Applying Random Forest Regression Algorithm in EVI Data for Soybean Yield Estimation.

Retrieving soybean yield information is important for agricultural management. However, trying to obtain this information can be challenging especially in large areas. Remote sensing information can be helpful, since large areas can be covered with satellite information. In addition, special indexes can be related to vegetative growth and canopy health. One example is the Enhanced Vegetative Index (EVI) from the Moderate Resolution Imaging Spectroradiometer (MODIS). Regression techniques can be used to identify EVI relationship with soybean yield. The Random Forest algorithm was used as a regression tool to translate EVI (Terra & Aqua, 8 days of temporal resolution) to soybean yield in the Paraná state (Southern Brazil). Observed yield from 86 farm areas during crop-year 2013-2014 were used, totaling 1253 MODIS pixels (250 by 250 m). From that, 70% (878) were randomly selected for training the model and 30% (375) were used for evaluation through MAE, EM, RMSE, MAPE and Willmott concordance index. The EVI information was related with important phenological dates such as sowing, vegetative peak, and harvest to create different variables. Each pixel has a value of EVI on each specific date (e.g. on the sowing date, at vegetative peak and at other times after sowing) totaling 39 variables. A recursive feature selection (RFE) was used to define the five best variables for one regression and another regression with all variables was also made. For the regression the Random Forest algorithm in R was used. The results shows that using the five best predictors MAE = 285.6, ME 18.8, RMSE = 423, MAPE = 10.6% and Dr = 0.6839. For all predictors MAE = 238.1, ME = 20.5, RMSE = 359, MAPE = 8.4% and Dr = 0.7365. This shows that it is possible to obtain low error yield estimation for soybean using all predictors.