Change Detection of Small Water Bodies in Alluvial Gold Mining Satellite Imagery

Monitoring change detection of the land surface is one of the critical concepts in the mean of natural resource management, deforestation, and conservation. One of the most causes of land degradation, deforestation, and water pollution is artisanal and small-scale mining (ASM) which is basically related to water bodies. To monitor the water bodies that are potential ASM ponds and detect the changes satellite imagery is commonly used. For fine-scale change detection, satellite images can be challenging because of atmospheric conditions on spectral data and representing individual objects with connecting pixels.

In this project, we investigate the change detection in the context of a current conservation challenge, artisanal-scale gold mining (ASGM) performance of two machine learning approaches (supervised and semi-supervised learning) on Sentinel-2 images. We obtained Sentinel-2 imagery and created an open- source labeled (binary and multiclass: ‘increase, decrease, water existence/absence’) land-cover change dataset for the Madre de Dios (MDD) region in Peru, a hotspot of ASGM activity, as well as active ASGM areas in other countries (Venezuela, Indonesia, and Myanmar). As supervised learning, we used E- ReCNN model and for semi-supervised learning, we used the SVM-STV approach with histogram matching preprocessing steps on multiple channels and additional Lab color space improved performance and reduce the influence of atmospheric effects.