Reshaping Hyperspectral Data into a Two-Dimensional Image for a CNN Model to Classify Plant Species from Reflectance

Abstract

Leaf-level hyperspectral-based species identification has a long research history. However, unlike hyperspectral image-based species classification models, convolutional neural network (CNN) models are rarely used for the one-dimensional (1D) structured leaf-level spectrum. Our research focuses on hyperspectral data from five laboratories worldwide to test the general use of effective identification of the CNN model by reshaping 1D structure hyperspectral data into two-dimensional greyscale images without principal component analysis (PCA) or downscaling. We compared the performance of two-dimensional CNNs with the deep cross neural network (DCN), support vector machine, random forest, gradient boosting machine, and decision tree in individual tree species classification from leaf-level hyperspectral data. We tested the general performance of the models by simulating an application phase using data from different labs or years as the unseen data for prediction. The best-performing CNN model had validation accuracy of 98.6%, prediction accuracy of 91.6%, and precision of 74.9%, compared to the support vector machine, with 98.6%, 88.8%, and 66.4%, respectively, and DCN, with 94.0%, 85.7%, and 57.1%, respectively. Compared with the reference models, CNNs more efficiently recognized Fagus crenata, and had high accuracy in Quercus rubra identification. Our results provide a template for a species classification method based on hyperspectral data and point to a new way of reshaping 1D data into a two-dimensional image, as the key to better species prediction. This method may also be helpful for foliar trait estimation.