A multi-source remote sensing image fusion and object detection network based on improved depthwise separable convolution and a multi-scale feature extraction module was proposed to address the limitation of convolutional downsampling in feature extraction and the problem of traditional feature-level fusion methods failing to fully leverage the complementary advantages of multi-source remote sensing data. A dual-branch separable convolution module was designed to enhance deep semantic feature representation through depthwise convolution and residual connections, thereby improving discriminative performance under complex backgrounds. Furthermore, a global-local adaptive feature fusion module was constructed, where feature maps were decomposed into different dimensional components using separable convolution to capture global structures and local details separately. These features were then fused via an adaptive mechanism to achieve cross-source information complementarity and multi-scale feature collaboration. Experiments on the VEDAI multi-source dataset demonstrated that the proposed method achieved a mean average precision (mAP) of 82.80%, which was 2.00 percentage points higher than that of ICAfusion, while also outperforming YOLOrs, YOLOfusion, SuperYOLO, and MF-YOLO. The network shows high effectiveness in feature-level fusion of multi-source remote sensing images and yields significant performance improvements in object detection tasks.