Classification of agricultural land cover using satellite imagery with deep learning

Food insecurity is a dreadful global phenomenon with its most devastating impacts experienced in developing countries such as Madagascar. Lately, crop maps have been invaluable tools to developed complex agricultural systems for weed control, yield estimation, crop monitoring and other innovative applications that can drastically promote food security. A cost-effective way of producing accurate crop maps at large scales is the use of outputs from deep learning classifiers built from satellite image data. However, previous works carried out in the development of satellite-based deep learning model for agriculture landscapes discrimination in developing countries are very few due to absence of appropriate ground truth record for the creation of training dataset. In this work, we applied deep learning algorithms on customized multispectral Sentinel-2 L1C image data, generated from Joint Experiment for Crop Assessment and Monitoring (JECAM) ground truth record, to develop classifiers for land cover and crop discrimination using Antsirabe in Madagascar as a region of interest. We further explored Big Transfer (BiT), a state-of-the-art transfer learning method, employing Eurosat land cover dataset to build deep classifiers for Land Use and Land Cover (LULC) discrimination. Using class-based evaluation metrics, bidirectional long short-term memory (LSTM) classifier exhibited the highest performance for both land cover and crop discrimination tasks followed by a hybrid of convolution neural network (CNN) and LSTM model. Our result further shows that depth of the model architecture and the abundance of image samples employed for upstream trainings are contributive to improving discrimination performances of resulting downstream models; however, the latter demonstrates higher significance in improving model performance for downstream discrimination tasks. In the future, we intend to exploit spectral indices paradigm to address class imbalance problem, which is inherent in the imagery dataset, for improved LULC and crop discrimination. Keywords: Food Security, Land Cover, Land Use, Deep Learning, Satellite Imagery, Remote Sensing, Crop Mapping