Design and development of a low-cost hyperspectral imaging setup via a machine-learningbased approach

Abstract This article presents a potential solution for developing a low-cost hyperspectral imaging (HSI) setup by preserving pertinent information acquired from a conventional hyperspectral imaging setup. Conventional hyperspectral images (HSI) of three different types of leaves, gongura (Hibiscus sabdariffa), amaranthus (Amaranthus viridis), and banana (Musa acuminata) were acquired with 204 wavelengths/bands. The spectra are processed using linear discriminant analysis (LDA) to find a set of signature wavelengths for leaf classification. Afterwards, 20 visible range wavelengths (440 to 700 nm) were found to be incorporated into a low-cost setup involving a monochromator, beam steering elements, and a smartphone camera with associated machine learning (ML) classifiers. For building the datasets, we extracted 90 spectra from the images captured using the conventional HSI setup under the full spectra range (397 nm to 1003 nm). Similarly, 90 spectra were extracted from the images captured from the low-cost setup under the 20 signature wavelengths. For further experimentation, we split the datasets into Dataset-A, containing 70% of the total spectra, and the remaining 30% in Dataset-B for HSI as well as low-cost setup. Dataset- B was reserved to evaluate the robustness of the classifiers on an unseen dataset. LDA surpasses the other classifiers in leaves classification for HSI as well as low-cost setup. For the HSI setup, LDA achieved a 100% average score for performance matrices (classification accuracy, precision, Recall, and F1-score) on dataset A as well as on dataset B. Moreover, for the low-cost setup, LDA achieved 98.33% ± 4.99% classification accuracy, 98.89% ± 3.33% precision, 98.33% ± 4.99% recall, and 98.22% ± 5.33% F1-score on dataset A. Additionally, LDA achieved 96.29% classification accuracy, 96.67% precision, 96.29% recall, and 96.28% F1-score for dataset B. The promising results indicated that the low-cost set-up closely emulates the HSI system’s performance in terms of performance metrics, delivering a low-cost HSI system for various applications in the future.