In response to the problems that segmentation is low in accuracy and detection is slow at speed in detecting drone aerial images in key areas by using existing semantic segmentation algorithm, an im proved PP-LiteSeg lightweight drone image semantic segmentation algorithm is proposed. The algorithm, first, is to design a composite attention fusion module in which the parameter free attention mechanism Si mAM is introduced into the unified attention fusion module to enhance global contextual information and improve the information richness of output features. Afterwards, the model parameters are reduced through replacing the calculation method of convolution in the backbone network from ordinary convolu tion to a combination of partial convolution and small-scale convolution kernels. At the same time, a new backbone network SDTCM_PNet is designed to further enhance the lightweighting of the model by chan ging the feature concatenation method of short-term dense connection modules in the multi-layer receptive field of the backbone network. The experimental results conducted on the self-collected drone aerial image dataset show that the algorithm proposed in this paper is valid. Simultaneously, the algorithm is to be de ployed and tested on embedded devices, and the algorithm also meets the needs of real-time.