through a convolutional neural network (CNN): A U-Net architecture is developed to eliminate the need for spatial filtering and empirical thresholds. The CNN is designed to dynamically adapt to local contexts and pixel-level variability, ensuring more robust and flexible detection capabilities to complex patterns. Ultimately, this study supports the operationalization of automated forest monitoring systems in the Congo Basin, enabling early detection of forest loss events within a 24-day sliding temporal window.