<p>Rapid urbanisation and increasing vehicular and industrial emissions have intensified risks to public health and urban sustainability. Reliable forecasting of the Air Quality Index (AQI) is essential for enabling timely interventions and evidence-based environmental policymaking. This paper presents a sustainability-aware recommender framework built on a multimodal deep learning architecture that jointly exploits visual environmental context from urban scene images and structured pollutant measurements for AQI forecasting, with a focus on PM<sub>10</sub> concentrations. A pretrained ResNet50 backbone is employed as a visual encoder to extract spatially informative features, which are fused with a lightweight dense encoder operating on tabular pollutant and contextual features. To ensure transparency, the system integrates explainable AI mechanisms: Grad-CAM is used to highlight spatially salient regions (such as haze bands and traffic congestion), while SHAP values quantify the contribution of individual pollutants and meteorological variables to the predicted AQI. A rule-based risk-aware recommender layer maps predicted AQI levels to actionable health and exposure guidelines. Experiments on a real-world multimodal dataset comprising 4215 image–sensor pairs collected across multiple Indian and Nepalese cities during February–March 2023 yield a mean absolute error (MAE) of 6.77 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\mu \text {g/m}^3\)</EquationSource> </InlineEquation>, a root mean square error (RMSE) of 8.22 <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\mu \text {g/m}^3\)</EquationSource> </InlineEquation>, and an <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(R^{2}\)</EquationSource> </InlineEquation> score of 0.98, outperforming strong tabular-only baselines such as XGBoost, support vector regression, and random forests. Robustness is further evaluated through location-aware 5-fold cross-validation, violin-plot analysis of fold-wise error distributions, and statistical significance testing against the best baseline model (<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(p < 0.001\)</EquationSource> </InlineEquation>). The resulting explainable and deployment-conscious framework supports scalable air quality forecasting and risk-aware recommendation in smart city platforms, facilitating proactive and sustainable urban air quality management.</p>

Explainable multimodal deep learning system for sustainable urban air quality forecasting and risk-aware recommendations

期刊论文
发表日期：

2026年3月28日

查看全文

Basamma Umesh Patil,
Chetan R,
Harshitha V

摘要

Rapid urbanisation and increasing vehicular and industrial emissions have intensified risks to public health and urban sustainability. Reliable forecasting of the Air Quality Index (AQI) is essential for enabling timely interventions and evidence-based environmental policymaking. This paper presents a sustainability-aware recommender framework built on a multimodal deep learning architecture that jointly exploits visual environmental context from urban scene images and structured pollutant measurements for AQI forecasting, with a focus on PM₁₀ concentrations. A pretrained ResNet50 backbone is employed as a visual encoder to extract spatially informative features, which are fused with a lightweight dense encoder operating on tabular pollutant and contextual features. To ensure transparency, the system integrates explainable AI mechanisms: Grad-CAM is used to highlight spatially salient regions (such as haze bands and traffic congestion), while SHAP values quantify the contribution of individual pollutants and meteorological variables to the predicted AQI. A rule-based risk-aware recommender layer maps predicted AQI levels to actionable health and exposure guidelines. Experiments on a real-world multimodal dataset comprising 4215 image–sensor pairs collected across multiple Indian and Nepalese cities during February–March 2023 yield a mean absolute error (MAE) of 6.77 \(\mu \text {g/m}^3\) , a root mean square error (RMSE) of 8.22 \(\mu \text {g/m}^3\) , and an \(R^{2}\) score of 0.98, outperforming strong tabular-only baselines such as XGBoost, support vector regression, and random forests. Robustness is further evaluated through location-aware 5-fold cross-validation, violin-plot analysis of fold-wise error distributions, and statistical significance testing against the best baseline model ( \(p < 0.001\) ). The resulting explainable and deployment-conscious framework supports scalable air quality forecasting and risk-aware recommendation in smart city platforms, facilitating proactive and sustainable urban air quality management.