A Sentinel-2 based multispectral convolutional neural network for detecting artisanal small-scale mining in Ghana: Applying deep learning to shallow mining

Artisanal Small-scale Mining (ASM) is a critical source of livelihoods for large areas of the Global South but it can bring with it many problems, including deforestation, water pollution and low worker safety. Timely and comprehensive management of ASM is crucial to ensure that it can take place safely and cleanly, supporting sustainable development. The informal nature of the sector presents challenges related to documenting the locations of ASM. Remote sensing methods have been used to detect ASM, although difficulties with accuracy, resolution and persistent cloud cover have been encountered. This paper proposes a method of ASM detection using a deep convolutional neural network model applied to open source Sentinel-2 multispectral satellite imagery. Firstly, the model is evaluated against both existing ASM detection methods and visual inspection of randomly sampled points. Secondly, the model is used to map mining and urban land use changes over a dataset spanning four years and 6 million hectares of southern Ghana, demonstrating the ability of this method to process very large areas. The omission and commission errors of less than 8% from the sampled points indicate that this model has achieved unprecedented levels of accuracy for the task of detecting ASM from satellite imagery. When applied to the case study area, the data on ASM trends over time demonstrate a correlation between the Ghanaian government's 2017 clampdown and ASM activities. The ASM land use category decreased by 6000 ha in 2017, despite a net increase of 15000 ha over the period 2015–2019. Additionally, the model was applied to quantify the extent of illegal mining related deforestation within Ghana's protected forests, measured at over 3500 ha, with 2400 of these lost since 2015. The results demonstrate that this methodology can detect ASM in Ghana with a high degree of accuracy at a minimal cost in terms of financial and human resources. The model shows strong generalisation abilities, offering exciting potential for using this methodology to further monitor and analyse ASM related land use changes worldwide.