Transfer learning strategy for plastic pollution detection in soil: Calibration transfer from high-throughput HSI...

Area of Science:

• Environmental Science
• Analytical Chemistry
• Spectroscopy

Background:

• Rapid soil analysis often relies on spectrometers and chemometric models.
• Calibration transfer from lab to portable devices is crucial for cost reduction.
• Conventional methods demand significant time for hyperparameter tuning and technique selection.

Purpose of the Study:

• To explore a transfer learning strategy for detecting soil plastic pollution.
• To transfer models from high-throughput hyperspectral imaging (HSI) to ultra-portable near-infrared (NIR) sensors.
• To identify an optimal calibration transfer algorithm and construct a transferable model.

Main Methods:

• Utilized near-infrared (NIR) analytical technique.
• Employed transfer learning, including Direct Standardization (DS) and Repeatability file (Repfile) for pre-processing.
• Applied Easy Transfer Learning (EasyTL) for modeling, compared with Support Vector Machine (SVM).
• Used Maximum Mean Discrepancy (MMD) to select optimal transfer algorithms.

Main Results:

• The Repfile-EasyTL model demonstrated superior accuracy and reduced time costs compared to other models.
• This approach requires fewer parameters and less dependency on standard samples.
• MMD distance effectively indicated the optimal calibration transfer algorithm prior to modeling.

Conclusions:

• The Repfile-EasyTL transfer learning strategy is a promising solution for accurate and efficient soil plastic pollution detection.
• This method significantly lowers costs and time associated with model calibration transfer.
• MMD serves as a valuable indicator for selecting the best transfer algorithm, streamlining the process.