Optimization of deep learning model for monitoring heterogeneous airborne PM concentrations using digital holographic microscopy

Airborne particulate matter (PM) is a global health concern and PMs with diameters less than 2.5 µm can permeate the lungs, causing illnesses. Current PM monitoring methods have limitations of being bulky, expensive, and require specialization. Hand-held devices are still costly while not measuring overly high/low concentrations accurately. Therefore, there is a need to make a compact and handy device that quickly and accurately measures PM concentrations. This study employs a novel digital in-line holographic microscopy setup with a smartphone to capture holographic images of PM signals. The speckle signals of heterogeneous PM_2.5 and PM10 recorded by a smartphone can be separated by 2D Gaussian filtering and used to optimize a deep learning network. The network consists of a deep autoencoder to regression layers which were trained on preprocessed speckle signals of homogeneous PM_2.5 captured by the smartphone. Using the high-pass filter on speckle signals of homogeneous PMs to extract 2.5 µm PM signals, the model was optimized and applied to predict heterogenous PM concentrations. Prediction results based on the optimized model can be utilized to both validate the method of separating PM_2.5 from heterogeneous PM signals and to measure heterogenous PM concentrations in real-time.