Classifying Reflectance Targets under Ambient Light Conditions using Passive Spectral Measurements

Area of Science:

• Remote Sensing
• Spectroscopy
• Machine Learning

Background:

• Collecting spectral data under variable ambient light poses challenges for accurate target classification.
• Portable spectrometers require robust methods to compensate for changing light conditions.
• Accurate spectral measurements are crucial for applications in environmental monitoring and material identification.

Purpose of the Study:

• To evaluate ambient light compensation methods for classifying grayscale targets using portable spectrometers.
• To compare the effectiveness of different spectral data processing techniques under varying light.
• To determine the optimal approach for maximizing classification accuracy in challenging lighting environments.

Main Methods:

• Utilized two sets of UV-VIS-NIR spectrometers, one uncalibrated for raw intensity and one calibrated for irradiance.
• Investigated three ambient light compensation methods: integration time maximization (M1), normalization by integration time (M2), and irradiance division (M3).
• Employed partial least squares (PLS) regression for spectral data preprocessing followed by an ensemble of machine learning algorithms for classification.

Main Results:

• The method maximizing spectral sensitivity (M1) demonstrated the highest prediction accuracy for classifying known targets.
• Average prediction accuracy across all spectrometers and compensation methods surpassed 93%.
• Partial least squares regression effectively compressed spectral data and prevented overfitting.

Conclusions:

• Ambient light compensation is critical for reliable spectral target classification with portable spectrometers.
• Maximizing spectral sensitivity offers a robust strategy for improving classification accuracy under variable lighting.
• The developed methods and machine learning approach provide a high-accuracy solution for remote spectral analysis.