Development of a Deep Neural Network (DNN) Model for Feature Selection from Satellite Images

Advances in space-based observation, using remotely sensed data, have proved to be an important tool to monitor the globe, including the areas inaccessible to humans. The Sundarbans deltaic region, witnessing the confluence of vast expanses of tropical mangrove forests, tidal rivers, and estuaries, is one such area. Considered as one of the richest biodiversity hotspot zones on earth, home to a large spectrum of biodiversity (flora and fauna), including endangered or threatened species, this forest plays a critical role in land reclamation, coastal habitat protection, and local socioeconomics. However, the forests have been experiencing changes due to climatic forces and anthropogenic activities. Monitoring these changes is crucial for adopting precise management practices. In this work, Landsat 8 images were used to identify the land use and land cover in the Sundarbans. For classification, a new Deep Neural Network (DNN) model is proposed. A comparative analysis of the Overall Accuracy (OA) of the proposed DNN model with two popular Machine Learning models, Random Forest and XGBoost showed 98.9%, 97.0%, and 98.1% OA, respectively. SHapely Additive exPlanations were used for each model to obtain important features. It was observed that Near-Infrared, Short Wave Infrared 1, Blue, and Enhanced Vegetation Index were the most important features. The proposed DNN model outperformed the RF and XGBoost models with these four important features, achieving 98.5% accuracy. In comparison, it was concluded that deep learning techniques are more effective in feature selection from remote sensing images.