Band selection pipeline for maturity stage classification in bell peppers: From full spectrum to simulated camera data

This paper describes a workflow for classifying the maturity of bell peppers using hyperspectral imaging and machine learning. The approach involves using spectral reflectance to determine the number of maturity stages, followed by a classification task using the optimal bands for accurate classification. The study explores a realistic scenario using simulated camera responses and investigates the use of real sensors with their spectral sensitivities and noise. Four classifier algorithms (RBFNN, PLS-DA, SVM, and LDA) were employed to predict the maturity stage based on spectral reflectance. The best results were achieved with the LDA algorithm, which was used in the optimization process for band selection. The optimized bands in the VISNIR range (400-1000 nm) were found to be [783.5, 844.1, and 905.4] nm, with an accuracy of 90.67% for spectral data. For camera responses with intermediate-level noise, the best bands were [760, 820, and 900 nm], achieving an accuracy of 81%. Overall, using three bands yielded satisfactory and practical results for real-world implementation.

Automated summary

This paper presents a workflow for classifying the maturity of bell peppers using hyperspectral imaging and machine learning. The approach utilizes spectral reflectance to determine the number of maturity stages and employs optimized bands for accurate classification, achieving satisfactory results.