Estimate of sugarcane productivity using machine learning algorithm from time series of WFI/CBERS-4 and WPM/CBERS-4A time series

Abstract

Brazil is the world's largest sugarcane producer, meaning productivity monitoring is crucial to enable mills to plan for future seasons and make decisions during the harvest. This results in time, labor, and resource savings. This study aims to leverage machine learning models and spectral data from the Wide Swath Multispectral and Panchromatic Camera (WPM/CBERS-4A) and the Wide Field Imager (WFI/CBERS-4) to estimate sugarcane productivity at various stages of crop development. The study focuses on 6229 sugarcane fields in the Araçatuba region of São Paulo. Productivity data for these fields from the 2020 to 2022 harvest, were provided by a partnering sugarcane mill. A time series of spectral data (bands and the vegetation indices) from the two sensors were used, combined with meteorological, water balance and agronomic data. From these variables, two distinct datasets were created, one for each sensor (WPM and WFI). Sixteen empirical models were developed for each dataset, each representing different stages of sugarcane development, both for plant cane (PC) and ratoon cane (RC), giving 32 models. The models were created iteratively, where each model was developed from the input data set of the previous and current month, allowing estimates throughout crop development along with choosing the most important variables. Data processing included sensor cross-calibration, cloud removal, vegetation index calculation, and integration of meteorological data. The models were trained for different phenological stages of the crops, considering variables such as precipitation, solar radiation, and the number of cuts. The Random Forest model was chosen for its robustness in dealing with large volumes of data, ability to capture complex relationships between variables, and resistance to overfitting. Additionally, the Nemenyi post-hoc test for mean comparison was applied. Upon analyzing the results, it was observed that for both the WFI and WPM models, the optimal stage to begin productivity estimations is from the 6th/7th month for plant cane and the 3.5th/4th month for ratoon cane. Before these time points, the predicted data statistically differed from those observed. Models using data from the complete crop development cycle yielded the best results, which achieved a Coefficient of determination (R²) of 0.63, a modified Willmott index (dₘₒd) of 0.70, and a root mean square error (RMSE) of 10.34 t.ha⁻¹. Thus, data from the WPM/CBERS-4A and WFI/CBERS-4 sensors integrated with additional data demonstrated a promising potential for predicting sugarcane productivity.

Plant phenotyping relevance

衛星センサー時系列と機械学習を用いてサトウキビ圃場の生産性（収量形質）を推定し、複数モデルを開発・評価しているため、単なる農業実験の routine 測定ではなく、植物形質推定ワークフローが中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.