Spectral reflectance reconstruction via extreme learning machine: Cross‐dataset evaluation and generalisation analysis

Abstract

This study was designed to reconstruct 31‐band spectral reflectance from CIE XYZ values using an extreme learning machine (ELM), and its within‐dataset and cross‐dataset performance was evaluated. The Sino dataset, containing 10,770 samples, was divided into an independent 20% subset for hyperparameter selection and an 80% subset for formal evaluation. By means of a full grid search over hidden‐layer size, activation function and regularisation coefficient, an ELM with 350 hidden neurons, radbas activation and a regularisation coefficient of 0 was selected. On the independent evaluation subset, a mean ΔE ₀₀ of 0.082 ± 0.002, an RMSE of 0.005 ± 0.000 and a spectral angle of 0.020 ± 0.001 rad were achieved by the optimised ELM. Among the seven reconstruction methods, RBF achieved the lowest reconstruction errors, whereas ELM required less computational time than RBF and BP while maintaining relatively low reconstruction errors. Cross‐dataset performance differed markedly with the training‐test direction. Sino‐trained models produced mean ΔE ₀₀ values of 7.936 ± 0.361 on Munsell and 6.829 ± 0.348 on Coloro, whereas Munsell‐ and Coloro‐trained models transferred to Sino with substantially lower colour‐difference errors. For the Sino‐trained models, large reconstruction errors occurred over broad wavelength ranges rather than being concentrated in a single spectral region. Using XYZ information under both D65 and illuminant A reduced the within‐dataset reconstruction errors, but similarly low errors were not obtained for all cross‐dataset directions. The MI _ab results showed poorer cross‐illuminant consistency for the transfer directions with larger spectral reconstruction errors.