Application of hyperspectral imaging in environmental monitoring: air pollution classification and detection

Over the past few years, air pollution has become a worldwide concern. Particulate matter (PM2.5) with a diameter smaller than 2.5 μm possesses the capability to travel through the atmosphere and deliver perilous substances to the lungs of humans through inhalation, thereby causing substantial health complications. Hence, this research aims to conduct a spectral analysis of PM2.5 images derived from four distinct regions, including trees, roofs, and roads, in order to classify the images as "Good," "Normal," or "Severe" in accordance with the pollution's severity. In pursuit of this objective, what we believe to be a novel algorithm for converting RGB images to hyperspectral images (cHSI) in order to extract spectral information was devised. In the absence of a designated dataset for evaluating the snapshot cHSI algorithm, a dataset comprising 15,137 images was compiled specifically for this study. Following the division of the dataset into training and testing sets, two distinct three-dimensional convolutional neural network (3DCNN) models were trained using the traditional RGB and snapshot HSI. Subsequently, the predictive accuracy of the models was evaluated. RGB-3DCNN and HSI-3DCNN were developed in order to compare the performance of the cHSI model to that of the traditional RGB method using snapshot cHSI images and traditional RGB images, respectively, as inputs for the models. The replacement of the RGB-3DCNN model with the cHSI-3DCNN model resulted in improved accuracy in all four regions of air pollution. That may enhance precision by as much as 9% across a range.