AI Meets Expert Judgment to Map Safer Waste Dumps in Indian City

In the crowded municipality of Raiganj in West Bengal, India, roughly 22 metric tons of municipal solid waste accumulate every single day, much of it ending up in open dumps that threaten the Kulik River along the city’s southern edge and the dense slum communities that line its banks. A new study published in Discover Green Chemistry offers a way out of this impasse, combining the decades-old Analytic Hierarchy Process with three machine learning algorithms to produce a ward-by-ward map of where dumping is safe, where it is risky, and where it must be banned outright. The work, led by Prolay Mondal of Raiganj University and Sunil Saha of the National Atlas and Thematic Mapping Organisation in Kolkata, arrives at a moment when the global waste crisis has become impossible to ignore.

The stakes are enormous. According to the United Nations Environment Programme’s Global Waste Management Outlook 2024, municipal solid waste generation worldwide is expected to climb from about 2.3 billion tons in 2023 to 3.8 billion tons by 2050. Open dumps and landfills already account for nearly 20 percent of all anthropogenic methane emissions, and their hidden environmental and public health costs are estimated at 361 billion dollars per year. Asia’s waste output is projected to grow from 802 million tons in 2016 to more than 1.1 billion tons by 2030, with South Asia bearing a disproportionate share of the methane-emitting dumpsites. In India alone, between 62 and 70 million tons of municipal solid waste are generated annually, and roughly 70 percent of it is openly dumped or otherwise poorly managed.

The consequences of poor siting decisions are visible in places like the Ghazipur landfill in Delhi, which rises more than 65 meters and emits thousands of tons of methane each year. Surveys cited in the study indicate that up to 84 percent of residents living near that landfill suffer from respiratory problems. Against this backdrop, choosing where to place a dumpsite is not a trivial administrative task but a multi-criteria decision problem that must balance environmental constraints, hydrology, topography, social exposure, and economics. That is precisely the kind of problem the Raiganj researchers set out to formalize with data rather than intuition.

The methodological core of the study is a fusion of expert-driven and data-driven weighting. The Analytic Hierarchy Process, developed by Thomas Saaty in 1980, breaks a complex decision into a hierarchy of criteria and sub-criteria, then assigns weights through pairwise comparisons on a scale from one, meaning equally important, to nine, meaning extremely important. The researchers built a 21-by-21 pairwise comparison matrix covering every variable in their analysis, normalised it, and derived criterion weights using Saaty’s method. Critically, they verified that their judgments were internally consistent: the Consistency Index came out at 0.02163 and the Consistency Ratio at 0.0134, far below the 0.10 threshold that would signal unreliable comparisons. For a matrix of 21 variables, the random index was approximated at 1.98 using Monte Carlo simulation.

On the machine learning side, the team trained three classifiers on a binary dataset of 80 sampled locations described by 21 variables, all resampled to a common 30-meter resolution. Random Forest, an ensemble of decision trees, was configured with 100 estimators and a maximum depth of five. Logistic Regression served as an interpretable baseline, with all inputs standardised to zero mean and unit variance and a regularisation parameter of 1.0. Support Vector Machine, equipped with a Radial Basis Function kernel, was chosen for its ability to handle nonlinear, high-dimensional data even with small samples, an important consideration given Raiganj’s modest 80-point dataset. The classes were moderately imbalanced, with about 50 susceptible and 30 non-susceptible observations, so the team used class weightings and imbalance-aware metrics such as weighted precision, recall, and F1-score, validated through fivefold cross-validation.

The most novel element is the aggregation step, which the authors call an Adaptive Confidence-Weighted Ensemble with Geometric Mean Modulation. Rather than simply averaging the weights produced by AHP and the three machine learning models, the method first computes the variance of each model’s weights across the dataset and assigns reliability scores inversely proportional to that variance. It then calculates a confidence-weighted average for each variable, multiplies this by the geometric mean of the weights from all sources to capture consensus, and normalises the resulting ensemble weights to sum to one. The final susceptibility map is generated through a Weighted Linear Combination in a geographic information system, translating the fused weights into spatially explicit zones of low, moderate, and high dumping susceptibility.

