Fully Connected Deep Convolutional Neural Network and Improved SURF for Land Cover Classification

Developing an accurate Land Cover (LC) mapping for efficient land management is crucial to update cover maps for producing sustainable crops. However, extracting accurate features from the LC maps to identify the regions is a challenging task. To overcome this, a Deep Leaning (DL) approach, Fully Connected Deep Convolutional Neural Network and an Improved Speeded Up Robust Features (FCDCNN-ISURF) is proposed for effective feature extraction and selection. Initially the Sentinel Satellite Imagery (Sentinel-2) dataset is pre-processed using Land Surface Water Index (LSWI) which gives the details about soil moisture level from the images. The features are extracted using ISURF for key points detection and description, which is fast and stable to scaling, rotation and illumination. The extracted features are then classified using FCDCNN which classifies the features by pixel-wise prediction. The experimental results on Sentinel-2 dataset showed better LC classification with 98.23% accuracy and 97.70% precision when compared to the results of existing methods such as Shapely Additive Explanation - Convolutional Neural Network (SHAPs-CNN) and Deep Relearning based Recurrent Neural Network (DRRNN) in LC classification.