A Remote Sensing Road Extraction Model Based on Improved Deeplabv3+ with Adaptive Multi-Scale Feature Fusion Pyramid

In the domain of road extraction from remote sensing images, semantic segmentation models often face challenges such as high computational complexity, low accuracy in road extraction in complex scenes, and poor road connectivity extraction. To address these issues, we propose a remote sensing road extraction model based on the MDA-Deeplabv3+ with an adaptive multi-scale feature fusion pyramid(AMFFP). To reduce computational load and lower the model’s complexity, the MobilenetV2 network is introduced to replace the Xception network in the backbone. Additionally, strip pooling is incorporated at the network’s bottom to simulate elongated roads, and a dense atrous spatial pyramid pooling (DenseASPP) is used to increase the network’s receptive field, which is employed to enhance road connectivity extraction. Finally, an adaptive multi-scale feature fusion pyramid is applied to the low-level features of the network to integrate information from different scales. Experiments conducted on the Massachusetts roads datasets and CHN6-CUG datasets demonstrate that the MFP-Deeplabv3+ model achieves improved performance metrics compared to the original model, resulting in more continuous and complete road extraction in the output images.