Vis-NIR spectra-image transformation based on circular spectral mapping for measurement of particulate matter...

The fine particulate matter (PM_2.5), one of the most concerned airborne pollutants, significantly impacts air quality and human health. The potential hazard of PM_2.5 related to its concentration, while the traditional methods for PM concentration measuring were expensive, time-consuming, while low-cost sensors often suffer from poor accuracy and stability. Therefore, there is a great need for a rapid, precise and stable measurement method for filter-based PM_2.5.

We propose a novel spectra-image transformation and fusion method for filter-based PM_2.5 measurement using a portable visible near infrared (Vis-NIR) spectrometer. Traditional machine learning models based on spectra alone achieved low accuracy (R²_p < 0.8). To improve performance, we introduced the circular spectral mapping (CSM) method to transform PM_2.5 spectra into CSM images, which were processed using ResNet-18, ShuffleNet V2, and MobileNet V2 networks with an attention mechanism module. The optimal model, ShuffleNetV2_Attn, improved R²_p to 0.9935. To furtherly improve the model stability, the numerical and graphical feature fusions were conducted, and the ShuffleNetV2_Attn was selected as optimal feature extractor of CSM images. The machine learning models were built based on fusion features, and the optimal model was the partial least squares (PLS) model based on fusion features extracted by successive projections algorithm (SPA), of which the R²_p, RMSEP and mean absolute percentage error (MAPE_p) were 0.9947, 6.0213 μg/m³ and 4.17 %, demonstrating high accuracy and stability overall concentration range.

The proposed spectra-image transformation and fusion method greatly improved the accuracy and efficiency of filter-based PM_2.5 measurement. It overcome the limitations of spectral-based machine learning methods, which often fail to capture full-band characteristics, and provides a new approach for integrating numerical and graphical spectral information.