A team from the Centre for Global Change at Sol Planje University presents LSTM-GAM-xAI, a hybrid deep learning and generalized additive model enhanced with LIME explainability and causal analysis. It forecasts concentrations of PM2.5, PM10, O₃, NO₂, NO, NOₓ, SO₂, and CO across 5- and 10-day timesteps with lower MSE than benchmarks, for improved regional air quality management.

Key points

• Integrates LSTM deep learning with a generalized additive model layer to capture nonlinear and temporal pollutant dynamics.
• Employs LIME post-hoc explainability to quantify feature contributions (e.g., NO₂, PM₂.₅) for each air pollutant forecast.
• Validates on synthetic Kimberley datasets across 5- and 10-day timesteps, outperforming LSTM, BiLSTM, GRU, BiGRU, 1DCNN, Random Forest, and XGBoost by lowest MSE.

Why it matters: This hybrid explainable AI framework sets a new standard for accurate, interpretable air quality forecasts, empowering data-driven environmental policy and health protection.