Characterization of common bean production regions in Brazil using machine learning techniques

Abstract

PROBLEM

Understanding the interactions between genotype, environment, and management is crucial for guiding the development of new cultivars and defining strategies to maximize yield under specific environmental conditions.

OBJECTIVE

This study aimed to classify and characterize homogeneous production regions for common beans in Brazil by leveraging simulated yield data and machine learning techniques. The goal was to identify the environmental factors that define these homogeneous regions and to develop a spatiotemporal sowing calendar for rainfed (wet and dry seasons) and irrigated (winter season) production.

METHODS

The CSM-CROPGRO-Dry Bean model was used to simulate the yield of common beans in Brazilian municipalities during the wet (rainfed) season (sowing between August and December - 275 municipalities), dry (rainfed) season (sowing between January and April - 251 municipalities) and winter (irrigated) season (sowing between April and July - 59 municipalities), utilizing soil data, daily climate data (1980 to 2016), management information (sowing date and irrigation/rainfed), and genetic coefficients to reflect the performance of the BRS Estilo cultivar related to phenology, growth and yield components. To create homogeneous environmental groups and associate them with specific environmental features, we applied machine learning techniques, including K-means clustering and decision tree analysis.

RESULTS

According to the results, we identified three distinct homogeneous regions — high, medium, and low yields — for each cultivation season (wet, dry, and winter). During the wet season, regions with yields between 2326 and 3500 kg ha⁻¹ were classified as high-yield, those between 1404 and 2325 kg ha⁻¹ as medium-yield, and those between 500 and 1403 kg ha⁻¹ as low-yield. In the dry season, high-yield regions had yields ranging from 2492 to 3500 kg ha⁻¹, medium-yield regions from 1484 to 2491 kg ha⁻¹, and low-yield regions from 500 to 1483 kg ha⁻¹. For the winter season, high-yield regions achieved yields between 2972 and 3500 kg ha⁻¹, medium-yield regions between 2252 and 2971 kg ha⁻¹, and low-yield regions between 634 and 2251 kg ha⁻¹. For rainfed seasons (wet and dry), the water stress (WSPD) had a greater impact on yield than air temperature and global solar radiation. While in the winter season, air temperature was the most relevant factor. Overall, in the wet season, delayed sowing contributed to increased yield, especially in the state of Paraná. In the dry season, delayed sowing caused a reduction in yield, particularly in the Midwest and Southeast regions. In the winter season, yield varied less significantly between sowing dates, except in the state of Mato Grosso, where harmful increases in air temperature were observed in the later months.

CONCLUSIONS

The integration of crop simulation models with machine learning tools is valuable for defining and characterizing homogeneous regions for common bean production. This approach has identified three distinct yield regions—high, medium, and low—for each crop season (wet, dry, and winter). By distinguishing these regions, this methodology supports breeding programs in developing cultivars optimized for specific environments and provides insights into how environmental factors influence crop performance. Additionally, it helps optimize sowing dates to align with favorable conditions, particularly during the wet and dry seasons, thereby contributing to reduced yield losses.

IMPLICATIONS

The classification and characterization of homogeneous production regions helps to better understand the genotype (G) x environment (E) x management (M) interactions and to adjust sowing dates to more favorable conditions, especially during the wet and dry seasons, contributing to reducing yield losses. Despite the limitations of crop modeling—such as not accounting for biotic factors and waterlogging—addressing water and air temperature stress presents an even greater challenge. In this context, crop modeling plays a crucial role in identifying effective adaptation strategies.