Detecting Change due to Alluvial Gold Mining in Peruvian Rainforest Using Recursive Convolutional Neural Networks and Contrastive Learning

Alluvial gold mining has led to the removal of over 10,000 km2 of tropical rainforest and the creation of hundreds of mining lagoons in the Peruvian department of Madre de Dios. Some of this mining activity is illegal and understanding the dynamics of land-use change surrounding this activity will help monitor such activity and benefit conservation and natural resource planning efforts by the regional and federal government. New remote sensing platforms and computer vision methods allow for high-resolution evaluation of change that will provide benefits for land-cover mapping. Here we detail our work with two novel techniques. First, we utilize E-ReCNN, a new recurrent convolutional neural network model designed to detect annual changes in mining lagoons generated by alluvial gold mining. Results of tests of this model using 3, 6, and 10-channel Sentinel-2 data show accuracies of (Kappa: 0.92 (± 0.04), Jaccard: 0.88 (± 0.07), F1: 0.88 (± 0.05)). Second, we show how we can utilize information contained within PlanetScope high-resolution daily imagery to classify land cover more accurately with Sentinel-2 data as an outcome of contrastive learning, a method which uses self-supervised learning methods across multiple augmentations of imagery of the same location. The results of this work showcase the value of specific channels (RGB, RGB+SWIR1+SWIR2+NIR) for improving classification accuracy. Altogether, our work reveals new methods for detecting land cover change in both aquatic and terrestrial elements of a tropical landscape which can be used to help document the timeline of anthropogenic disturbance.