Deep and Ensemble Learning Based Land Use and Land Cover Classification

Monitoring of Land use and Land cover (LULC) changes is a highly encumbering task for humans. Therefore, machine learning based classification systems can help to deal with this challenge. In this context, this study evaluates and compares the performance of two Single Learning (SL) techniques and one Ensemble Learning (EL) technique. All the empirical evaluations were over the open source LULC dataset proposed by the German Center for Artificial Intelligence (EuroSAT), and used the performance criteria -accuracy, precision, recall, F1 score and change in accuracy for the EL classifiers-. We firstly evaluate the performance of SL techniques: Building and optimizing a Convolutional Neural Network architecture, implementing Transfer learning, and training Machine learning algorithms on visual features extracted by Deep Feature Extractors. Second, we assess EL techniques and compare them with SL classifiers. Finally, we compare the capability of EL and hyperparameter tuning to improve the performance of the Deep Learning models we built. These experiments showed that Transfer learning is the SL technique that achieves the highest accuracy and that EL can indeed outperform the SL classifiers. © 2021, Springer Nature Switzerland AG.