Automated karst spring catchment approximation for discharge modelling

Abstract. Karst spring discharge modelling is complicated by the limited availability of hydrogeologically delineated catchments, particularly in large-sample applications. We present an automated approach to approximate karst spring catchments from discharge, gridded recharge, and topography and evaluate its suitability for discharge modelling at 63 springs across the wider Alpine region. Recharge-area size is estimated from a simplified water balance and represented by a circular catchment whose position is informed by the topography. The approximations are evaluated against expert-based catchments in terms of geometry and derived model inputs and subsequently applied to discharge modelling with the lumped reservoir model KarstMod and a global long short-term memory (LSTM) network. The estimated recharge areas agree within ±50 % of the expert-based delineations for 65 % of the springs, while spatial overlap is more limited (median Intersection over Union: 0.34). Despite this imperfect geometric agreement, the approximations represent catchment-mean elevation, precipitation, and temperature more closely than extracting inputs solely at the spring location. For KarstMod, this translates into substantially improved discharge simulations: approximated and expert-based catchments both achieve a median Kling–Gupta efficiency (KGE) of 0.72, compared with 0.63 for the point-based representation. Differences are particularly pronounced for snow-influenced springs, where representation of catchment elevation affects temperature forcing and simulated snow accumulation and melt. The global LSTM is less sensitive to catchment representation, although greater sensitivity also emerges for snow-influenced springs. These results show that accurate reconstruction of hydrogeological catchment boundaries is not necessarily required to improve model inputs for lumped-parameter modelling. The proposed approximation does not replace expert hydrogeological delineation but provides a scalable, physically informed catchment representation for large-sample karst spring discharge modelling where such delineations are unavailable.