Use of supervised and unsupervised approaches to make zonal application maps for variable-rate application of crop growth regulators in commercial cotton fields

Abstract

Abstract Background Zonal application maps aim to represent field variability by means of variables of interest that can further be translated into management practices. Zonal application maps for crop growth regulators in cotton under variable-rate strategies are commonly based exclusively on vegetation index variability. However, saturation by vegetation indexes cannot be avoided by multispectral imagery and dense crop vegetation areas. The objective of this study was to compare zonal application maps for crop growth regulator application under variable-rate conditions via two approaches: (i) relying on field-measured crop height and (ii) not relying on field-measured crop height. During the agricultural season, we developed zonal application maps using an unsupervised framework, representing local variability based on field-collected data, satellite imagery data, soil texture and phenology. Posteriorly, using data from 3 agricultural seasons, we developed a supervised framework to predict plant height by using a machine learning algorithm based on remote sensing and phenology data to test the development of the same maps without field data on plant height. Results The results indicated good performance of the machine learning model for predicting plant height, but these predictions presented much lower variability than that found in field conditions. Ultimately, the predictions went through the same unsupervised process to construct maps, but without requiring field-measured data. We tested both approaches in three fields on two different dates each. Fields with intense textural variability provided the highest compatibility between the maps of both approaches. Conversely, the lowest compatibility was found in fields with lower soil textural variation. Conclusion There is potential for the use of predictive modelling to assist in the construction of zonal application maps, but the adherence to real patterns of crop growth variability found in the field is highly variable and must be assessed on a case-by-case basis.

Plant phenotyping relevance

リモートセンシングと機械学習により綿花の草丈という明示的な植物形質を推定し、圃場測定値との変動性を比較・検証しているため、管理マップ作成への応用を含む実質的な形質推定手法研究である。

Code and data availability

The paper's phenotype datasets (crop height, soil texture, satellite imagery) and analysis code are not publicly available; the authors state data may be obtained from the corresponding author on reasonable request. The only URLs in the text are generic library/references (SciPy, BayesianOptimization, Planet specs), no

No evidence-backed public reproduction asset is currently recorded.