Predicting coffee yield based on agroclimatic data and machine learning

Climate directly and indirectly influences agriculture, being the main responsible for low and high yields. Prior knowledge on yield helps coffee farmers in their decision-making and planning for the future harvest, avoiding unnecessary costs and losses during the harvesting process. Thus, we sought to predict coffee yield with regressive models using meteorological data of the state of Parana, Brazil. This study was carried out in 15 localities that produce Coffea arabica in this Brazilian state. The climate data were collected using the NASA/POWER platform from 1989 to 2020, while the data of arabica coffee yield (bags/ha) were obtained by CONAB from 2003 to 2018. The Penman-Monteith method was used to calculate the reference evapotranspiration and the climatological water balance (WB) was calculated based on Thornthwaite and Mather (1955). Multiple linear regression was used in the data modeling, in which C. arabica yield was the dependent variable and air temperature, precipitation, solar radiation, water deficit, water surplus, and soil water storage were the independent variables. The comparison between the estimation models and the actual data was performed using the statistical indices RMSE (accuracy) and adjusted coefficient of determination (R(2)adj) (precision). Multiple linear regression models can predict arabica coffee yield in the state of Parana 2 to 3 months before harvest. The maximum air temperature is the climate element that most influences coffee plants, especially during fruit formation (March). Maximum air temperatures of 31.01 degrees C in March can reduce coffee production. Wenceslau Braz, Jacarezinho, and Ibaiti presented the highest yields, with mean values of 32.5, 29.9, and 29.3 bags ha(-1), respectively. The models calibrated for localities that have Argisol had the highest mean accuracy, with an RMSE of 2.68 bags ha(-1). The best models were calibrated for Paranavai (Latosol), with an RMSE of 0.78 bags ha(-1) and R(2)adj of 0.89, and Ibaiti (Argisol), with RMSE and R(2)adj values of 3.09 bags ha(-1) and 0.83, respectively. Paranavai has a mean difference between the actual and estimated coffee yield of only 0.86 bags ha(-1). The highest deviations were observed in Wenceslau Braz (9.17 bags ha(-1)) and the lowest deviations were found in Paranavai (0.86 bags ha(-1)). The models can be used to predict arabica coffee yield, assisting the planning of coffee farmers in the northern region of the state of Parana.

Automated summary

Climate has a direct and indirect impact on agriculture, particularly on coffee yield. This study used regression models and meteorological data to predict the yield of Coffea arabica in Parana, Brazil. The research found that maximum air temperature is the most influential climate element on coffee plants during fruit formation. A multiple linear regression model was developed to forecast coffee yield 2 to 3 months before harvest.