Abstract: High-precision numerical simulation of typhoon field relies on accurate initial field and constrained field as input and real-time assimilation data. Compared with the inshore and land areas where typhoons are mature and observation methods are relatively abundant, the deep sea areas with rapidly intensification often lack effective conventional observation data, and cannot provide sufficient assimilation data for typhoon numerical simulation. Infrared remote sensing of satellite is an important technical means for deep sea meteorological observation, but its application in typhoon wind field is still limited to the method of artificial identification of cloud system characteristics for positioning and intensity determination, and no spatial mapping relationship between satellite remote sensing data and typhoon wind field has been established. A novel adaptive loss function for typhoon wind speed is proposed to solve the problem that conventional convolutional neural networks cannot consider the unbalanced distribution of high and low wind speed data in the typhoon horizontal wind field, and a deep sea typhoon near-ground horizontal wind field inversion model based on satellite nine-channel infrared remote sensing data is established. The research shows that the adaptive loss function accurately captures the distribution characteristics of high wind speed range in the near-ground horizontal wind field of the deep sea typhoon. The average wind speed error and wind direction error of the model proposed in this paper are 2.24 m/s and 11.5°. The inversion error of strong wind speed interval in typhoon field is reduced by 24.16% compared with the traditional convolutional neural network. At the same time, this method has good generalization and extrapolation ability, and improves the time resolution of reanalysis of wind field generated to 10 minutes, which is consistent with the remote sensing data, and provides a new idea for the wind field inversion of the reanalysis of the deep sea typhoon.