Early warning of leakage, especially small leakage, is significant for safety maintenance of thermal pipeline. Due to spatial resolution, the measuring accuracy of distributed fiber optic sensor for local temperature variation caused by small leakage is low and the measurements are quite different from the actual temperature field. Based on Brillouin optical time domain reflectometer(BOTDR), a new method to establish a mapping relationship between the BOTDR measurements and the actual temperatures is proposed. Laboratory experiments were carried out to simulate small leakage and achieve the measurements of gradient temperature fields. Feature extraction of the measured data is then conducted through Gaussian fitting. With artificial neural network(ANN), a mapping model of the actual and measured temperature features is established. The results demonstrate that: the designed experiment can accumulate enough prior data to derive an ANN model, based on which a mapping relation of the actual temperature field and the BOTDR measurements can be achieved to improve the measuring accuracy of BOTDR and provide a reference to propose warning strategy.