The use of Markov Random Field for tropical deforestation detection in dense C-band SAR time-series

By Jorn van der Ent (the Netherlands)

Abstract
Tropical forest have been increasingly threatened and stakeholders need accurate information to act. This thesis used MRF deforestation detection in C-band SAR data. With MRF, spatial and temporal context was used to improve classification accuracy. The SAR datawas used for a first rough classification, after which a spatial, temporal or spatial-temporal MRF was applied to improve the classification. This resulted in an estimated user’s accuracy of 63%, an producer’s accuracy of 75% and an overall accuracy of 97%. The improvement of spatial classification accuracy had a trade-off with the temporal accuracy as the mean time-lag increased to 8 days on average. Data acquired during 2015 and 2016 over central Sumatra has been used for a proof-of-concept demonstration showing deforestation rates of 7.1% and 7.4% respectively.