Comparison of expert-knowledge and machine learning approaches for mapping groundwater-dependent ecosystems in a regional setting in Central Mexico

Intensive groundwater abstraction, land-use changes, and climate variability have significantly altered natural discharge and flow patterns within groundwater systems, threatening long-term groundwater sustainability. These disruptions increase the risk of degradation in ecosystems that rely directly or indirectly on groundwater discharge, i. e. groundwater-dependent ecosystems (GDEs).

Mexico is particularly vulnerable to declining water table levels, a situation accelerated by gaps in groundwater management that fail to incorporate GDEs into decision-making processes. This issue is especially critical in northeastern Michoacán, home to two of the country’s largest lakes: Pátzcuaro and Cuitzeo lakes, which represent a key study area for studying growing threats to GDEs caused by pollution, climate change, and intensive groundwater abstraction. In order to preserve GDEs, along with their associated biodiversity and ecosystem services, accurate mapping is essential to secure their future integration into groundwater sustainability policies and conservation initiatives.

To address this issue, we compared four methods usually used in geospatial mapping: the Analytical Hierarchy Process (AHP), Weights of Evidence (WoE), and two machine learning models: Logistic Regression (LR) and Random Forest (RF), using environmental variables associated with GDE presence obtained from geospatial data and remote sensing products.

Model performance was evaluated using a validation dataset derived from local inventories and fieldwork conducted in 2024, applying Receiver Operating Characteristic (ROC) curves and the Area Under the Curve (AUC) metric. Results showed that RF (AUC = 0.82) and LR (AUC = 0.70) outperformed WoE (AUC = 0.61) and AHP (AUC = 0.59), with RF demonstrating the highest predictive accuracy and best performance in cross-validation folds.

The GDEs prediction map derived from RF highlights areas primarily along the shores of both lakes, where volcanic lithology contacts with lacustrine deposits, inducing groundwater discharge through springs that sustain wetlands. Additional GDEs areas occur along fault zones that enhance discharge within volcanic lithology near Morelia City and in perennial streams located at intermediate elevations.

The study faces limitations related to varying spatial resolutions, independent errors in geospatial datasets, and uneven data quality across local zones within the study area. Furthermore, the absence of direct field verification for areas with the highest predicted GDE potential constrains the overall impact of the study. Nevertheless, this research provides significant evidence of the advantages of using machine learning approaches in regions lacking detailed hydrogeological information, supporting the integration of GDEs into groundwater sustainability management.