Accurate identification of dust pollution is essential for environmental monitoring and management. However, the traditional U-Net model struggles with uneven particle scales and complex background interference in dust images, resulting in suboptimal segmentation performance. To address these challenges, we propose an improved U-Net model with a pyramid structure, integrating a dynamic multi-scale pyramid pooling module, a channel attention mechanism, a data augmentation strategy, and a dust-aware loss function. This design aims to enhance the model’s recognition capability and segmentation accuracy in complex dusty scenes. Experimental results show that the proposed model achieves an Intersection over Union (IoU) of 87.5% and a Dice coefficient of 93%, significantly reducing false detections in sparse dust regions and complex backgrounds. Additionally, the predicted masks align closely with manual annotations, enabling accurate quantification of dust coverage and providing reliable data support for pollution source localization, concentration assessment, and intelligent early warning systems.