Speaker
Description
Timely and accurate monitoring of deforestation is critical for managing tropical forests, yet individual Earth observation sensors have inherent limitations. Multispectral imagery provides rich spectral information on vegetation properties but is often hindered by cloud cover and limited canopy penetration. Synthetic Aperture Radar (SAR) complements optical data by being largely insensitive to atmospheric conditions and by providing insights into vegetation structure, but SAR-based monitoring can be affected by moisture variability, reduced contrast when residual vegetation remains after clearing, and the fading of deforestation signatures over time. The complementary nature of multispectral and SAR data has led to growing interest in multi-source fusion approaches. However, most existing methods rely on offline processing or decision-level fusion strategies, which do not fully exploit the temporal and spectral synergies between the two sources. Consequently, no operational forest loss monitoring system currently integrates both modalities within a unified framework.
In this context, the Bayesian Online Change Point Detection (BOCD) algorithm has been generalized to handle asynchronous, irregularly sampled, multi-source time series by incorporating a dynamically weighted fusion scheme. Specifically, a source’s relevance depends on how recently it was observed, gradually diminishing as the time since its last observation increases, following a physical decay model. The resulting method, named ms-BOCD, enables sensor-agnostic, interpretable, and Near Real-Time (NRT) detection of deforestation.
The method is applied to Sentinel-1 SAR and Sentinel-2 multispectral time series and evaluated for deforestation detection over sampled areas in the Brazilian Amazon, covering a total of 8,000 hectares. The reference data for validation are provided by MapBiomas Alerta and include clearings of various sizes, i.e., 0.1–1, 1–5, 5–20, and >20 hectares. Additionally, the method is compared against existing approaches, including VH-BOCD (BOCD using only Sentinel-1 cross-polarization), RADD, and GFW (Global Forest Watch), which aggregates alerts from multiple independent operational systems.
Results highlight an overall 26% improvement in detection performance of ms-BOCD compared to VH-BOCD and RADD, along with near-zero false alarms. Gains are particularly pronounced for larger deforested areas (+27% for >20 ha vs. +7% for 0.1–1 ha), which often correspond to parcels with residual vegetation remaining after clearing. This pattern is due to the slash-and-burn deforestation practice typical of the Amazon, where cut trees are left to dry before being burned at the end of the dry season, as removing large quantities of biomass is difficult in densely forested areas. Particularly, residual vegetation poses challenges for C-band SAR data, as it can cause volume scattering saturation, especially in cross-polarization channels, making multispectral data especially useful in these cases. Compared to ms-BOCD, GFW shows a 13% increase in false alarms, likely due to the aggregation of errors from independent alerting systems. Finally, temporal analysis indicates that ms-BOCD achieves faster detection than RADD while maintaining performance comparable to GFW.
Overall, ms-BOCD provides accurate, interpretable, and NRT detection of deforestation, demonstrating the potential of multi-source Bayesian fusion for operational tropical forest monitoring.