Applicability of farmlands detection in Djibouti from satellite imagery using deep learning

Abstract

This study attempted to investigate the applicability of Mask-R-CNN to a farmlands detection in Djibouti using Sentinel-2. As training data, a total of 670 data were created, of which 580 were used for training and ninety were used for testing. Training was performed on RGB image and vegetation index (NDVI, NGRDI, NDWI) image from Sentinel-2 using Mask R-CNN. To compare a training model, SVM (Support Vector Machine) were also applied to the same Sentinel-2 images. In order to evaluate the performance, by defining the true positive (TP) as a truly detected object, the false positive (FN) as non-detected objects, the false positive (FP) as falsely detected objects, the precision and recall rate were calculated as performance indices. As a result, most of the indices obtained from the vegetation index image showed higher accuracy than that of the RGB image in both Mask R-CNN and SVM models. The recall rate of SVM is extremely low, which means that false positives are very frequent. Therefore, Mask R-CNN, which has low false positives is considered more useful than SVM the detection of farmlands. Due to the different distribution and location of farmland in urban and rural areas of Djibouti, the trained Mask R-CNN using vegetation index image was tested excluding urban areas with dense farmland. The recall rate increased from 6.5% to 19.6%, indicating that the accuracy of the model is expected to be improved by considering the locational characteristics of farmland.