Collecting remotely sensed spectral data under varying ambient light conditions is challenging. The objective of this study was to test the ability to classify grayscale targets observed by portable spectrometers under varying ambient light conditions. Two sets of spectrometers covering ultraviolet (UV), visible (VIS), and near-infrared (NIR) wavelengths were instrumented using an embedded computer. One set was uncalibrated and used to measure the raw intensity of light reflected from a target. The other set was calibrated and used to measure downwelling irradiance. Three ambient-light compensation methods that successively built upon each other were investigated. The default method used a variable integration time that was determined based on a previous measurement to maximize intensity of the spectral signature (M1). The next method divided the spectral signature by the integration time to normalize the spectrum and reveal relative differences in ambient light intensity (M2). The third method divided the normalized spectrum by the ambient light spectrum on a wavelength basis (M3). Spectral data were classified using a two-step process. First, raw spectral data were preprocessed using a partial least squares (PLS) regression method to compress highly correlated wavelengths and to avoid overfitting. Next, an ensemble of machine learning algorithms was trained, validated, and tested to determine the overall classification accuracy of each algorithm. Results showed that simply maximizing sensitivity led to the best prediction accuracy when classifying known targets. Average prediction accuracy across all spectrometers and compensation methods exceeded 93%.
Keywords: ambient light compensation; machine learning; reflectance target classification; remote sensing; spectroscopy.