Exploiting JECAM Database for Agriculture Land Cover Classification of Antsirabe Site Using Sentinel 2 Imagery with Deep Learning

Abstract

Zero hunger, the goal 2 of Sustainable Development Goals (SDGs), can only be achieved when food is available, affordable and accessible to the people. Food insecurity, a phenomenon where either or all of these ingredients for zero hunger are absent, remains a critical global issue that warrants coordinated strategies at regional scale; most especially for crop farming which serves as the major source of food for most humans. Therefore, efficient Land Use and Land Cover (LULC) classification is a pivotal tool in the development of apposite strategies for combating food insecurity. Open satellite missions like Sentinel 2 offer a cost effective way for acquiring regional imagery dataset for LULC classification; however, the relevance of such dataset is dependent on the quality of ground truth data from which the imagery dataset is created. Qualitative ground truth data are usually obtained through ground surveys which come at extra costs, warranting the need for elaborate community ground truth geo-database constructed from joint ground surveys. Such database is absent in the tropical belt that is mostly made up of developing countries where higher impacts of food insecurity are experienced. This remained the case, until recently when JECAM (Joint Experiment for Crop Assessment and Monitoring) database was developed for six countries in the tropical belt. JECAM database is an elaborate geodatabase that consists of 27,074 agricultural LULC polygons (20,257 crops and 6,817 non crops). In this study, we built three deep learning models for agricultural LULC classification using the entire 13 bands of the satellite imagery dataset. Class-based performance evaluation metrics were used to evaluate the performances of the deep learning models on test set. LSTM (Long Short-Term Memory) model exhibited the highest capability for LULC class discrimination, followed by 2D-CNN (2 Dimension Convolution Neural Network) Autoencoder model, then the 2D-CNN model. In the future, we intend to exploit spectral indices and transfer learning paradigm to address class imbalance problem, which is inherent in the imagery dataset, for improved LULC class discrimination.