i Detection of Smoke and Dust Aerosols Using Multi-sensor Satellite ...

values of dark surface is smallest, generally less than 20 K. In Figs. 4.4d and 4.4f, the
logarithm of single reflectance of red band is introduced to separate dust from surface
scenes. Using logarithm can enhances the resolving power of this index, especially in
small reflectance range. Obviously, the Ln (R1) of most dust pixels is larger than -1.2
over bright surface, and -1.6 over dark surface.
The detail thresholds of each index are summarized in the Table 4.4. The thresholds
are calculated statistically from all training data, namely selected dust events in China
during 2001-2007. The errors caused by using the defined thresholds of each test are also
listed in Table 4.4. The largest error is appeared at the separation of dust from cloud over
dark surface in reflective solar spectrum, which is up to 12.484%. This big error caused
possibly by light dust pixels suspended over water. Considering that some pixels can be
counted repeatedly, the total errors of both over bright and dark surfaces are not equal to
the summation of errors from each test.
4.2.2 Flowchart
The flowchart of dust detection algorithm is shown in Fig. 4.5. The L1B
measurements are input into cloud module to filter out cloud first by employing BTD (12,
11) and NDDI. The remaining pixels are then divided into two branches based on
pre-stored surface type information: dust over bright surface and over dark surface. The
BTD (3.7, 11) and the logarithm of reflectance at band 1 are applied in both branches
with respective thresholds. At the end, a same “noise-elimination” process is executed to
filter out discrete pixels.
54