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

Abstract Rapid detection tasks in soil environment are generally implemented by various spectrometers and chemometric models. To reduce costs for model construction, calibration transfer from laboratory spectral instruments to portable devices has recently received extensive attention. In different application cases of model transference, most conventional methods require extra time to tune hyperparameters and to select calibration transfer techniques. Based on the near-infrared (NIR) analytical technique, this work aimed at exploring a transfer learning strategy to detect plastic pollution levels in the soil by transferring the model from a high-throughput hyperspectral image (HSI) system to an ultra-portable NIR sensor. Transfer learning was explored to diagnose the proper calibration transfer algorithm and construct the transferable model. For transferable model construction, conventional calibration transfer algorithms (Direct Standardization (DS) or Repeatability file (Repfile)) served as a pre-processing step, and non-parametric transfer learning algorithm (Easy Transfer Learning (EasyTL)) was explored in the modeling step. Supporting vector machine (SVM) was carried out as a typical modeling algorithm for comparison. For transformation algorithms selection, a distance metric algorithm, maximum mean discrepancy (MMD), was performed on spectral feature matrices before and after DS or Repfile transformation. On three transfer tasks, the results indicated that the Repfile-EasyTL model was a promising solution with higher accuracy, much lower time costs, less parameters, and dependency on the increase of standard samples than other models (SVM, DS-SVM, Repfile-SVM, EasyTL, DS-EasyTL). Moreover, MMD distance presented the great potential to serve as an indicator to vote the optimal calibration transfer algorithm before the modeling step.