Transparent non-destructive estimation of moisture content in vegetables using hyperspectral imaging, multi-target regression pipeline approach and explainable AI

Abstract The estimation of moisture content (MC) has become a significant part of assessing the nutrient value, freshness, and shelf life of post-harvested fruits and vegetables. Hence, early intervention can help maintain quality and reduce postharvest spoilage. However, manual assessment of MC requires skilled manpower and specialised equipment, which makes it costly and prone to errors. This study presents an innovative approach that integrates explainable AI (XAI) and a machine learning model to overcome such challenges. The study develops the multi-target regression (MTR) pipeline for simultaneously automated MC estimation in various agricultural samples. Hyperspectral imaging (HSI) data of broccoli, leek, and mushrooms were utilised for experimentation purposes. The MTR pipeline comprises three stages: classification, feature extraction, and regression. Random forest classifier, Partial least square, and Stacking regression models (SRM) were used for classification, feature extraction, and regression purposes. The study used 788 VIS-NIR reflectance spectra obtained from the HSI data of broccoli, leek, and mushrooms. The proposed MTR pipeline was evaluated using the coefficient of determination (R 2 ), root mean squared error (RMSE), mean squared error (MSE), and mean absolute error (MAE) scores. The overall performance score of the MTR pipeline for MC prediction on unseen datasets in the VIS to NIR range was 90.17 R 2 , 0.07 MSE, 1.91 RMSE, and 1.33 MAE. The individual proposed SRMs achieved R 2 scores of 81.31 ± 7.36% for broccoli, 98.09 ± 0.34% for leek, and 98.54 ± 1.21% for mushroom. The corresponding RMSE scores were 0.65 ± 0.13%, 0.32 ± 0.05%, and 3.08 ± 1.18% for broccoli, leek, and mushroom, respectively. The proposed pipeline outperforms the existing state-of-the-art (SOA) methods reported in the literature. The study highlighted the importance of wavelengths for MC estimation in agricultural samples, utilising the Explainable AI (XAI) Shapley Additive Explanations (SHAP) tool. The promising result of the study demonstrates that the involvement of XAI with the proposed pipeline facilitates transparent MC prediction in crops, ensuring supplies of higher-quality products to consumers.