Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Assessing the Effect of PCA-Based Dimensionality Reduction on Machine Learning Performance in Hyperspectral Optical Imaging

Parisa Parand, Mahmoud Samadpour

TL;DR

Hyperspectral imaging yields many correlated bands, creating a dimensionality challenge for regression tasks like soil moisture estimation. The study applies PCA to reduce a 125-band dataset to two principal components that retain >$99\%$ of the variance, followed by training a Random Forest regressor, achieving $R^2=0.947$ on a held-out test set. The two-component PCA representation yields clearer structure and maintains predictive performance while reducing computational load, as visualized by PCA projections. This work provides a practical, scalable template for integrating dimensionality reduction into hyperspectral ML workflows and motivates exploring alternative feature extraction methods across diverse sensing scenarios.

Abstract

Hyperspectral optical imaging provides rich spectral information for estimating continuous environmental and material parameters; however, its high dimensionality and strong feature correlation pose significant challenges for machine learning models, especially when ground-truth datasets are limited. In this study, we investigate a hyperspectral dataset composed of 150 spectral bands with soil moisture as the target variable. To address the curse of dimensionality, Principal Component Analysis (PCA) was employed as a baseline dimensionality reduction technique. The optimal number of principal components was determined to be two, retaining more than 99% of the total variance. This selection was supported by the analysis of the covariance matrix, eigenvalue distribution, and the scree plot. Projecting the data onto the first two principal components enabled improved visualization and interpretability compared to the original high-dimensional feature space. The reduced representation also revealed a clearer separation of target values, effectively decreasing data complexity. To evaluate the impact of dimensionality reduction on predictive performance, a Random Forest regression model was trained to estimate soil moisture from the PCA-transformed data. The model achieved a coefficient of determination (R2) of 94.7 %, demonstrating that PCA-based feature reduction can enhance computational efficiency while preserving strong predictive capability in hyperspectral machine learning workflows.

Assessing the Effect of PCA-Based Dimensionality Reduction on Machine Learning Performance in Hyperspectral Optical Imaging

TL;DR

Hyperspectral imaging yields many correlated bands, creating a dimensionality challenge for regression tasks like soil moisture estimation. The study applies PCA to reduce a 125-band dataset to two principal components that retain > of the variance, followed by training a Random Forest regressor, achieving on a held-out test set. The two-component PCA representation yields clearer structure and maintains predictive performance while reducing computational load, as visualized by PCA projections. This work provides a practical, scalable template for integrating dimensionality reduction into hyperspectral ML workflows and motivates exploring alternative feature extraction methods across diverse sensing scenarios.

Abstract

Hyperspectral optical imaging provides rich spectral information for estimating continuous environmental and material parameters; however, its high dimensionality and strong feature correlation pose significant challenges for machine learning models, especially when ground-truth datasets are limited. In this study, we investigate a hyperspectral dataset composed of 150 spectral bands with soil moisture as the target variable. To address the curse of dimensionality, Principal Component Analysis (PCA) was employed as a baseline dimensionality reduction technique. The optimal number of principal components was determined to be two, retaining more than 99% of the total variance. This selection was supported by the analysis of the covariance matrix, eigenvalue distribution, and the scree plot. Projecting the data onto the first two principal components enabled improved visualization and interpretability compared to the original high-dimensional feature space. The reduced representation also revealed a clearer separation of target values, effectively decreasing data complexity. To evaluate the impact of dimensionality reduction on predictive performance, a Random Forest regression model was trained to estimate soil moisture from the PCA-transformed data. The model achieved a coefficient of determination (R2) of 94.7 %, demonstrating that PCA-based feature reduction can enhance computational efficiency while preserving strong predictive capability in hyperspectral machine learning workflows.

Paper Structure

This paper contains 9 sections, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Research Methodology
  3. Dataset Description
  4. Data Preparation
  5. Dimensionality Reduction and Machine Learning Model
  6. Results and Discussion
  7. Dimensionality Reduction using PCA
  8. Visualization of Principal Components
  9. Conclusion

Figures (5)

  • Figure 1: Distribution of soil moisture in the reference dataset, showing most measurements concentrated near 32%.
  • Figure 2: Correlation heatmap of sample hyperspectral bands and soil moisture (correlation scale: $-1$ to $+1$).
  • Figure 3: Distribution of soil moisture in train and test samples.
  • Figure 4: Scree plot showing dominant contribution of PC1 and PC2.
  • Figure 5: (a) PC1 vs. PC2 and (b) PC2 vs. PC3 projections with soil moisture color mapping.