Employing broadband spectra and cluster analysis to assess thermal defoliation of cotton

• We examined thermal defoliation of cotton plants with multispectral data and clustering analysis.
• Clustering worked well in dividing the data into appropriate groups.
• The best cluster structure occurred for the normalized difference vegetation index.

Growers and field scouts need assistance in surveying cotton (Gossypium hirsutum L.) fields subjected to thermal defoliation to reap the benefits provided by this nonchemical defoliation method. A study was conducted to evaluate broadband spectral data subjected to unsupervised classification for surveying cotton plots subjected to thermal defoliation. Ground-based reflectance measurements of thermally treated and non-treated cotton canopies were collected at two study Sites (Site 1 and Site 2) with a handheld hyperspectral spectroradiometer. The hyperspectral data were merged into eight broad spectral bands: coastal blue (400–450 nm), blue (450–510 nm), green (510–580 nm), yellow (585–625 nm), red (630–690 nm), red-edge (705–745 nm), near-infrared (770–895 nm), and panchromatic (450–800 nm). Also, a broadband normalized difference vegetation index (NDVI) was created with the red (630–690 nm) and near-infrared bands (770–895 nm). For each study Site, two datasets were analyzed: (1) two-class case (thermally treated cotton observations and non-treated cotton observations) and (2) five-class case (thermally treated cotton observations and non-treated cotton observations and three additional classes created with the weighted average of the thermally treated cotton observations and non-treated cotton observations). The clustering algorithm referred to as CLUES (CLUstEring based on local Shrinking) was employed to automatically group the data into clusters without the user selecting the number of clusters. Cluster validation was determined with the average silhouette width; also accuracy was assessed with contingency matrixes. Clustering analysis worked well in dividing the data into appropriate groups, with the best cluster structure occurring for the NDVI. User’s and producer’s accuracies for the NDVI were greater than 86%, indicating an excellent classification. Findings support future endeavors to assess airborne and satellite-borne systems equipped with sensors sensitive to the wavelengths deemed useful in this study and unsupervised classification techniques that automatically determines the numbers of clusters to evaluate thermal defoliation of cotton fields.