Using Sentinel-2 imagery and an M-RCNN to detect canopy and deforestation

Abstract

Continued deforestation is a global problem with major consequences for climate change. Several UN-REDD-recognized canopy disturbance methods are available, all of which currently rely on numerous pre-processing steps before an output is created, making them time-consuming methods of surveillance in situations where timely intervention to halt deforestation is essential. This study trains an M-RCNN using Sentinal-2 imagery with transfer learning to produce high-resolution deforestation. Monitoring which can be frequently and rapidly updated. By testing different hyperparameters, an M-RCNN tool was developed to produce optimal imagery change detection within the confines of the available computational capacity. This study has shown that, with a more powerful GPU and by building on training techniques, machine learning is a reliable and straightforward object and change recognition tool which could be used for rapid detection of deforestation. Future developments for this study would be to train the model using SAR so that cloud mitigation techniques are not required.