To address the problems of large number of model parameters and poor detail extraction in DeepLabv3+ for optical remote sensing image road extraction task , a light-weight road extraction model L-DeepLabv3+ is proposed to improve DeepLabv3+. Firstly , the number of model parameters is reduced by replacing the backbone network with MobileNetv2 ; secondly , an improved void space convolutional pooling pyramid module is designed in the coding layer. This module enhances the model feature expression capability by embedding a channel space parallel attention module and YOLOF module , and replaces the normal convolution with deep separable convolution to further reduce the number of model parameters ; Finally , Dice_loss and Focal _loss are combined as loss functions to solve the positive and nega- tive sample imbalance problem. The experimental results show that L-DeepLabv3+ achieves 68. 40% intersection ratio and 82. 67% pixel accuracy for road extraction on DeepGlobe Road dataset , and the number of model parameters is on- ly 5. 63 MB , and the FPS reaches 72. 3 , which is a significant improvement compared with other models , and achieves a better balance between model accuracy and light weight.