Abstract: ‍ ‍The detection of low slow small targets in complex radar environments is a challenging task， and the use of deep learning and data feature fusion methods is an effective means to solve this problem. This paper optimizes the Radar Map fusion Detection Network （RMDN） based on the original algorithm， mainly focusing on distinguishing the importance of radar and map information in the detection process. The specific optimization content is to reduce the network depth of the map feature extraction module， add a channel attention mechanism， and allow the neural network to autonomously learn the weights of radar and map information features， so that the neural network can better utilize map information to assist in the detection of radar targets. Based on this optimization， this paper redesigns the Radar Map Fusion Detection Network RMDN-V2. The main idea of the algorithm is to use satellite remote sensing maps to provide background environmental information as an aid to radar signal detection. By incorporating feature information from the target background into the detection decision， it improves the accuracy and robustness of target detection， reduces the sensitivity to strong clutter and moving objects， and improves the performance of target detection algorithms in complex environments. In the original RMDN algorithm， there was no distinction between the importance of radar information and map information. The two occupy the same position in the algorithm， which is inconsistent with the role of satellite remote sensing maps only providing background environment information. This leads to the neural network overly relying on map information， causing the network to believe that the target will only appear in specific terrain areas， ultimately resulting in overfitting. In the newly proposed RMDN-V2 algorithm， targeted design is made to address this issue， with a focus on distinguishing the importance of radar information and map information in the entire detection process. This enables neural networks to autonomously learn the weights of radar information and map information features， allowing them to better utilize map information for auxiliary detection of radar targets， rather than just remembering special terrain areas， alleviating the existing overfitting phenomenon. In addition， this article further analyzes the impact of input data RD peak area map size and peak map size selection on the detection performance of RMDN class algorithms. The experimental results of the final drone radar test data show that the targeted optimization carried out in this article is effective. The RMDN-V2 algorithm has shown good detection performance for both types of drone target detection， especially for the smaller and more difficult to detect DJI Mavic 2 Pro. The detection performance of RMDN-V2 has been significantly improved， with a false alarm rate of$\mathrm{1}\times {\mathrm{10}}^{-\mathrm{6}}$， the detection rate of RMDN-V2 for DJI Mavic 2 Pro is about 30% higher than RMDN and about 50% higher than CFAR detection. Through comparison， it was found that the detection performance of the algorithm in this article is also better than some mainstream deep learning radar target detection algorithms currently available. This article proposes a new algorithm to solve the current problem of low slow small target detection.