Predictive framework of plant height in commercial cotton fields using a remote sensing and machine learning approach

Abstract

Cotton is an important crop in the Brazilian agricultural business. This crop management has high technological and economic demand, thus the search for the efficient use of inputs such as crop growth regulator (to balance vegetative and reproductive growth) has become important in applied research areas such as precision agriculture. The objective of this work was to evaluate a predictive framework for cotton height based on field and remote sensing data and machine learning techniques, since this is an important variable in management planning of crop growth regulators on large scale cotton production. Field data from two agricultural seasons (2020/2021 and 2021/2022) from 16 fields and approximately 2800 hectares of commercial planted area were used as input along with Planet's satellite imagery. Machine learning techniques were used to answer the following main questions: (i) Can algorithms accurately predict crop height using two entire seasons of this data and remote sensing? (ii) Which remote sensing data and machine learning algorithm presents the best performance in the predictions? The algorithms could predict cotton height for the entire season with R-squared ranging from 0.31 to 0.87. Considering the main phenological phases, R-squared of the predictions ranged from -0.2 to 0.83. The inclusion of days after emergence showed an improvement in the performance of all models, decreasing the mean absolute error. For the main phenological phases, the best model performances were found from emergence to first flower appearance. Difference vegetation index and near infrared bands were the best predictors, considering the entire season and phenological phases. Random forest regressor presented the best overall performance, considering R-squared and mean absolute error. The use of two variables (remote sensing and time) proved to be sufficient to achieve one of the best performances.