A machine learning approach is effective to elucidate yield-limiting factors of irrigated lowland rice under heterogeneous growing conditions and management practices

In the context of the heterogeneous field conditions for lowland rice production in sub-Saharan Africa, yield is adversely affected by a diverse array of complexly related factors, including environmental conditions, soil fertility, and fertilizer management, making it difficult to unravel the factors with greatest impact. In the central highlands of Madagascar, yield is constrained by low temperatures, insufficient rainfall, phosphorus deficiency, inadequate fertilizer input, and suboptimal agricultural practices. Identifying the driving factors by interpreting reliable predictive algorithms for the yield will greatly improve the optimization of rice productivity. In this study, we aimed to apply machine learning to comprehensively elucidate the complex yield-limiting factors for lowland rice production across the central highlands of Madagascar. A machine learning algorithm was applied to comprehensively evaluate the factors affecting yield using a total of 578 data points, i.e., P-applied and non-P-applied plots in 289 farmers’ fields, consisting of their yield, weather, soil, and crop management across the central highlands of Madagascar. There were large variations in yield ranging from 0.07 to 10.26 t ha−1. Machine learning explained the large yield variation well, especially in the Random Forest model, which showed the highest yield prediction accuracy. As relevant factors, high rice yields were associated with high growing degree days and high cumulative rainfall around 2 weeks before and after the heading date. Then, earlier transplanting in high elevation sites gave the optimum growing degree days, and the use of greater shoot-to-root ratio of seedlings during transplanting has the potential to enhance rice yields in the lowland rice system. Machine learning was significantly effective in identifying the individual effects of various factors controlling rice-yields under the heterogenous growing conditions of lowland-rice grown by smallholder farmers in the central highlands of Madagascar. Further accumulation of crop management datasets will be needed in order to untangle the multifactorial components that reduce crop yields, even for small-scale farms. •Machine Learning (ML) reveals field-scale yield limiting factors.•Key growing degree days (GDDheading) impacts yield.•Early date of transplanting boosts GDDheading for better yield and alleviates cold damage.•Transplanting seedlings with greater shoot-to-root ratio positively affects yields.