|
The application of neural network algorithms for the quantification of Volatile Organic Compounds (VOCs) concentrations, derived from infrared absorption spectral data, has been shown to achieve superior accuracy compared to conventional least-squares regression techniques. The current neural network models in use generally have issues with low precision and stability, which affect the consistency and credibility of the inversion results. In this study, in order to resolve the aforementioned challenges, we present the pioneering application of the MiniRocket (minimally random convolutional kernel transform) model for the quantitative analysis of VOCs concentrations utilizing hyperspectral data derived from satellite platforms. This model surpasses traditional neural network methodologies by virtue of its automated feature extraction, enhanced computational efficiency, near-deterministic processing, and superior predictive accuracy. The near-deterministic nature of MiniRocket's transformation process ensures reproducibility, as it guarantees identical outcomes given the same input data across diverse computational settings. We employed a training dataset consisting of 120 infrared hyperspectral data with a spectral resolution of 1 cm-1 and a spectral range of 2.8 to 14.3 μm. Additionally, we utilized a validation dataset comprising 80 sets of test data with randomly assigned concentrations. Experimental results indicate that the MiniRocket model achieves a mean error of prediction (MES) of 6.2×10-3 parts per million (ppm) for the estimation of pollutant gas concentrations, with a processing time reduced to 0.02 seconds. These outcomes not only underscore the model's superior predictive accuracy but also highlight its unparalleled computational efficiency when compared to other existing models in the field.
|
