Towards a new approach to estimate soybean yield at the field level

• Autor(a) principal: Wei, Marcelo Chan Fu
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/11/11152/tde-12022021-120048/
• topic: Glycine max; Glycine max; Agricultura de precisão; Componentes de produtividade; Image processing; Linear regression; Precision agriculture; Processamento de imagens; Regressão linear; Yield componentes

Soybean is one the most cultivated crops in the world. Looking to improve soybean production in a more sustainable way, it is imperative for farmers to make better management practices, improving the economic return and reducing the negative impact on the environment. Thus, yield data becomes an essential data layer that can support farmers to boost the management practices related to the soybean crop. Commonly, soybean yield data are obtained from combine harvester equiped with yield monitor or by agrometeorological models. Both present limitations regarding the data usage. For example, yield monitor data quality are affected by the sensor system that require data filtering process and agrometeorological yield models are limited to the amount of predictor variables required and its spatial and temporal resolution. Thus, aware of these limitations, it becomes an opportunity to investigate new methods to estimate soybean yield at the field level considering soybean yield and yield component data availability and the advances of technological applications in agriculture. The objectives of this study were to: (a) analyze the relationship among soybean yield and its components (number of grains - NG and thousand grains weight -TGW) in a worldwide range from available data and propose a yield estimation model based on yield components and (b) make an exploratory analysis on two dimensions (2D) methods to obtain data related to soybean yield from plants at the R8 phenological stage. Initially, it was conducted a literature review to gather soybean yield and its components data to compose the training dataset. Linear regression models based on soybean yield components were fitted on the training dataset and evaluated on a validation dataset composed of 58 samples collected at the field level. To conduct the exploratory analysis of image processing techniques on soybean, images were taken at the field level from a consumer-grade camera, then basic image processing techniques were applied followed by the application of a Boolean-based algorithm to detect the soybean components. As a result, it was generated three linear regression models: the first based on the TGW, the second based on NG and the third based on TGW and NG. The yield model based on NG presented the highest prediction accuracy, indicating that NG can be a potential yield component to be used as predictor variable. The exploratory analysis of the application of image processing techniques on RGB images provided potential results that support further investigation to improve image processing to gather NG data. Aiming to improve the detection of the yield components through image processing, it was found important steps that must be applied before application of the Boolean-based algorithm such as thresholding. In this study, it is proposed a new soybean yield estimation model relying on the use of one predictor variable (NG) and also, it is presented a potential image processing method that allows gathering NG data from soybean RGB images taken at the field level.