Neural network-based detection of tropical moist forest disturbances using standardised Sentinel-1 time series

Details

Supervisors

Defourny, Pierre ; Delhez, Baptiste

Faculty

Faculté des bioingénieurs

Degree label

Master [120] : bioingénieur en sciences et technologies de l'environnement, à finalité spécialisée

Abstract

enThe transition from the 20th to the 21st century has been marked by an increasing recognition of the impact of human activities on the planet and its ecosystems. Among the global challenges related to the achievement of sustainable development goals, the conservation and systematic monitoring of tropical forests remain critical. Anthropogenic changes in forest cover significantly compromise the provision of essential ecosystem services in these regions. This research focuses on the development of Neural Networks for the near-real time detection (and monitoring) of tropical forest disturbances caused by smallholder agriculture in the Congo Basin, based on dense standardised Sentinel-1 C-SAR time series. Two regions of interest are defined for this study: Dekese serves as the calibration region of interest for the developed neural networks, while Likati is used to assess model transferability and to perform temporal validation. The results highlight a clear trade-off between spatial precision and detection sensitivity, underscoring the difficulty in simultaneously achieving accurate localisation and comprehensive disturbance identification. The approach enabled timely detection of disturbances, allowing the generation of near-real-time forest disturbance alerts that are crucial for effectively monitoring short-duration events, such as those associated with smallholder agriculture. Such alert products can serve as valuable tools to promote more effective forest governance and long-term tropical forest ecosystem monitoring, ultimately contributing to the achievement of related sustainable development goals.