Deep learning approach to detect charcoal kilns in medium resolution satellite imagery: a step towards disentangling the drivers of deforestation in Mozambique

Abstract

The need for a better understanding of the drivers of deforestation and forest degradation is made clear in UN REDD+ policy. Charcoal production is thought to be one of the main drivers of degradation across Sub-Saharan Africa (SSA).

However it is particularly difficult to detect due to the fact it is usually small scale and results in forest degradation. Consequently the spread and intensity of the industry is poorly understood. We put forward a methodology to autonomously identify charcoal kilns in Southern Mozambique using a state of the art deep learning algorithm. We took a pre-trained Faster R-CNN Inception v2 coco model and fine-tuned it with charcoal kilns from Sentinel-2 10m NIR imagery. The parameters were optimised to fit the unique detection task. The resulting accuracy assessment produced an f1-score of 0.76 with a precision of 82.5% and a recall of 70.2%. This is a similar accuracy to other modified Faster R-CNN small object detectors. In our study the model performance was significantly improved by ensuring the algorithm produced enough proposal regions to detect the very high density of kilns within each tile. Furthermore adjusting the size and scale of these proposal boxes (anchors) increased the f1-score by 0.05. Although the results are specific to the case study it proves that there is potential to extrapolating the technique across SSA. This could help forest managers monitor illegal activity in near real time, as well as better inform policy such as UN REDD+.