GF-4 can provide stable data for disaster prevention and mitigation, and its mid-infrared sensor can be well applied in rapid-fire monitoring. Because of lacking far-infrared, the spectral information that GF-4 provided is supplementary data as usual. Affected by a single band, the commission error and omission error of the adaptive threshold method is high. Therefore, to probe the potential of GF-4 data and improve the accuracy of fire point recognition, this study analyzed the characteristics of GF-4 data and proposed a fire point detection method with brightness temperature difference correction based on dual temporal image. The method mainly includes three parts: brightness temperature compensation acquisition based on Kriging interpolation on temporal scale, adaptive threshold segmentation on a spatial scale based on contextual information, and fire point detection, with two images-before and during the fire event. Firstly, the difference between the two images is processed. Moreover, we use this difference of non-polluted pixels in the dynamic neighborhood around the potential fire point as the sampling data for spatial interpolation and then substitute the result of the previous step into the first image. Finally, using discrimination conditions for fire point discrimination and false alarm elimination get the final results. The study also compares three spatial interpolation acquisitions: Inverse Distance Weigh, Simple Kriging and Ordinary Kriging. From the fitting results, the Ordinary Kriging can reflect the volatility of the pixel area and has a certain smoothing effect to avoid peaks of background brightness temperature, which is the better method. The study area contains two fires in Qinyuan, Shanxi Province and New Barhu Right Bannerin, Inner Mongolia. Results show that compared with the traditional single time phase algorithm, introducing brightness temperature difference correction data can better fit the background brightness temperature, reducing the commission error to 3% and obtaining comprehensive evaluation index F_β above 0.9. This developed method could be used to support automatic fire point detection and extraction in future studies.