Derivation of surface models using satellite imagery deep learning architectures with explainable AI

Global mapping of urban morphology and human settlement is an area of great interest, with significant scientific and humanitarian value. Current approaches for high-resolution surface modeling using aerial LiDAR are impressive and useful, but it is impractical for application on a global scale. This is due to technical challenges (efficiency of coverage, etc), as well as cost. These are significant obstacles, particularly in the less wealthy regions of the world, many of which are experiencing the greatest increase in population and physical infrastructure. Advances in satellite imagery and deep learning are now enabling the lower resolution alternative, but one which is more scalable with global application. This paper examines the efficacy of four state-of-the-art deep learning architectures (DeepLabv3+, SegNet, U-Net, and U-Net++) for application to Digital Surface Modeling (DSM), particularly for urban districts. The study considered New York City as the test area, using satellite based Radar backscatter images and visible and near infrared images for surface model development, along with high resolution LiDAR based surface height model of the city. We performed a quantitative evaluation using metrics such as Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and R-Squared, and a qualitative analysis to identify the best performing model. U-Net++ proved to be the most effective, with superior accuracy in capturing key urban features. In addition, we employed Explainable Artificial Intelligence (XAI) techniques, such as Gradient-Weighted Class Activation Mapping (GRAD-CAM) and Saliency Maps, to improve model interpretability, revealing important information about decision-making processes and highlighting critical input features. The results of this study contribute to the advancement of remote sensing techniques for urban analysis, with potential applications. Despite the overall success, the study also identified challenges, such as the loss of detailed building shapes in the predicted DSMs and the influence of optical shadowing and SAR-induced quiescence on model performance. • Comparison of four CNN models for DSM generation, with U-Net++ showing the best results. • DSMs generated from multispectral satellite imagery using deep learning. • XAI techniques (GRAD-CAM, Saliency Maps) applied to improve model interpretability. • Challenges identified: loss of building details and SAR distortions in DSM predictions.