Identifying soil drivers of rice productivity under fly ash and organic amendments using explainable machine learning

Declining soil quality and nutrient imbalances constrain rice productivity in tropical acidic soils. The agricultural reuse of fly ash (FA), an industrial by-product, offers potential as a soil amendment when combined with organic inputs, yet mechanistic understanding of its effects on soil-yield relationships remain limited. Traditional statistical methods often fail to decode non-linear soil-yield relationships, necessitating advanced machine learning (ML) approaches. A field experiment evaluated the integrated effect of FA (10-40 t ha^-1), FYM (5 t ha^-1), and NPK effects on soil physio-chemical and biological properties and identified key soil predictors driving rice productivity using explainable machine learning. The FA40 + FYM + NPK treatment achieved the highest grain yield (54.0 q ha^-1), outperforming NPK alone by 38.5 %. This treatment improved soil porosity (45.5 %), water-holding capacity (37.8 %), available nitrogen (212.9 kg ha^-1), available phosphorus (19.6 kg ha^-1), and microbial enzyme activities, including urease (22.9 μg NH₄⁺-N g^-1 hr^-1) and β-glucosidase (15.1 μg pNP g^-1 hr^-1). Machine learning interpretation revealed β-glucosidase, organic carbon, urease, available phosphorus, and clay content as dominant predictors of yield variation. Conditional partial dependence plots revealed synergistic interactions between β-glucosidase and organic carbon, and between urease and available phosphorus, indicating that carbon turnover and nutrient mineralization jointly regulated yield response. These findings demonstrate that the combination of FA (20-40 t ha^-1) with FYM and NPK can improve soil functionality and sustain rice productivity. Explainable modelling provides mechanistic insight for advancing soil health assessment and fertilizer strategies in acidic agroecosystems.