Comprehensive growth monitoring index using Sentinel-2A data for large-scale cotton production

Problem

Timely and accurate plant growth monitoring is crucial for precision crop management. Traditional remote sensing methods use a single agronomic parameter to evaluate crop growth status (GST), limiting accuracy.

Objective

To develop a comprehensive growth monitoring index (CGMI) based on multiple parameters.

Methods

A two-year field experiment in the Mosuwan Reclamation Region of Xinjiang, China was conducted to collect parameter characterization data for cotton growth, including leaf area index, canopy chlorophyll content, above-ground biomass, and boll numbers, and their contributions and interrelationships in relation to yield were analyzed. Entropy and game theory weighting methods were used to establish the CGMI_EWM and CGMI_GT, and a sequential forward selection algorithm (SFS) was used to screen the most effective remote-sensing monitoring feature variables for the different reproductive stages. Partial least squares regression (PLSR), random forest (RF), and support vector regression (SVR) were used to develop an optimal model to comprehensively monitor cotton growth and draw a spatial distribution map.

Results

CGMI_EWM and CGMI_GT could effectively reflect GST. The correlation between CGMI_GT and yield based on a game theory combination weighting method was significantly higher than that between yield and each agronomic parameter. The correlation between CGMI_GT and yield (r = 0.75) was slightly higher at the initial boll stage than that of CGMI_EWM (r = 0.73), whereas at the initial boll-opening stage, the correlation between CGMIGT and yield (r = 0.74) was significantly higher than that of CGMI_EWM (r = 0.63). The weight coefficients used to construct the CGMI_GT exhibited stable performance in different years. Feature variables were selected to monitor the comprehensive growth of cotton at different stages based on the SFS algorithm. PLSR, RF, and SVR were used to estimate CGMI_GT. The RF algorithm had the best estimation performance in both the initial boll and initial boll-opening stages (R² = 0.63, root mean square error (RMSE) = 0.086, RE = 19.8 % vs. R² = 0.56, RMSE = 0.107, RE = 24.1 %). A comprehensive cotton growth distribution map in the Mosuwan Reclamation Region was drawn using the optimal model, and growth was comprehensively evaluated. Areas with good cotton growth were concentrated in the north, and there was a decreasing trend from north to south.

Conclusions

We provide a new comprehensive evaluation tool for cotton growth status large-scale, real-time monitoring.

Significance

Our results promote differentiated management, improve crop yield prediction accuracy, and aid in the formulation of cotton price strategies.