Predição de produtividade de trigo por meio de dados espectrais e altura estimada da planta obtidos por meio de aeronave remotamente pilotada

Abstract

The field monitoring of wheat crop, from sowing to harvest, is essential to control any problems that may occur during crop development. The wheat yield prediction is a way to provide the farmers information about how the crop will perform, helping to identify possible problems in advance. The plant height can also be an important indicator of yield, since it can be related to the crop performance. The extraction of this information in the field can be impracticable when performed manually and the use of technologies that implement the Remote Sensing (SR) concepts, such as remotely piloted aircraft (RPA), can be an alternative to supply this need. Considering the recent appearance of RPA in agricultural applications and the limited number of studies involving plant height estimation using digital models, this study aimed to obtain a numerical model for the wheat yield prediction based on digital values of spectral bands and estimated plant height values extracted from RPA images. The experimental areas of this study were located in Ponta Grossa, Parana State, Brazil, and were sown with wheat in 2018 and 2019. The flights over the areas in 2018 were carried out by an RPA model eBee, using RGB and NIR cameras in different resolutions. The flight over the area in 2019 was carried out by an RPA model Phantom using an RGB camera with 5 cm/px resolution. The images obtained were processed in the Pix4D software, which uses algorithms based on Structure from Motion (SfM) and allows the creation of digital surface (DSM) and elevation (DEM) models. From the calculation of the difference between these models it was possible to estimate the height of the objects above the ground. Manual measurements of plant height and grain yield in the field were also performed in order to validate the models. To create the prediction models, three Data Mining (MD) algorithms were selected for use in preliminary tests. SMOReg, an algorithm based on Support vector machine (SVM) for applications involving regression, presented the best results and was selected to continue the tests. The results based on the plant height estimation indicated that the NIR camera proved to be more accurate for this purpose, obtaining a considerably lower average error of estimate compared to that of the RGB camera. Tests performed with the combination of digital RGB data and plant height estimated by the models reached a correlation index (r) of up to 0.97. Therefore, a prediction model that showed a high correlation with wheat grain yield was obtained in our study, using only SR data.