Forecasting gaps in sugarcane fields containing weeds using low-resolution UAV imagery based on a machine-learning approach

Abstract

Effective gap assessment is crucial for guiding sugarcane farmers in decisions about replanting versus maintaining ratoons. This study explores the use of low-resolution multispectral aerial imagery to enhance cost-efficiency and field management practices. Reflectance images captured during the germination phase were employed to develop predictive models, assessing five machine learning algorithms for their effectiveness in detecting sugarcane in fields with unmanaged weed populations. The optimal buffer distance for predicting canopy size during the tillering phase was identified, and this model was applied to sugarcane areas during germination. Gap identification was achieved by intersecting buffered sugarcane areas with planted rows. The Linear Discriminant Analysis (LDA) model emerged as the most effective, utilizing reflectance bands from the red, green, blue, and red-edge spectra, and achieving an accuracy of 84%. Notably, the blue reflectance band proved particularly important for distinguishing between sugarcane and non-sugarcane classifications. The gap detection model achieved a mean absolute error of 6.19%. These findings provide valuable insights for farmers, sugar mills, service providers, and other stakeholders, enabling informed decision-making regarding ratoon management. This research supports the strategic allocation of machinery and labor, thereby enhancing operational efficiency in alignment with the planting season.