Field prediction of calorific value and ash content of shrub biomass energy using Vis-NIR canopy spectra and deep...

In the field, it is important to quickly assess the calorific value and ash content of large-scale shrub biomass energy materials. In this work, multi-scale attention-enhanced one-dimensional convolutional neural network (MSA-1DCNN) and conditional Wasserstein generative adversarial network (WGAN) were employed to predict them using field canopy visible-near-infrared spectra. The key WGAN parameters, including target-value perturbation magnitude and generated sample size, were also compared. Results showed that MSA-1DCNN improved the prediction-set R_p² of calorific value from 0.290 to 0.650 and reduced RMSEP from 0.238 to 0.167 compared with the baseline 1D-CNN. WGAN-based augmentation further improved deep learning prediction performance. The mean prediction-set R_p², RMSEP, and RPD were 0.718, 0.150, and 1.912 for calorific value and 0.842, 0.031, and 2.558 for ash content, respectively. It provided technical support for rapid, non-destructive, large-scale evaluation of combustion performance in the field using small sample specimens.