Researchers in Odisha, India, have combined three machine learning models with classical irrigation indices to judge whether surface water in the Mahanadi River Basin is fit for watering crops, and their results suggest that most of the water they sampled is suitable, with a handful of stations standing out as problem sites. The study, published in the journal Green Technology, Resilience, and Sustainability, is described as original research and was made available as an open access article on 19 September 2025.
The work was led by Abhijeet Das, who set out to map surface water quality in the basin using what the paper calls integrated multiproxy methods. The team collected nine water samples from nineteen sampling places, classified over a period from 2017 to 2023, and evaluated them with standard methods recommended by the American Public Health Association. Sampling was concentrated in the dry season, when cation concentrations are higher, anions are more stable and water levels are lower. Quality control included double analysis of blank samples, pre-calibration of instruments and an ion balance error check, with all samples falling within the ten percent range regarded as trustworthy for hydro-chemical analysis.
The chemistry of the samples followed a clear ordering. Among the cations, magnesium dominated, followed by calcium, sodium and potassium, while among the anions bicarbonate came first, ahead of chloride and sulphate. The pH values ranged from 7.73 to 7.92, an alkaline state that the authors note is typically found in gently alkaline soils or uncontaminated rainfall, and most areas fell below the World Health Organization drinking limit of 6.6 to 8.5. Electrical conductivity ranged from 138 to 295 microsiemens per centimetre, with a mean of about 186.9, which the authors describe as good conductivity. Calcium ranged from 9.43 to 28.18 milligrams per litre, magnesium from 14.83 to 28.72, sodium from 1.58 to 4.63 and potassium from 2 to 10.1, all below their respective WHO reference values.
Two parameters broke the pattern. Chloride ranged from 9.65 to 271 milligrams per litre, and sulphate from 4.97 to 217, and at one site, labelled V-(9), both exceeded the WHO limits of 250 and 200 milligrams per litre respectively. The authors attribute the elevated chloride to overfertilization linked to intensive farming in the affected region, and the high sulphate trend to weathering, mineral breakdown and excessive evaporation. Excluding that site, they conclude that surface water quality is appropriate for drinking as well as non-irrigation use, although they also note that the constructed physicochemical results show particular areas of low surface water quality in the downstream parts of the study block.
To judge irrigation fitness, the team applied six agricultural indices: the sodium adsorption ratio, sodium percentage, soluble sodium percentage, magnesium adsorption ratio, potential salinity and Kelly’s ratio. The sodium adsorption ratio values ranged from 0.41 to 16.59, with the highest reading at V-(9). At eighteen separate places the ratio was below ten, which the paper classifies as perfect for irrigation with no effect on the soil, while values above ten signal substantial physical soil concerns that make it harder for plants to absorb water. For sodium percentage, the maximum value was 91.0 and the minimum 22.95, and the samples split into 42 percent acceptable, 36.84 percent permissible, 10.52 percent questionable and 10.52 percent unacceptable for irrigation, with elevated values at stations V-(2), (8), (9) and (19).
The soluble sodium percentage reached a maximum of 92.0 and a minimum of 23. Fifteen of the nineteen tested samples, or 78.94 percent, were suitable for irrigation, while the remaining four samples, at stations V-8, 9, 2 and 19, reduce the permeability of water flow by closing soil pores and causing specific soil aggregates to scatter, which the authors say makes irrigation unsuitable in those spots. The magnesium adsorption ratio ranged from 18.34 to 85.55, averaging 39.58, and 89.47 percent of samples fell below fifty, the threshold for suitability, while 10.52 percent exceeded it at stations V-(9) and (19), where rising soil alkalinity is expected to harm crop yields.
Potential salinity, calculated from chloride and half the sulphate concentration, ranged from 0.13 to 10. Under the classification used, values below five are excellent to good, between five and ten are good to injurious, and above ten are injurious to unsatisfactory. The authors report that 94.73 percent of all samples were deemed suitable and classed as excellent to good, 10.53 percent fell into the good-to-injurious category, and one location was considered not acceptable. Kelly’s ratio, which compares sodium with calcium and magnesium, ranged from 0.32 to 8.5, with 78.95 percent of samples categorized as excellent and the remaining 21.05 percent doubtful or not suitable for farming. The authors attribute these patterns to both geological and human-caused sources.
The machine learning component is where the study aims to break new ground. The authors trained three models, an artificial neural network, a random forest and a support vector machine, executed in MATLAB, to predict the six irrigation indices from nine water quality variables. Performance was judged with the coefficient of determination, root mean square error and mean absolute percentage error. For the sodium adsorption ratio, the three models achieved R-squared values of 0.96, 0.99 and 1.0 respectively, and for sodium percentage the values were 0.99, 0.93 and 1.0. Soluble sodium percentage gave 0.99, 0.95 and 1.0, magnesium adsorption ratio 0.99, 0.94 and 1.0, potential salinity 0.99, 0.94 and 1.0, and Kelly’s ratio 0.99, 0.91 and 1.0. Across these metrics the support vector machine consistently outperformed the other two, and the authors conclude that the overall model performance trend for most indices follows the order SVM, then ANN, then RF.
The models also ranked the importance of the input variables, and the rankings differed by target and by model. For the sodium adsorption ratio, the support vector machine ranked sodium first, followed by chloride, pH, electrical conductivity, sulphate, bicarbonate, potassium, magnesium and calcium, while the random forest put calcium first. For sodium percentage the SVM ranking started with sodium and calcium, whereas the random forest again placed calcium on top. The authors suggest that merging these algorithms with surface water suitability forecasts will cut the cost and duration of analysis for future researchers, and that the approach can be replicated locally and globally with similar techniques.
Beyond the numbers, the study reports governance deficiencies in the region that increased the vulnerability of surface water, citing insufficient monitoring procedures, low institutional capacity, low awareness and ineffective application of rules and regulations. The authors recommend a multimodal strategy for surface water management that incorporates socioeconomic, cultural and technical elements, along with cooperative efforts, legislative changes, community involvement and efficient enforcement. They also propose routine water quality monitoring across the watershed, stricter anti-pollution laws for industries, modernized waste and sewage infrastructure, public health education initiatives and cooperation between local communities, non-governmental organizations and government agencies. The study does carry limitations the authors themselves flag: the three models are not inherently consistent with human reasoning, and explaining black box models through proxy instruments may not adhere to rigid scientific criteria, so they call for future models that combine high accuracy with natural interpretability. The research, they write, provides valuable insights for policymakers and water resource managers in Odisha.
Subject of Research: Machine learning assessment of surface water suitability for irrigation in the Mahanadi River Basin, Odisha
Article Title: Pollution source apportionment and application of machine learning approaches in surface water suitability for irrigation based on hydro chemical analysis
Article References: Das, A. (2025). Pollution source apportionment and application of machine learning approaches in surface water suitability for irrigation based on hydro chemical analysis. Green Technology, Resilience, and Sustainability, 5(1), Article 4. https://doi.org/10.1007/s44173-025-00021-9
Image Credits: AI Generated
DOI: 10.1007/s44173-025-00021-9
Keywords: surface water, irrigation suitability, machine learning, Mahanadi River Basin, hydrochemistry, support vector machine, artificial neural network, random forest, water quality, Odisha, sodium adsorption ratio, GIS
News Source: Blake Davidson. (October 30, 2026). Machine learning maps irrigation safety of surface water in Odisha’s Mahanadi basin. Scienmag.
