MULTISPECTRAL IMAGE BASED WATER QUALITY ASSESSMENT OF AQUACULTURE PONDS USING FIELD MEASUREMENTS AND MACHINE LEARNING

Abstract

Aquaculture ponds require regular water-quality monitoring, but conventional field and laboratory measurements provide limited spatial coverage and are difficult to repeat across a large number of ponds. This study proposes a multispectral image-based framework for water-quality assessment of aquaculture ponds in Haryana, India, by integrating Sentinel-2 imagery, field observations and machine learning. First, aquaculture ponds are identified and extracted from Sentinel-2 images using NDWI, MNDWI, thresholding and morphological image-processing techniques. Field measurements of temperature, pH, electrical conductivity, dissolved oxygen, turbidity, chlorophyll-a (Chl-a) and total suspended solids (TSS) are then linked with corresponding satellite observations. Pond-level spectral bands, band ratios, spectral indices and texture features are extracted rather than relying on individual pixels. Feature analysis is used to identify relevant image variables, followed by comparative modelling using multiple linear regression, Random Forest, support vector regression and XGBoost. The validated models are used to estimate selected optically related water-quality indicators and generate spatial water-quality information for individual ponds. The main novelty of the proposed work is the integration of automated aquaculture pond extraction with field-calibrated multispectral features and explainable machine learning within a single pond-level assessment framework. Unlike approaches focused only on pond mapping or individual parameter retrieval, the proposed framework establishes a link between field-measured water condition and spatial satellite observations. The study aims to provide a practical and scalable approach for targeted aquaculture water-quality monitoring in Haryana.