How Soil Quality Affects Long-Term Rice Productivity

Sustaining rice productivity in intensive rice-rice systems requires comprehensive soil management, with diagnosis of key soil physical, chemical, and biological indicators that need attention. In a 16-year long-term experiment (established in 2005-06 and ongoing) of the irrigated double rice system of Eastern India, we investigated the effect of key soil drivers on rice productivity.

The experiment assessed the effect of control (no N fertilizer application), imbalanced fertilization (N/NP/PK), balanced and recommended NPK and 150% NPK, NPK with lime, micronutrient additions (Zn with/without S or B), and integrated nutrient management with FYM (with/without lime), Composite surface soil samples (0-15cm) were collected after harvest of the 32^nd rice season for evaluation of soil physical, chemical, and biological properties. Rice grain yield after the 32^nd season was recorded at 14% grain moisture.  

To identify key soil drivers, an interpretable machine learning framework was used, specifically a conditional random forest-based yield model, permutation-based variable importance, and accumulated local effect (ALE) plots. The model described the yield variability very well (mean RMSE 305 kg ha^-1, R² 0.88, MAE 254 kg ha^-1). Variable importance screening highlighted total K, protease, and urease activities, as well as permanganate-oxidizable carbon (POC), as dominant predictors. ALE-based effect sizes suggested these properties accounted for ~400 (total K), ~250 (protease), ~200 (urease), and ~140 (POC) kg yield variability.

Overall, the results indicate that potassium dynamics are a primary constraint in intensive rice-rice systems, with risks associated with continuous K mining, and emphasize the importance of routine monitoring of biological activity indicators for long-term sustainability.

Keywords: Conditional random forest; Soil quality index (SQI); Long-term fertilizer application; K-dynamics; Soil enzymes; Cattle manure