URBAN LAND COVER CLASSIFICATION USING DEEP NEURAL NETWORKS BASED ON VHR MULTI-SPECTRAL IMAGE AND POINT CLOUD INTEGRATION

The recent development in computer processing and memory performance, and the growing availability of satellite imageries have made deep learning and convolutional neural networks more effective with wide range applications in urban research fields. The urban environment comprises several significant uses of remote sensing, such as feature extraction, building identification, transportation system management, and infrastructure monitoring. AI-based algorithms, especially convolution neural networks, are characterized by their ability to achieve higher performance for satellite image classification by analyzing multiple accessible data and utilizing all available features. In this paper, a Convolutional Neural Network (CNN) model, based on GoogLeNet, was applied for land cover classification in urban environment. The suggested model was designed, trained, validated, and implemented in a built-up study area with building, road, and vegetation classes. To enhance the classification capabilities required for accurate discrimination of land cover classes, WorldView-2 satellite image has been integrated with point cloud data which enable deriving digital surface model, intensity, normal vector, surface variation, and vertical layers. The model assessment process involved comparing the classification results to the Object Based Image Analysis (OBIA) method to verify and validate the approach. The proposed CNN model achieved an overall accuracy of 83.25% for the classification process and 87.00% for the post-classification refinement outperforming the OBIA classification results of 80.75% and refinement accuracy of 83.50%.