Proof of concept to detect alluvial gold mining with Sentinel-1 CSAR

Mineral extraction is an important part of the economy in the Democratic Republic of Congo (DRC). Artisanal and small-scale alluvial gold mining has been a source of livelihood, but as it is mostly operated informally, it is also a driver of illicit activities, child labor, conflict financing, human abuse and of course environmental degradation like water contamination, deforestation and erosion.

As part of its cooperation with the DRC, the German Federal Institute for Geosciences and Natural Resources (Bundesanstalt für Geowissenschaften und Rohstoffe (BGR)) therefore wishes to improve the control and monitoring of mining activities in DRC.

One approach for better monitoring in cloud persistent regions is the use of satellite synthetic aperture radar, i.e. Sentinel1 CSAR data from the European Copernicus program. In this study, a 30x30 km2 region north of the town Bunia in the Ituri Province in DRC has been monitored using the whole Sentinel-1 time-series from 2017 to 2022.

Alluvial gold mining practice comes with clearing of vegetation and deforestation, excavating soil and washing the soil which causes flooding of the area and the development of water pools. This results in a strong reduction in radar backscatter, specifically in VH polarization.

VH backscatter γ°(VH) is a function of volume scattering which reflects mainly the vegetation. VH backscatter is high over forest, very low over bare soil and even lower over water. New developed alluvial gold mining areas can therefore be detected by thresholding the difference in backscatter Δγ°(VH) between two time periods t₀ and t₁.

Instead of using single day scenes, speckle noise was reduced by producing quarter-yearly averaged mosaics of the whole time series, resulting in a time series of gold mining detection on a three-month basis from 2018-2022. The time series are then compared to very high resolution Planet Labs time series and validated with a set of 31 ground observations.

Varying the threshold of the VH backscatter decrease between 3dB and 5 dB shows that lower thresholds lead to an earlier detection as they already detect the first process of vegetation loss whereas higher threshold mainly detect the newly formed water pools at a later stage of gold mining activities, though with less false detections.

The optimal choice of parameters (time interval, integration time, and threshold) is relevant for the overall purpose of the monitoring, e.g. near-real time detection, manual selection of sites to visit by authorities, overall estimation of disturbed forest.