Deep learning applications in hyperspectral imaging for agriculture: image reconstruction and model design for quality prediction

Monjur, Ocean

Permalink

https://hdl.handle.net/2142/132785 Copy

Description

Title

Deep learning applications in hyperspectral imaging for agriculture: image reconstruction and model design for quality prediction

Author(s)

Monjur, Ocean

Issue Date

2025-12-03

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

Kamruzzaman, Mohammed

Committee Member(s)

• Rausch, Kent D.
• Malvandi, Amir

Department of Study

Engineering Administration

Discipline

Agricultural & Biological Engr

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Hyperspectral imaging reconstruction
• Agricultural applications
• Computer vision
• Hyperspectral dataset

Abstract

Non-invasive techniques, such as hyperspectral imaging (HSI), are crucial for analyzing the detailed chemical and structural composition of agricultural products. By capturing both spectral and spatial information simultaneously, HSI enables advanced analysis of key quality attributes in agricultural products. Despite the benefits, the adoption of HSI remains limited due to the high cost and complexity associated with collecting and processing hyperspectral data. To address these challenges, this thesis presents two complementary contributions. The first is Agro-HSR, a large-scale RGB-to-Hyperspectral image reconstruction dataset of sweet potatoes, comprising 1322 RGB-HSI pairs. For a subset of 141 samples, agro-product quality attributes, including Brix, dry matter content, and firmness, are also provided. The goal of this dataset is to promote the use of deep learning models in converting standard RGB images into HSI, thereby reducing the overall cost of data acquisition for HSI. The reconstructed spectra from the best performing reconstruction model had R^2 scores of 0.52, 0.88, and 0.85 in Brix, dry matter content, and firmness, respectively, all of which closely follow the original spectra results. The second contribution is Agro-Net, a Convolution-Attention Fusion model designed to extract complementary features from hyperspectral images for accurate agro-attribute prediction. Agro-Net outperforms traditional approaches in predicting the firmness of potatoes and the fertility of eggs, highlighting the benefits of leveraging both spectral and special features from HSI to improve agro-attribute prediction. Combined, these contributions advance the application of deep learning in agricultural hyperspectral imaging, enabling both practical data accessibility and effective predictive modeling.

Graduation Semester

2025-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/132785

Copyright and License Information

Copyright 2025 Ocean Monjur

Owning Collections