The variable set is unusually rich compared with earlier studies that typically used six to twelve factors. It spans five spectral indices derived from satellite imagery, including the Normalized Difference Built-up Index, the Normalized Difference Vegetation Index, the Normalized Difference Water Index, short-wave infrared reflectance, and land surface temperature, alongside six environmental factors such as slope, elevation, soil texture, and distance to water bodies, and ten anthropogenic and socio-economic variables including distances to roads, temples, markets, nursing homes, government offices, slums, and the existing open dump, plus night-time lights and population distribution. When the ensemble results were broken down, population distribution, distance from slums, fused nightlight proxy, and distance from the existing open dumping site emerged as the dominant determinants of susceptibility, with individual contributions exceeding 26 percent in several cases.

Performance comparisons confirmed what earlier literature had suggested: Random Forest outperformed its rivals, achieving an accuracy of 0.89, a precision of 0.88, a recall of 0.85, and an F1-score of 0.86. Logistic Regression delivered respectable balanced scores with 0.84 accuracy, while SVM lagged behind on every metric, suggesting it struggled to capture the underlying pattern in the small dataset. The AHP-only model, evaluated against the same samples, achieved 0.81 accuracy. These figures align with a broader pattern in the field: hybrid AHP-machine learning studies have repeatedly outperformed single methods, with reported accuracies of 92 percent for an AHP-Random Forest application in Croatia and 89 percent for an AHP-GIS approach in Iran, while standalone machine learning studies in Egypt, Tehran, Macedonia, Kolkata, and Serbia have reported accuracies ranging from 80 to 90 percent.

When the ensemble model was applied across Raiganj’s 25 municipal wards, it classified roughly 23.73 percent of the city’s 10.76 square kilometers as low susceptibility, 40.82 percent as moderate, and 35.45 percent as high. The high-susceptibility zones cluster in the central and western wards, including Wards 15, 12, 8, and 13, where dense population, proximity to the Kulik River, and closeness to roads, temples, and government offices combine to make dumping hazardous. Low-susceptibility land, about 2.55 square kilometers, is concentrated in the southern and peripheral wards such as Wards 23, 24, and 25, which have lower population densities and fewer sensitive land uses. Moderate zones, covering about 4.39 square kilometers in wards like 1, 2, 10, and 18, represent transitional mixed-use areas where dumping might be tolerated only under strict conditions.

Translating these zones into policy, the researchers propose a three-tier regulatory framework: legal dumping zones in the low-susceptibility southern wards, conditional use zones in the moderate belt with mandatory buffer distances of 500 to 800 meters, environmental surveillance, and structured community participation, and outright prohibition in the high-susceptibility wards near the river and critical infrastructure. They also recommend monitoring land use and land cover through remote sensing, tracking vegetation stress via NDVI, NDWI, and land surface temperature, and enforcing waste transport route compliance to prevent overflow and unsafe handling. The authors acknowledge the framework’s limitations, noting that the model lacked information about actual site conditions, but argue that its outputs are sufficient to establish safe waste routes and enforce buffer distances. Because the approach relies on publicly available satellite data and standard open-source tools, the researchers say it can be adapted to virtually any fast-growing urban area facing the same dilemma between mounting waste and shrinking safe ground.

Subject of Research: Hybrid AHP and ensemble machine learning approach for optimising municipal dumpsite selection in Raiganj, India

Article Title: A hybrid AHP and ensemble machine learning based approach for optimising dumpsite selection in Raiganj municipality for sustainable urban waste management

Article References: Mondal, P., & Saha, S. (2026). A hybrid AHP and ensemble machine learning based approach for optimising dumpsite selection in Raiganj municipality for sustainable urban waste management. Discover Green Chemistry, 1(1), Article 5. https://doi.org/10.1007/s44509-026-00004-4

Image Credits: AI Generated

DOI: 10.1007/s44509-026-00004-4

Keywords: dumpsite selection, Analytic Hierarchy Process, Random Forest, Logistic Regression, Support Vector Machine, machine learning, GIS, solid waste management, Raiganj, West Bengal, susceptibility mapping, urban sustainability