Explainable artificial intelligence and deep reconstruction of hyperspectral images for advancing sweetpotato quality evaluation

Ahmed, Md Toukir

Permalink

https://hdl.handle.net/2142/129679 Copy

Description

Title

Explainable artificial intelligence and deep reconstruction of hyperspectral images for advancing sweetpotato quality evaluation

Author(s)

Ahmed, Md Toukir

Issue Date

2025-04-07

Director of Research (if dissertation) or Advisor (if thesis)

Kamruzzaman, Mohammed

Doctoral Committee Chair(s)

Kamruzzaman, Mohammed

Committee Member(s)

• Grift, Tony E.
• Rausch, Kent D.
• Wang, Yi-Cheng

Department of Study

Engineering Administration

Discipline

Agricultural & Biological Engr

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Explainable AI
• Hyperspectral imaging
• Image reconstruction
• Machine learning
• Sweetpotato quality

Abstract

Sweetpotato (Ipomoea batatas L.) is valued for its rich nutritional content, economic benefits, and versatility in both food and industrial applications, making its quality assessment essential. However, traditional methods for evaluating sweetpotato quality are often time-consuming and destructive. Hyperspectral imaging (HSI) has emerged as a powerful tool to assess internal composition and texture by capturing detailed spatial and spectral data. However, HSI generates high-dimensional data, which presents substantial computational challenges. To address these challenges, chemometrics is employed for effective data reduction and interpretation. In this context, explainable artificial intelligence (XAI) has become increasingly important in making predictive models more transparent and interpretable, thereby enhancing the reliability and adoption of artificial intelligence (AI)-driven quality assessments. Concurrently, advances in deep learning have opened new avenues for reconstructing hyperspectral data from standard RGB images, offering a more accessible and cost-effective alternative to traditional HSI systems. This approach, though underexplored in agricultural applications, holds significant potential for improving the practicality and efficiency of sweetpotato quality evaluation, making it a promising area for further research and development. This research aims to utilize the implementation of XAI and HSI reconstruction techniques to enhance the accuracy, transparency, and accessibility of sweetpotato quality assessment, ultimately contributing to the optimization of agricultural practices and industrial applications. In the first part of the study, XAI was integrated with hyperspectral imaging to enhance the assessment of three important quality attributes in sweetpotatoes, i.e., dry matter content (DMC), soluble solid content (SSC), and firmness. Sweetpotato samples of three different varieties, including “Bayou Belle”, “Murasaki”, and “Orleans”, were imaged using a portable visible near-infrared hyperspectral imaging (VNIR-HSI) camera, with a 400-1000 nm spectral range. The extracted spectral data were used to select key wavelengths, develop partial least squares regression (PLSR) models, and utilize shapley additive explanations (SHAP) values to ascertain model effectiveness and interpretability. The regression models (dry matter: R2p = 0.92, RMSEP = 1.50% and RPD = 5.58; soluble solid content: R2p = 0.66, RMSEP = 0.85obrix, and RPD =1.72; firmness: R2p = 0.85; RMSEP = 1.66N and RPD = 2.63) developed with key wavelengths were used to generate prediction maps to visualize the spatial distribution of response attributes, facilitating an improved evaluation of sweetpotato quality. Multivariate modelling techniques such as PLSR are commonly used for their simplicity, speed, and performance in industrial spectroscopic applications. While these models handle mild nonlinearities well, they often struggle with extrapolation. Therefore, the second part of the study aimed to utilize HSI and convolutional neural networks (CNN)-based regression to predict the firmness of various sweetpotato varieties by extracting spectral data from images captured with a VNIR-HSI system (400-1000 nm). The hyperparameters of CNN were fine-tuned using bayesian optimization (BO), which resulted in an 18.42% reduction in the prediction root mean squared error (RMSE) compared to the traditional PLSR model. Additionally, the SHAP method was applied to interpret the CNN model and assess the contribution of variable wavelengths. The CNN model based on important wavelengths was used to visualize spatial distribution of firmness in sweetpotato samples. Though HSI has emerged as a promising tool for many agricultural applications, the technology faces difficulty in being directly used in a real-time system due to the extensive time needed to process large volumes of data. Consequently, the development of a simple, compact, and cost-effective imaging system is not possible with the current HSI systems. Therefore, the overall goal of the third part of the study was to reconstruct hyperspectral images from RGB images through deep learning for agricultural applications. Specifically, this part of the study used hyperspectral convolutional neural network - dense (HSCNN-D) to reconstruct hyperspectral images from RGB images for predicting SSC in sweetpotatoes. The algorithm reconstructed the hyperspectral images from RGB images, with the resulting spectra closely matching the ground-truth. The PLSR model based on reconstructed spectra outperformed the model using the full spectral range, demonstrating its potential for SSC prediction in sweetpotatoes. In the final part of the study, three different hyperspectral reconstruction algorithms, such as HSCNN-D, hierarchical regression network (HRNET), and multi-Scale transformer plus plus (MST++), were compared to assess the DMC of sweetpotatoes. Among the tested reconstruction methods, HRNET demonstrated superior performance, achieving the lowest mean relative absolute error (MRAE) of 0.07, RMSE of 0.03, and the highest peak signal-to-noise ratio (PSNR) of 32.28 decibels (dB). Some key features were selected using the genetic algorithm (GA), and their importance was interpreted using XAI. PLSR models were developed using the RGB, reconstructed, and ground truth (GT) data. The visual and spectra quality of these reconstructed methods was compared with GT data, and prediction maps were generated.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129679

Copyright and License Information

Copyright 2025 Md Toukir Ahmed

Owning Collections