Hyperspectral reflectance and machine learning for multi-site monitoring of cotton growth

Abstract

Hyperspectral measurements can help with rapid decision-making and collecting data across multiple locations. However, there are multiple data processing methods (Savisky-Golay [SG], first derivative [FD], and normalization) and analyses (partial least squares regression [PLS], weighted k-nearest neighbor [KKNN], support vector machine [SVM], and random forest [RF]) that can be used to determine the best relationship between physical measurements and hyperspectral data. In the current study, FD was the best method for data processing and SVM was the best model for predicting average cotton (Gossypium spp. Malvaceae) height and nodes. However, the combination of FD and RF were best at predicting cotton leaf area index, canopy cover, and chlorophyll content across the growing season. Additionally, results from models developed by both SVM and RF were closely related to pseudo-CHIME satellite wavebands, where in-situ hyperspectral data were matched to the spectral resolutions of a future hyperspectral satellite. The information and results presented will aid producers and other members of the cotton industry to make rapid and meaningful decisions that could result in greater yield and sustainable intensification.

Keywords

• Remote sensing
• Cotton
• SVM
• Random forest
• CHIME