OptTransEnsembleNet: deep ensemble learning framework for land use land cover classification

Precise land use land cover (LULC) classification plays a crucial role in urban and regional planning. Policymakers ought to have an in-depth understanding of the LULC of the landscape. Classical remote sensing methods of LULC classification are time-consuming and expensive. In recent years, deep learning (DL) has shown remarkable performances in remote sensing applications especially like LULC classification. In spite of its promising result, generalization ability of DL models, optimization of hyperparameters, and accuracy with limited datasets are major concerns of DL algorithms. In this study, an OptTransEnsembleNet DL framework which integrates deep transfer learning, deep ensemble learning, and optimization is proposed. Initially transfer learning approach is employed to train the 4 base learners on EuroSAT spectral indices derived dataset which is generated from benchmark EuroSAT dataset. The pretrained models InceptionV3, MobileNetV2, Xception, and InceptionResNetV2 are used as feature extractors and further custom layers are added. The hyperparameters of these layers of base learners are optimized with the grid search optimization algorithm. Next, the aforementioned 4 models are used as base learners in deep ensemble learning and a weighted average ensemble is created for LULC. The optimal weights for the weighted average ensemble are found using the grid search algorithm. Overall, OptTransEnsembleNet-ensemble of 4 base learners produce an overall accuracy (OA) of 96.40% which is higher in comparison to individual models such as MobileNetV2 (92.15%), Xception (93.52%), InceptionV3 (91.50%) and InceptionResNetV2 (93.80%). Likewise, the model is applied on Jagatsinghpur district in Odisha state of India for classification followed by generation of its LULC map. The proposed model helps to improve the generalization performance and robustness of a DL model. This study could be beneficial for urban planners and administrators for resource management.