Management zones and space-time prediction of soybean yield variability: machine learning techniques applied to soil physical quality parameters

• Ano de defesa: 2023
• Autor(a) principal: Pereira, Gislaine Silva
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: ACP; Agrupamentos; Aquacrop; Árvores de classificação; Classification trees; Clusters; CROPGRO; PCA; Random forest; Soja; Soybean
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/11/11152/tde-04012024-105156/

Methods and tools for precision agriculture are the key to ensuring increased soybean production. In this respect, knowledge of intra-field variability is the key to helping the agricultural producer in the decision-making process. Although methods for modelling production are based on regional conditions and/or agroecosystem models that do not represent local scales. The aim of this thesis is to use machine learning techniques to improve data quality for predicting yield at the management-zone level. The research was divided into three chapters that use techniques and methods focused on precision agriculture to validate the need to guarantee greater support to the farmer at the local level. The first chapter sought to use machine learning to improve the quality of data and the information from high-resolution sensors in generating management zones (MZs). In addition to validating the differences between and within MZs related to soil factors. The hypothesis of this first chapter was centred on the need to use principal component analysis (PCA) to improve the quality of MZ prediction based on observed data. The second chapter aimed to estimate soybean yield in each MZ over several years based on maps of soil water and crop development. One hypothesis for the chapter was the need to confirm the existence of the variability of intra-regional yield. The second hypothesis focused on testing the quality of near infrared reflectance (NIR) surfaces to represent crop development compared to using vegetation index (NDVI). The third hypothesis was that the machine learning technique Random Forest (RF) affords better quality yield prediction due to its efficiency in working with unbalanced data compared to the conventional method of multiple linear regression analysis (MLR). The aim of the third chapter was to understand the sensitivity of crop models (Aquacrop and CROPGRO) in estimating soybean yield at the management-zone level, especially as a function of available soil water. The hypothesis of this chapter was in the ability of crop models to show less variability when estimating yield based on the variations in soil water. The results of Chapter 1 showed that the PCA techniques afforded higher-quality clustering compared to the conventional method of normalization, besides ensuring greater stability in defining the number of MZs. Soil variables were fundamental for validating the specific characteristics of each region using the classification tree technique. The results of Chapter 2 showed the differences between digital soil water surfaces as a function of the MZs, demonstrating the importance of different management practices in each region, even at the local level. NIR reflectance improved quality predictions of soybean yield in each region compared to the use of NDVI. The RF method afforded higher-quality estimates compared to the MLR method. The results of Chapter 3 showed that the Aquacrop and CROPGRO models showed variable performance when estimating soybean yield in each zone in occurrence of wet and dry years. More studies should be carried out using crop models to predict soybean yield at local level. In this way, was possible to highlight the importance of evaluation on a local scale, with the use of machine learning methods and digital mapping to support precision agriculture. The use of MZs is the adequate to understanding the variability of soil and plant factors that will later influence planning for the localized use of inputs, impacting yield at same field. For future studies, the use of local sensors to continuously monitor variability of climate, soil and plant variability to improve precision of machine learning methods in agriculture.