Urban Land Cover Classification Using Deep Neural Networks Based on VHR MS Image and DSM

Abstract

Urban environment presents one of the most challenging research fields for remote sensing data analysis tasks due to the wide range of land cover materials, and the variety of land use classes. Urban feature extraction, e.g. buildings and roads, has a significant impact on several applications such as urban planning and management, monitoring human activities, and change detection. Effective feature extraction procedures require both high-quality data and reliable processing methodology. Recent progress in remote sensing techniques offers a wide range of Multi-Spectral (MS) images and Digital Surface Models (DSMs) for urban studies. Deep Learning (DL) approaches, especially Convolutional Neural networks (CNNs), have a notable performance in handling large amounts of data with the advantage of mapping the relationship between high-dimensional and nonlinear features. The current research employs the U-NET to develop a CNN model for urban feature extraction from multi-spectral images and DSMs. The proposed U-NET is trained, validated, tested, and applied for image semantic segmentation using integrated WorldView-2 multi-spectral image and LiDAR DSM. The classified urban features are subsequently refined based on the elevation values of the DSM. Four accuracy indices: correctness, completeness, quality, and overall accuracy are calculated to evaluate the obtained outcomes and check the model stability. Building extraction has attained an overall accuracy of 69.1% and 89.9% for classified and refined images respectively, whereas road extraction has obtained an overall accuracy of 89.9% and 90.7% for classified and refined images respectively. The U-NET model has achieved promising outcomes for image semantic segmentation, while the DSM added a notable enhancement during refinement.