Investigating Data Mining Techniques to Detect Dust Storms in MODIS Imagery

摘要

NASA's Moderate Resolution Imaging Spectroradiometer (MODIS) provides global data with the correct spectral properties for dust storm detection. Unfortunately, there are no well accepted dust storm detection algorithms or methodologies for this data set. Different researchers have devised different dust indexes along the lines of Normalized Difference Vegetation Index (NDVI). For example, Normalized Difference Dust Index (NDDI) uses the difference in the reflectance between band 7 (2.13 urn) and band 3 (0.469 urn) divided by the sum of the two bands. One researcher uses NDDI value greater than 0.28 to classify the pixel as dust whereas another uses a threshold of 0. Other indexes such as Sand Dust Index (DSI) use different spectral channels - in this case Band 20 (3.7 urn) and 29 (8.5 urn). Which of these indexes along with the suggested threshold are really useful? This paper explores this question. It compares these different indexes not only against one another but also against machine learning techniques such as unsupervised classification. Unsupervised classification techniques such as clustering not only provide the ability to automatically group pixels in different classes but also determine the appropriate thresholds.