A machine learning methodology for land use/land cover classification in tropical areas using medium resolution satellite imagery, case: Colombia

Abstract

ABSTRACT : In this work, we first present a methodology for preparing 10 m to 60 m spatial resolution Sentinel-1, Sentinel-2, and ALOS DSM imagery of forest/grassland areas in Colombia to train a DeepLabV3+ convolutional neural network model. Our preprocessing pipeline for the Sentinel-2 imagery comprises cloud and shadow removal, atmospheric correction, and topographical correction, resulting in mostly cloud-free mosaics of tropical areas. At first, we train the network on very low spatial resolution (500 m) labels of the Colombian Amazonas region resampled to 10 m (+100000 samples after augmentation). Then, we fine-tune the network on medium spatial resolution data (30 m) of northern Antioquia, also resampled to 10 m, resulting in faster convergence and higher accuracy despite the limited number of labelled samples (~5000 samples after augmentation). Our results validate recent proposals where low spatial resolution data is used for training neural networks, and motivate us to keep exploring this line of research.