Data-driven machine learning for pattern recognition supports environmental quality prediction for irrigated rice in Brazil

The sustainability of irrigated rice (Oryza sativa L.) production systems in Brazilian tropical region highly depends on the success of developing stable cultivars. To achieve this goal, many steps in product development must address the environmental variability and genotype by environment interactions (GE), which makes difficult the design and development of local-specific adapted cultivars. Thus, the adoption of new strategies for characterizing environmental-phenotype relations are the key for optimizing this process. In addition, it could also benefit post-breeding stages of seed production. To overcome this situation, we implemented a data-driven approach to link environmental characterization to yield clustering using historical data (1982-2017, 31 locations, 471 genotypes), 42 envirotyping covariables and machine learning (ML), combining two unsupervised (K-means and decision tree models, DTC) algorithms. Additionally, linear mixed models (LMM) were applied to explore the relations between the outcomes of our approach and GE analysis for irrigated rice yield in Brazilian tropical region. Four environments were identified: Very Low Yield (1.7 Mg.ha-1), Low Yield (5.1 Mg.ha-1), High Yield (7.2 Mg.ha-1), and Very High Yield (9.0 Mg.ha-1), considering all genotypes and regions. Our approach allows the prediction of environments (yield clusters) for a diverse set of growing conditions and revealed geographic and climatic causes of environmental quality, which differ according to each region and genotype group. From the LMM analysis, we found that the current relation between genetics (G), environmental variation (E), and GE for rainfed rice in Brazil is 1:6:2, but when we introduced our data-driven clusters (ME), the ratio decreased to 1:5:1. Consequently, the selection reliability for local adaptability across an extensive region increases. Our approach helps to identify mega-environments in Brazil that could be used as a target population of environments (TPE) of breeding programs. Additionally, it helps to identify more productive and stable seed production fields. HighlightsO_LIA nationwide environmental characterization and its relation to the genotype by environment interaction (GE) for grain yield of rainfed rice growing regions in Brazil. C_LIO_LIA data-driven approach capable to identifying clusters of yield levels and a machine learning approach to relate those clusters with environmental typologies. C_LIO_LIUnrevealed geographic and climatic causes of environmental quality for a group of genotypes or cultivar-specific predictions. C_LIO_LIThe strategy benefits diverse stages of breeding (multiple environmental trial analysis) and post-breeding (selection of fields for seed production) as an alternative approach to reduce costs and support decisions on cultivar planting locations. C_LI