Conference Agenda

Deforestation is a critical environmental challenge, driving greenhouse gas emissions, biodiversity loss, and disruption of hydrological and energy cycles. Monitoring these processes in Near Real-Time (NRT) is essential, particularly in tropical regions where rapid land-use changes threaten ecosystems and climate regulation. Optical monitoring has traditionally been the main approach, but persistent cloud cover in the tropics often limits its effectiveness. Synthetic Aperture Radar (SAR), unaffected by clouds or illumination, has emerged as a reliable alternative, with missions like Sentinel-1 offering free global acquisitions suitable for operational monitoring. Sentinel-1 provides C-band dual-polarization data (VV and VH), yet most operational systems exploit only one channel or process both independently before merging outputs. This strategy underutilizes the complementarity of the two polarizations and can lead to omission errors, when both channels are required to confirm a change, or commission errors, when a false alarm in one channel is enough to trigger detection. A methodological gap therefore remains in fully exploiting Sentinel-1 polarimetry for early forest loss detection while maintaining operational feasibility.

To address this limitation, the Bayesian polarimetric detector pol-BOCD has been developed to jointly process VV and VH Sentinel-1 time series for NRT deforestation monitoring. The method extends Bayesian Online Changepoint Detection (BOCD) to the polarimetric domain by explicitly modeling the statistical uncorrelation between co- and cross-polarized channels in natural environments. The joint data likelihood is factorized into two univariate distributions, allowing the independent contributions of each channel to be integrated within a single hidden Markov chain, without increasing computational complexity beyond that of the univariate BOCD.

The approach is evaluated in two tropical biomes with heterogeneous land cover and undergoing intense forest conversion: the Cerrado and the Brazilian Amazon. The Cerrado is a biologically rich savanna facing extensive agricultural expansion and exhibiting strong seasonality that complicates SAR-based monitoring. The Amazon rainforest, by contrast, is increasingly affected by small, fragmented clearings that are ecologically disruptive yet difficult to detect. Validation relies on the MapBiomas Alerta reference dataset, covering 8,000 ha in the Cerrado and 13,400 ha in the Amazon in 2020, and including small clearings (0.1–2 ha) as well as larger ones (3 to 50 ha).

Results demonstrate that pol-BOCD improves the detection of small-scale clearings in the Cerrado and large-scale clearings in the Amazon by approximately 10% compared to VH-only BOCD, which in turn outperforms VV-only BOCD (-23% compared to pol-BOCD). In contrast, improvements are marginal for large clearings in the Cerrado and small ones in the Amazon. These outcomes are attributed to the distinct deforestation practices in the two biomes. In the Cerrado, large-scale disturbances—often associated with mechanized soy cultivation—are executed systematically, leaving behind uniformly bare soil. In such conditions, VH polarization is more effective due to the sharp reduction in volume scattering, rendering VV polarization comparatively less informative. In contrast, smallholder agriculture and subsistence farming typically involve artisanal clearing methods, which often leave substantial residual vegetation. Under these less uniform conditions, VV polarization contributes more significantly, complementing VH in detecting subtle structural changes due to its greater sensitivity to ground scattering. In the Amazon, an opposite pattern takes place: large clearings typically involve cutting down the forest, allowing it to dry, and then burning the biomass, as removing extensive vegetation from dense rainforest areas is logistically challenging. A further comparison, shows that pol-BOCD outperforms both the union and intersection of single-polarization results, which suffer from high false alarms and low true detections, respectively. Against operational methods, pol-BOCD surpasses GLAD-L in the Cerrado (+47%) and RADD in the Amazon (+34% for small clearings), benefiting from the higher spatial resolution of the Sentinel-1 input data due to the omission of speckle filtering during pre-processing.

Overall, the findings highlight the potential of joint Sentinel-1 polarimetric processing in pol-BOCD to enhance NRT deforestation detection, particularly in areas where residual vegetation introduces textural complexity in radar signals, underscoring its usefulness for operational monitoring across diverse biomes.