Corn grain yield forecasting by satellite remote sensing and machine-learning models

• Autor(a) principal: Pinto, Antonio Alves [UNESP]
• Data de Publicação: 2022
• Outros Autores: Zerbato, Cristiano [UNESP], Rolim, Glauco de Souza [UNESP], Barbosa Júnior, Marcelo Rodrigues [UNESP], Silva, Luis Fernando Vieira da, Oliveira, Romário Porto de [UNESP]
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• DOI: 10.1002/agj2.21141
• Texto Completo: http://dx.doi.org/10.1002/agj2.21141; http://hdl.handle.net/11449/242103

This study aimed to evaluate the performance of six machine-learning models in forecasting corn (Zea mays L.) grain yield before harvest using, as input, variables in the models, some of the most-used vegetation indices (VIs) and spectral bands in the literature, as well as using data at 770 and 980 sum of degree days (SDD). The field study was carried out in a commercial area in the 2017–2018 and 2018–2019 harvests. Spectral data were obtained from Sentinel-2 satellite images and were used as input variables in the proposed models: artificial neural networks (ANN), k-nearest neighbors (KNN), random forest (RF), and support vector machine (SVM). The maximum R2 and minimum values of mean absolute error (MAE) and RMSE were 0.89, 0.33, and 0.42 t ha−1, respectively, for the RF algorithm using all input variables. The results obtained in the present study show that it is possible to predict corn grain yield 80 d before harvest using only VIs for the crop. Testing the various combinations of spectral bands and VIs resulted in obtaining the GREEN band and the VI global environment monitoring index (GEMI) as the best predictor variables in the present study. The use of more than one SDD did not improve the performance of the models tested. The models developed using data at 980 SDD obtained the best precision and accuracy performance both in the scenario with all model input variables and with the two best predictors. The KNN algorithm obtained the best performance in the precision and accuracy metrics for most of the scenarios studied in the present work.