LAND COVER CLASSIFICATION OF HIGH RESOLUTION IMAGES FROM AN ECUADORIAN ANDEAN ZONE USING DEEP CONVOLUTIONAL NEURAL NETWORKS AND TRANSFER LEARNING

Abstract

Different deep learning models have recently emerged as a popular method to apply machine learning in a variety of domains including remote sensing, where several approaches for the classification of land cover and use have been proposed. However, acquiring a suitably large data set with labelled samples for training such models is often a significant challenge to tackle, that leads to suboptimal models not being able to generalize well over different types of land cover. In this paper, we present an approach to perform land cover classification on a small dataset of high-resolution imagery from an area in the Andes of Ecuador using deep convolutional neural networks and techniques such as transfer learning, data augmentation, and some fine-tuning considerations. Results demonstrated that this method can achieve good classification accuracies if it is backed with good strategies to increase the number of samples in an imbalanced dataset.