Integrating physical index and self-organizing mapping for aerosol dust detection (PISOM) over Himawari-8 AHI satellite images

Abstract

Dust detection from satellite images has been explored with both physical dust index method and machine learning method. However, both methods have their own limitations. The dust index method is threshold dependent and uses limited satellite observations, whilst the machine learning method requires a large quantity of training data and is generally physically uninterpretable. Besides, previous studies give definite detection results, i.e., 1 and 0 representing dust and non-dust. In actual situations, dusts could be mixed with clouds, and thin dust plumes can be regarded as a mixture of dusts and land surfaces. To tackle these problems, this study proposes a new method that integrates physical, machine learning, and analytical methods, namely Physical Index and Self-Organizing Mapping (SOM) integrated detection method (PISOM). It first uses physical indices to extract all dust-like pixels. Then the preliminary detection results were refined with a well-trained SOM model. Finally, it calculated the particle-type ratios for each dust-like pixel with an analytical method. The PISOM has been testified for dust aerosol detection over Northern China using Himawari-8 AHI satellite images. It was cross-compared with the two classical physical dust indices of brightness temperature difference (BTD) and normalized difference dust index (NDDI) on 5732 samples of heavy dust (HD), thin dust (TD), desert surface (DS), and thin cloud (TIC), which were manually extracted from historical dust storm cases. The results show that PISOM has the highest accuracy score (0.930) than the BTD (0.890) and NDDI (0.700). The confusion matrix of PISOM for the four types reveals that the PISOM misclassifies 29 and 325 samples of HD and TD as TIC. An in-depth examination reveals that these misclassified samples are mixtures of HD and TD with TIC, which indicates the effectiveness of PISOM in interpreting the mixture cases. Moreover, application on typical dust storm cases demonstrates that the PISOM performs well over different regions and under various illumination conditions. Importantly, the quantified dust detection results make the PISOM application-oriented, i.e., the results can be used to estimate dust-affected areas or quantify the probability of dust ratio over particular pixel(s).