Non-destructive estimation of biomass characteristics: Combining hyperspectral imaging data with neural networks

Hyperspectral image analysis is a quick and non-destructive way to determine the physical and chemical properties of odorous biomasses and feedstocks. This research investigated the feasibility of predicting characteristics using integrating hyperspectral imaging (HSI), principal component analysis (PCA), and artificial neural network (ANN). Further, the potential of bio-H2 production was studied by integrating these methods and structural equation modeling (SEM). Using PCA, we found that the most significant spectra were 575 nm, 602 nm, 638 nm, 737 nm, 882 nm, and 950 nm (within the 400–950 nm range). While the ANN model performed well in predicting total phenolic compounds and chemical oxygen demand, it performed poorly in predicting total carbohydrates, cellulose, and hemicellulose. The ANN model's R2 and RMSE for predicting bio-H2 production were 0.98 and 0.38, respectively, indicating high accuracy for the ANN model. The causal relationships among the parameters were determined using SEM (R2 > 0.92). As found, 575 nm and 900 nm spectra were discovered to had significant positive effects on cellulose content and bio-H2, and 602 nm and 882 nm spectra had significant adverse effects on bio-H2 production and positive effects on total phenolic compounds. The results confirmed that the integrated method of HSI-PCA-ANN-SEM was completely successful for studying the potential of bio-H2 production.