We first develop a generalized self-supervised multi-channel despeckling method for Sentinel-1. The method exploits the TOPSAR acquisition mode, in which images are acquired in overlapping bursts. We show that the two observations from a burst overlap are independent, thanks to the difference in squint angle between them. This lets us apply the Noise2Noise framework and train a network directly on real data, without any simulated or clean reference image. By operating on the covariance matrix, the method jointly estimates all parameters of interest, including the correlation between channels. The filter outperforms the state of the art on both simulated and real data, while better preserving fine textures.

We then study the value of these parameters for forest monitoring. The C-band interferometric phase carries little height information. The coherence, however, correlates with forest height in simple conditions, namely flat terrain and a non-dense canopy. Integrating it into Kanop’s height estimation pipeline over the Landes forest improves the metrics, but the gain is too small relative to the computational cost to retain in production.

We finally apply the filter to deforestation detection with the Bayesian Online Changepoint Detection (BOCD) algorithm. BOCD was designed to operate on unfiltered time series, since classical despeckling reduces speckle at the cost of resolution. Our filter instead suppresses speckle while preserving resolution. Filtering the time series before feeding it to BOCD consistently improves detection, for both single- and dual-polarization variants.