Leveraging Remote Sensing and AI Algorithms for Predicting Inflow and Water Levels in Transboundary Reservoirs on Eastern Reivers of Pakistan

Abstract

This study integrates remote sensing technologies and artificial intelligence (AI) algorithms to predict inflow and water levels in transboundary reservoirs, a crucial endeavor for managing shared water resources in regions characterized by complex hydrological dynamics. Effective management of these reservoirs is essential for ensuring water and food security, mitigating natural disasters, and supporting economic stability, particularly in countries like Pakistan, which often faces uncertainties regarding reservoir levels and inflows during flood seasons from adjacent transboundary sources. The primary objective of this research is to capture variations in transboundary catchments using remote sensing and AI-based modelling approaches. The focus is on the Bhakra Dam, situated on the Sutlej River in Himachal Pradesh, India, where heavy rainfall, snow melt, and dam operations significantly influence the dam's capacity, leading to potential flooding in Pakistan. This study conducts an inflow, outflow, and storage analysis using various AI algorithms, including Linear Models (LM), Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGB), Partial Least Squares (PLS), Random Forest (RF), Ridge Regression (RG), and Gaussian Processes (GP). Historical data from 2014 to 2020 regarding inflow, outflow, water levels, and storage from Indian dams was sourced from the Central Water Commission of India. Additionally, daily rainfall data from the Global Precipitation Measurement (GPM) v6 and land surface temperature (LST) and emissivity data from MODIS were utilized. A single C-band synthetic aperture radar (C-SAR) from Sentinel-1 was employed to assess surface area and extent of the dam lake. The performance of the machine learning algorithms was evaluated using the R package Hydrostat, employing a comprehensive set of statistical metrics, including Nash-Sutcliffe Efficiency (NSE), R-squared (R ²), Knowledge Graph Embedding (KGE), Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Residual Standard Deviation (RSR). To avoid redundancy, these metrics are reported succinctly. The Model Efficiency Index (MEI), with values approaching 1 indicating strong agreement between observed and predicted data, was used as a key indicator of model performance.

Among the tested models, Extreme Gradient Boosting (XGB) demonstrated the highest accuracy for predicting inflow, Random Forest (RF) excelled in outflow prediction, and Ridge Regression (RG) was most effective for storage estimation. The results demonstrate a high MEI, indicating the accuracy of AI algorithms for predicting water levels, storage, inflow, and outflow. Furthermore, a correlation model efficiency matrix was utilized to analyze relationships between variables such as level differences, area differences, temperature, and average rainfall, revealing a near-perfect positive correlation (0.99) between storage and level differences. This study underscores the potential of integrating remote sensing and AI technologies to enhance water resource management in transboundary contexts.

Potential areas of research with a specific emphasis on comparing their performance against machine learning algorithms involves comparing the predictive capabilities of machine learning algorithms with those of physically-based models for reservoir inflow forecasting. Additionally, research efforts could be directed towards leveraging remote sensing data in regions with limited data availability, particularly in developing countries.