DSM generation from VHR spaceborne imagery using deep learning-based land-cover-aware GCP extraction

Accurate digital surface model (DSM) generation from very high-resolution satellite stereo imagery remains constrained by the labor-intensive requirement for manual ground control point (GCP) measurement. This work presents a fully automated deep learning framework that eliminates this bottleneck by combining SuperPoint feature detection with LightGlue matching to generate both tie-points (TP) between stereo image pairs and GCPs via correspondence with reference ortho-images. Both rely on the transformer architecture, on self- and cross-attention, and outperform traditional methods like SIFT, SURF, or ORB in terms of accuracy and speed. A critical innovation integrates land-cover-aware filtering to mitigate systematic height errors arising from geometric misalignments between reference ortho-images and digital elevation models, particularly in urban environments where inaccurate ortho-rectification produce unreliable GCP heights. The land cover itself is classified with a CCN-based architecture, specifically, a U-Net derivation with categorical focal loss and custom augmentation strategy in the training phase. Applied to a tri-stereo Pléiades dataset over Eisenstadt, Austria, the method automatically generates around 2000 filtered GCPs that achieve sub-meter planar accuracy (0.5m RMS in easting/northing) and approximately 2m vertical accuracy after block adjustment of rational polynomial coefficient sensor models. The proposed land-cover-aware method yields a lower RMS error in height of 0.55m w.r.t. the automated method. By replacing manual GCP measurement with deep learning-based automation and land-cover-aware quality control, this work completes the end-to-end workflow for rapid, high-precision DSM production across diverse terrain types, with direct applicability to urban mapping, change detection, and continental-scale monitoring missions.