Estimation of soybean yield from machine learning techniques and multispectral RPAS imagery

Abstract

Throughout the plant development, the physiological processes of growth, as well as the handling of water and fertilizers represent sources of variation in the spectral response of the crops. Thus, in the context of precision agriculture, information on the phenological stage of cultivation can be decisive in accurately estimating agronomic variables through the use of aerial images obtained by RPAS (Remotely Piloted Aircraft System) platforms. Therefore, the aim of this study was: i) to investigate the potential of using RPAS embedded with multispectral sensors in order to obtain data on a soybean irrigated field with a central pivot system ii) to identify the best stage of development of irrigated soybean in order to obtain multispectral images aiming to the yield prediction; iii) to test different vegetation indices (VIs) for estimating the yield of soybean crop using artificial neural networks. The treatments were established using different metric water potentials in the soil and the analysis was comprised by 30 sample units. Multispectral images were acquired through a Sequoia® camera aboard the Phantom 4® Pro platform, during seven phenological crop stages. In addition to the spectral bands, nine VIs were taken as predictors based on the response variables derived from the site survey. The selection of the best growth stage was based on the highest level of association, represented by the Spearman correlation coefficient between predictors and each response variable. The Multi-Layer Perceptron (MLP) algorithm was used to adjust the predictive model, whose performance was assessed in terms of training and testing. To have independent measurements of the performance of the proposed algorithm was used by k-fold cross-validation. The results obtained from the correlation between the image-derived data and the soybean yield indicate that the ideal monitoring window for irrigated soybean can be found at the end of the vegetative stage (V6). The selection of the phenological crop stage had a positive impact on the predictions. The MLP models showed a good adjustment and good generalization capability. In turn, grain yield was the most complex agronomic parameter for modeling, with correlations of 0.70 and 0.92 for training/testing, and excellent results in other statistical tests, reinforce the great combining capacity between remote sensing via RPAS and machine learning (ML) in applications aimed at precision agriculture. This approach offers a useful tool for evaluating grain yield in center-pivot-irrigated soybean crops and emphasizes the need of observation regarding the phenological crop stage as a factor of analysis in the prediction of grain yield via remote sensing.