Earthquake prediction is one of the key areas of earthquake research. Thermal infrared abnormity, which is the abnormally increased land surface temperature, is universal before earthquake and has complex nonlinear relation with the three elements of earthquake. Combining the advantage of neural network, this paper provides a method for earthquake prediction by taking thermal anomaly as information source and constructing a neural network to carry out the test. Based on the MODIS data which has synthesis of eight days with 1km resolution, taking a 10°×10° rectangle, whose center is the epicenter, as research area, and a two-month time before earthquake as the time range, we used RST algorithm to extract thermal anomaly information before earthquake. Considering the time-space relationship between thermal anomaly information and the fault zone, thinking carefully about the information of the neural network input neurons, we constructed BP neural network and used 100 earthquake cases with magnitudes larger than 5, and 70 random samples without earthquake in the research region for training and simulating. According to the statistical analysis, the prediction accuracy is 80%, missing prediction rate is 20%, and false prediction rate is 13.3%. Prediction accuracy of magnitude with error within magnitudes of 2 is 69%, prediction accuracy of earthquake origin time with error within 30 days is 87.5%, and prediction accuracy of epicenter location with error within 3°is 81.2%. The result shows that BP neural network-based earthquake prediction is feasible by using thermal infrared abnormal precursor extracted by RST method. However, in this experiment, the determination of the start time of thermal abnormity, the origin point and range of research area are based on the known epicenter location and time. In fact, the result depicts a non-linear relationship between earthquake and thermal abnormity, and the accuracy of prediction reflects the correlation degree. Therefore, the prediction accuracy of future earthquake may be not as large as our result. For future earthquake prediction, accurate selection of research area and neuron number of hidden layer in neural network has great influence on prediction result.