Improved Stereo Matching Technique for Digital Elevation Model Generation in Mountainous Terrain

Abstract Geospatial applications like terrain analysis, infrastructure design, and area mapping require accurate Digital Elevation Models (DEMs) derived from satellite stereo imagery. Classical photogrammetric approaches using Rational Polynomial Coefficients (RPCs) and Ground Control Points (GCPs) remain widely adopted for precise elevation modeling. Recently, Deep Learning (DL) based stereo matching methods have emerged as promising alternatives, enabling automated DEM extraction in complex terrain conditions. The proposed work presents a systematic, application-oriented comparison between conventional photogrammetric workflows and DL assisted stereo pipelines for DEM generation using CARTOSAT-1. Quantitative assessment was conducted against high-quality reference data using standard accuracy metrics, including Root Mean Square Error (RMSE), Mean Absolute Error (MAE), terrain-specific error statistics, cross-sectional elevation profile analysis, and terrain error mapping. Results indicates, incorporating GCPs significantly improves geometric accuracy, achieving vertical RMSE values within a few meters. DL based stereo approaches that operates w/o GCPs, demonstrate competitive performance in overall elevation estimation. Among the evaluated networks, GANet achieves the lowest error and highest structural consistency, especially in steep and vegetated regions. While RPC-based photogrammetric methods with GCPs provide the highest absolute accuracy. DL shows great potential for automated DEM generation where GCPs are scarce, highlighting the complementary strengths of both approaches.