Transfer learning for identifying rainwater harvesting sites in training data-scarce catchments

Machine learning (ML) based approaches for rainwater harvesting (RWH) site suitability mapping often face limited data for training ML models. This study investigates whether the knowledge learned from data-rich catchments can be effectively transferred to the hydrologically and geomorphically similar and distinct basins. The transfer-learning (TL) framework is tested across three Indian catchments of different climatic conditions in the states of - Odisha, Maharashtra, and Tamil Nadu-using seven key influential predictors derived from LiDAR DEMs, geological and soil maps, and GR4J-simulated annual streamflow. A total of 18 experimental cases are designed, viz., intra, direct transfer, adaptation-based transfer, and multi-catchment combinations. Feature similarity between the source and target catchments is quantified using Kullback-Leibler (KL) divergence to systematically assess conditions for transfer feasibility. Four ML models-Support Vector Machine (SVM), Random Forest (RF), XGBoost (XGB), and KNN (K-Nearest Neighbour)-are trained over 100 randomized iterations using curated suitable and unsuitable RWH locations. Results show that performance of direct transfer declines sharply when dominant catchment features exhibit high divergence. RF and XGB are more resilient to cross-domain variability, while SVM and KNN are highly sensitive to feature mismatches. Incorporating a 20% target-domain sample substantially improves kappa and F1 scores by 25-40% in Odisha, 20-30% in Maharashtra, and 35-45% in Tamil Nadu, respectively, and stabilizes feature importance between the source and target domains. Multi-source training, which aggregates data from all three catchments, achieves the highest accuracy (>0.95) and kappa (>0.85). KL divergence analysis confirms that transferability depends on geomorphic similarity; that is, lower divergence enables effective transfer, whereas higher divergence reduces generalization. Overall, this study demonstrates that TL is feasible and beneficial for RWH site mapping, provides a practical and transferable tool for sustainable RWH planning in data-limited regions and also builds foundation for developing wide area scalable models by continuously updating the weights of influential thematic layers as new training data becomes available.