Nawale et al. (2025) Lake recharge projections with autoencoders and hydro-climatic indicators

Identification

• Journal: Modeling Earth Systems and Environment
• Year: 2025
• Date: 2025-10-03
• Authors: Sanchita Nawale, Santosh Murlidhar Pingale, Meenal Mategaonkar
• DOI: 10.1007/s40808-025-02619-1

Research Groups

• Narsee Monjee Institute of Management Studie’s Mukesh Patel School of Technology Management and Engineering, Mumbai, India
• National Institute of Hydrology Roorkee, India

Short Summary

This study models groundwater recharge dynamics for Nakane Lake, India, from 2024–2030 using an autoencoder integrated with hydro-climatic and remote sensing data, demonstrating its superior accuracy in forecasting a cumulative groundwater depth increase of approximately 0.92 m by December 2030.

Objective

• To analyze the role of Nakane Lake in groundwater recharge and its impact on surrounding groundwater levels.
• To assess the influence of lake water percolation on groundwater recharge.
• To model lake–groundwater interactions using historical and projected data from 2024 to 2030.
• To apply machine learning methods, including autoencoders and Principal Component Analysis (PCA), to capture nonlinear recharge dynamics.
• To forecast lake water depth using climate indices to support sustainable water resource management in the Dhule region of Maharashtra, India.

Study Configuration

• Spatial Scale: Nakane Lake in Dhule, Maharashtra, India (20° 53′ 9′′ N latitude and 74° 43′ 3′′ E longitude). Data resolution for global surface water mapping was 30 m.
• Temporal Scale: Modeling and projection period from 2024 to 2030. Historical data used includes Global Surface Water (GSW) from 1984–2018, population data from 1990–2040, and water level trends from 2011–2020.

Methodology and Data

• Models used:
  • Deep Learning: Autoencoder (optimized with grid search, Adam optimizer, ReLU activations, L2 weight decay, dropout).
  • Machine Learning: Random Forest (RF), XGBoost, Gradient Boosting Decision Tree (GBDT), Bayesian Ridge (BR), K-Nearest Neighbors (KNN), Extra Trees (ET), Adaptive Boosting (AB), Bagging.
  • Clustering/Dimensionality Reduction: Fuzzy Logic (for water demand clustering), Principal Component Analysis (PCA), Gaussian Mixture Models (GMM) (optimized with BIC and AIC), K-Means clustering.
  • Statistical Analysis: Shapiro-Wilk test, Linear Regression, Lin’s Concordance Correlation Coefficient (LCCC).
  • Hydrological Index: Standardized Precipitation-Evapotpiration Index (SPEI-12).
• Data sources:
  • Remote Sensing: NASA POWER (Prediction of Worldwide Energy Resources) for hydro-climatic variables (temperature, precipitation, potential evapotranspiration); Global Surface Water (GSW) dataset (from Landsat 5, 7, 8) for surface water bodies; National Remote Sensing Centre (NRSC) and Bhuvan portal for spatial parameters (slope, soil type, LULC).
  • Observation/Reanalysis: Central Ground Water Board (CGWB) for inflow, outflow, hydrogeological data, lake statistics, actual evaporation, and groundwater levels; British Columbia dataset for detailed flow and temperature; Census of India for population data; United States Geological Survey (USGS) for rainfall data.
  • Local Data: Dhule Municipal Corporation for essential data.

Main Results

• The autoencoder model demonstrated exceptional performance in capturing complex, nonlinear relationships in hydrological data, achieving an R² of 0.999 for groundwater recharge projection.
• Among nine machine learning algorithms tested, the autoencoder consistently outperformed all others, showing nearly perfect R² and LCCC values.
• Optimization of the autoencoder using grid search cross-validation led to a 26.95% reduction in RMSE and a 6.74% increase in R².
• Projections indicate a consistent upward trend in groundwater levels in the Nakane Lake region, with groundwater depth increasing from 0.022 m in January 2024 to 1.51 m by December 2030, representing a cumulative gain of approximately 0.92 m over the six-year period.
• Seasonal variations were observed, with significant drops in lake depth during mid-year dry months (e.g., 0.453 m in July 2024) and recoveries during post-monsoon months (e.g., 0.611 m in December 2030), highlighting substantial seasonal groundwater recharge.
• The study identified four base flow clusters (Low, Lower-Medium, Upper-Medium, High) using GMM, which accurately illustrate different base flow levels.

Contributions

• This study presents a novel hybrid modeling approach combining fuzzy clustering with unsupervised autoencoder models for groundwater recharge forecasting, which is a first attempt in this context.
• It improves groundwater recharge forecasting accuracy by capturing nonlinear dynamics and incorporating real-world uncertainties in water demand using fuzzy logic, especially in climate-sensitive, semi-arid lake regions.
• The research provides a robust framework for simulating lake-groundwater interactions under future climatic scenarios (2024–2030), addressing a gap in existing literature.
• It offers a paradigm shift for sustainable water management strategies in similar semi-arid areas by demonstrating the value of intelligent modeling for anticipating groundwater trends.

Funding

Not explicitly stated in the paper. The Dhule Municipal Corporation was thanked for providing essential data.

Citation

@article{Nawale2025Lake,
  author = {Nawale, Sanchita and Pingale, Santosh Murlidhar and Mategaonkar, Meenal},
  title = {Lake recharge projections with autoencoders and hydro-climatic indicators},
  journal = {Modeling Earth Systems and Environment},
  year = {2025},
  doi = {10.1007/s40808-025-02619-1},
  url = {https://doi.org/10.1007/s40808-025-02619-1}
}