Combating Food Insecurity in Madagascar Using Sentinel-2 Imagery with Deep Learning Land Cover and Crop Classifiers

Abstract

Food insecurity is an ugly global phenomenon with its most devastating impacts felt in developing countries such as Madagascar. Recently, crop maps have been valuable tools to build complex agricultural systems for yield estimation, crop monitoring, weed control and other applications that can promote food security. A cost effective means of generating accurate crop maps at large scales is the use outputs from deep learning classifiers developed from satellite imagery; however, works done in the development of satellite-based deep learning model for agriculture landscapes classification in developing countries are very little due to absence of suitable ground truth database for the creation of training dataset. In this study, we applied deep learning methods on multispectral Sentinel-2 L1C imagery, created from Joint Experiment for Crop Assessment and Monitoring (JECAM) ground truth database, to develop models for land cover and crop classification using Antsirabe in Madagascar as an area of interest. Bidirectional long short term memory (LSTM) model exhibited the highest performance for both land cover and crop classification tasks followed by a hybrid of convolution neural network (CNN) and LSTM model.