Trial of Detection Accuracies Improvement for JJ-FAST Deforestation Detection Algorithm Using Deep Learning

JICA-JAXA Forest Early Warning System in the Tropics (JJ-FAST) monitors tropical forest in 77 countries by using PALSAR-2/ScanSAR time-series data. The deforestation detection algorithm used in JJ-FAST shows overall user's accuracies of 71.1%, but the user's accuracies was 53.3 % for Latin America. Four convolutional neural network models (CNN) have been developed and tested to improve deforestation detection accuracies in Peru and Brazil. By applying the suggested CNN models, averaged user's accuracies are improved from 45 % to 77 - 85 % in dry season, and from 34 % to 48 - 57 % in rainy season. Averaged relative F1 are improved from 0.6 to 0.67 - 0.75 in dry season, and from 0.49 to 0.52 - 0.63 for rainy season. It is shown that adding CNN operation improves the deforestation detection accuracies. Intermediate CNN outputs were visualized to gain insights about what features of the image does each layer learn for improving the accuracies.