The observation characteristics of MODIS data with high temporal resolution and medium spatial resolution can play an important role in fire detection. However， MODIS fire detection is currently in areas with high heterogeneity， and there are many false detections of fire， and cold fire are easily missed. To solve this problem， in order to fully mine the relevant information in MODIS data， realize the high-precision identification of fire points. A MODIS fire detection algorithm using deep learning technology is proposed. Acquisition of a large number of samples with high quality and broad representation is the prerequisite for deep learning to achieve accurate detection of fire. In order to increase the number of fire samples and ensure the quality of wildfire samples， use the American ground wildfire data set as real fire samples to accurately match them with MODIS data in time and space， and build a fire detection sample library based on deep learning methods . According to the analysis of the radiation transfer process， the wave band and band combination with good identification for fire detection are determined as the input source. Based on the constructed sample data set and information source， build a DNN （Deep Neural Network） fire detection model. Application experiments were carried out in three typical scenarios and compared with MODIS fire products. The results show that the improved method reduces the average brightness temperature of 4um by 2 K in the extraction of cold fire in agricultural areas， and the ratio of correct to wrong changes is positive. Compared with MODIS products， the false fire points extracted in the suburban area are significantly reduced， and the false fire points near the 4um brightness temperature of 325 K are excluded， and the false detection rate is reduced by 19.89%.