Coal fires are widely prevalent in the north of China. They have already caused huge losses in
resources and pose a serious environmental problem. To monitor and extinguish coal fires, the first step is
to detect their location and scale. Because of the huge amounts of heat energy released by coal fires, the
resulting thermal anomalies can be detected by using thermal infrared remote sensing technology. On
nocturnal aerial images it is relatively easy to discern coal fires, because the effect of solar radiation is
insignificant. However, nocturnal aerial images are not available as often as Landsat TM daytime images
for such a large area as the north of China. In this paper, we first give a briefing of the basic principle in
reducing solar radiation on TM thermal IR image. Then, neural network is used to set up a mathematical
model of ground temperature. In view of the special character of artificial neural network used in this
application, we offer the batch learning approach and adjust active momental factor. The result achieved
by reducing solar radiation on TM thermal IR image is as good as airborne nighttime thermal infrared
image for detecting coal fires. So this method is very practical and greatly economizes the cost of aerial
remote sensing image.