Detecting smoke over land using imaging satellite data

Detecting smoke over land using imaging satellite data

SHARE

US: A group of scientists, Yong Xie, John J. Qu, Xiaoxiong Xiong and Yan Zhang have developed a multispectral approach that relies on remote-sensing measurements by NASA’s Moderate Resolution Imaging Spectroradiometer, making it possible to distinguish polluting smoke from clouds. The scientists have identified smoke by filtering out other types of scenes based on differences in their spectral features. A scaled index was applied to define the confidence level of smoke pixels and to characterise the smoke intensity. Their algorithm provides a quick means of automatically localizing the smoke in an entire MODIS swath.

According to the scientists, satellite remote sensing is already in use to track and monitor smoke in a timely manner from space, but most of the ongoing implementations of the technique are unsuitable for bright surface areas or in regions where smoke is mixed with clouds. Hence, they developed a new technique.

Yong Xie and John J. Qu are associated with George Mason University. Xiaoxiong Xiong and Yan Zhang are working at NASA Goddard Space Flight Center.

Table 1 lists the scene types, threshold tests, and thresholds used for smoke detection over land. In the table, B denotes the MODIS band, BT indicates brightness temperature and BTD is the brightness temperature difference.

Table 1. The threshold tests and thresholds used for detecting smoke over land.

Scene type Threshold test Threshold
  B26 and 0.035
Cloud BTD (31, 22) and −10 K
  BT31 293 K
Soil (B3−B7)/(B3+B7) 0.10
  (B3−B8)/(B3+B8) −0.15
Vegetation and water and B8 0.17
  and B1 0.14

The most important step in detecting smoke is to accurately separate it from clouds. In our algorithm, three simultaneous tests examine the BT of B31 (11μm), the BTD between B31 and B22 (3.9μm), and the reflectance of B26 (1.38μm). These bands are insensitive to smoke but sensitive to the presence of clouds. The normalised ratio of B3 (470nm) and B7 (2.13μm) (see Table 1), whose value is scaled to [0, 1], is applied to filter out soil and bright surface pixels. The reflectance of B8 (412nm) is designed to filter out water and vegetation pixels.

The average geometrical radii of smoke particles are small, within the range of 0.01–0.05μm. According to the Rayleigh scattering theory, the two blue bands (B3 and B8) are very sensitive to the existence of smoke. Thus, they propose that the normalised ratio of these bands—ratio (B3, B8)—be considered a key parameter in the algorithm. The ratio defines the confidence level of smoke or non-smoke pixels and serves as an indicator of smoke intensity. A close linear relationship between ratio (B3, B8) and the reflectance of B3 indicates that the stronger the smoke, the larger the ratio value (it is generally larger than −0.15). Normalisation expands the ratio from [−0.15, 0.05] to [0, 1]. The pixels with ratios larger than 0.05 are treated as heavy smoke pixels and given the highest possible confidence value of 1. A 5×5 pixel box is used to single out pixel outliers based on the continuity of the smoke.

The algorithm described here provides a rapid means of automatically detecting smoke over land from space using MODIS satellite remote-sensing measurements. The scientists propose a normalised ratio of two blue bands to characterise the smoke intensity and to define the confidence level of the results. Smoke was detected successfully in many locations using this approach. Nonetheless, as a next step, they plan to improve the algorithm by integrating multisensor measurements to detect biomass smoke more effectively and accurately.

Source: SPIE