Identification of Smoke Plumes from Satellite Images
This work was completed in Vancouver while Victoria Wan, a UWO student supervised by John Braun, was working under the supervision of Charmaine Dean at SFU. It is a joint effort by the student and supervisors across these two institutions, as well as collaborators at the British Columbia Ministry of Healthy Living and Sport and the UBC School of Environmental Health.
The health effects of forest fire smoke is challenging to determine because exposure is episodic, and because of the sparseness of monitoring networks found in semi-rural areas where risk of smoke exposure is considered high.
In 2003, the location of British Columbia forest fires concentrated in the southern interior region (see map below). The satellite image below shows the location of air quality monitors (represented by the orange dots) in the region around Kelowna; the districts of highest population density in this region are identified by the two circled areas, with the lower area being the city of Kelowna itself.
The sparseness of the air quality monitors, especially in semi-rural areas, cannot provide spatially representative data.
Satellite imagery is one major untapped resource for estimating smoke exposure as it offers a novel and data-rich tool to provide visual information on smoke plumes. In the image below, fires are identified in red.
Consider the red, green blue colour-space data values at pixel-level from satellite images; the figure below displays transformed values from 50 images that capture the Kelowna region in British Columbia during the fire season in 2007.
The colour signature at the lower transformed colour scale represents the land class and that at higher end of the colour scale represents smoke or cloud class.
Assuming that the data at pixel level arise from normal mixture distributions, it is possible to create a classification scheme which distinguishes land from smoke or cloud based on a such a parametric assumption. In comparison to the parametric mixture model approach, the use of the so-called non-parametric data sharpening technique 1 may also be applied to obtain estimates of forest fire smoke plumes.
Data sharpening was developed to provide bias-reduction for density estimation. It moves data points near the peaks and troughs to become more tightly aligned, which reduces the bias from underestimating the peaks and overestimating the troughs. This non-parametric approach can used be to convergence to obtain local modes in the data.
In the images above, the areas coloured in yellow are the predicted ânon-landâ areas using data sharpening, while pink areas are the predicted ânon-landâ areas using the mixture model method. From our investigations, it seems that data sharpening in general does not pick up the wispy part of the smoke areas while the mixture model over-predicts the area of smoke.
These smoke plume identification techniques are being improved by incorporating spatial correlation in the data and knowledge of the locations of the fire centres.