Deep or Shallow? A Comparative Analysis on the Oil Species Identification Based on Excitation-Emission Matrix and...

Area of Science:

• Environmental Science
• Analytical Chemistry
• Marine Biology

Background:

• Increased petroleum activities heighten the risk of marine oil spills, threatening ocean ecosystems.
• Excitation-emission matrices (EEMs) offer a viable method for identifying oil types in spills.
• The comparative performance of deep learning versus shallow learning models for EEM analysis remains underexplored.

Purpose of the Study:

• To compare the effectiveness of four machine learning models in identifying oil species from EEMs.
• To evaluate the performance of deep convolutional neural networks (DCNNs) against traditional models like random forest (RF), support vector machine (SVM), and back propagation neural network (BPNN).

Main Methods:

• Collected fluorescence EEMs of common oils using a tuneable xenon lamp and spectrometer.
• Trained and tested RF, SVM, BPNN, and DCNN models using the generated EEM datasets.
• Evaluated models based on identification accuracy, computational requirements, and processing time.

Main Results:

• SVM, BPNN, and DCNN models achieved over 93% accuracy in identifying tested oil types.
• Deep learning models (BPNN, DCNN) showed no significant accuracy improvement over the SVM model.
• SVM demonstrated a superior balance between high accuracy and computational efficiency.

Conclusions:

• Support Vector Machine (SVM) is a highly suitable model for oil species identification in marine spills, balancing accuracy and efficiency.
• While deep learning models perform well, their computational demands may not justify the marginal accuracy gains over SVM.
• This comparative analysis provides practical guidance for selecting appropriate machine learning models for oil spill identification and response.