Leveraging remote sensing and machine learning to detect cocoa-driven deforestation in landscape-specific contexts

Details

Supervisors

Defourny, Pierre ; Delhez, Baptiste

Faculty

Faculté des bioingénieurs

Degree label

Master [120] en sciences agronomiques et industries du vivant, à finalité spécialisée: Geo-Information Science and Earth Observation for Environmental Modelling and Management

Abstract

Deforestation, the alarming loss of forests, has become a global issue, with millions of hectares of forest lost each year, which has implications for ecosystems, biodiversity, and climate change. Ghana contributes to the global deforestation rates mainly due to cocoa expansion, mining activities, urbanization, logging and others like them. Cocoa expansion is on the rise due to the economic benefits farmers derive from it, as a result it is driving the loss of forests which thrive well in the same landscape. The European Union Deforestation Regulation aims to reduce the rate of cocoa-driven deforestation by ensuring that farmers are able to prove that their cocoa pods are not from a deforested area. This study aims to evaluate the performance of remote sensing approaches and machine learning techniques in detecting cocoa-driven deforestation in landscape-specific contexts in Ghana. It compares Sentinel 2, Planet NICFI and a Pan-Sharpened Sentinel 2 image to understand which best maps the landscape. It also looks into a change detection analysis to identify patterns of cocoa-driven deforestation. Based on the best model for LULC mapping of the landscape, it is evaluated in a different area to evaluate its performance. The overall accuracy of the Pan-sharpened Sentinel 2 image was 96.5%. This was selected as the best in mapping the cocoa-forest mosaic landscape. There were 431.77 ha of open forests that have been converted to cocoa between 2022 and 2023. The overall accuracy of the model when tested in a different landscape was 88.5%. This approach is efficient in detecting cocoa-driven deforestation in Ghana.