Machine learning-supported cross-excitation laser-induced fluorescence lidar for the classification of oil spill.

Accurate identification of the spilled oil type was crucial for implementing suitable emergency response measures and guiding further marine oil pollution remediation. A cross-excitation laser-induced fluorescence (LIF) lidar was developed to acquire fluorescence spectra from 152 distinct oil samples (1368 spectra, 0.15 nm spectral resolution, and 1 min integration) under different wavelength excitations (365, 385, and 405 nm). Based on LIF data the random forest (RF) and extreme gradient boosting (XGBoost) algorithms were used along with the shapley additive explanations (SHAP) for oil classification. The detection precision of XGBoost (0.922) was significantly higher than RF (0.856) across the hold-out test set, as well as exhibited a considerable discrimination in the receiver operating characteristic (ROC) analysis. SHAP analysis revealed that oil classification is governed by two key spectral regions: the near-UV/blue (370-470 nm) and the green-yellow (500-570 nm) bands. Specifically, wavelengths such as 479, 504, 485, 454, and 472 nm acted as positive markers for crude oil, while green-yellow features (e.g., 533, 577, 600, 510 nm) were decisive for light fuel oil. This discriminative capability is due to the fundamental compositional differences in aromatic chemicals between oil types, which result in unique fluorescence responses under laser excitation. Simultaneously, the stage-wise additive structure of XGBoost proved capable of capturing subtle, narrow-band "wavelength-intensity-interaction" cues. This study demonstrated that the proposed use of the LIF-XGBoost framework could be an effective approach for accurate and interpretable oil classification, which offered a practical and robust pathway for emergency management of marine oil spill.