A transmittance-driven multi-task perception method for dust region analysis in photovoltaic images

Photovoltaic (PV) modules in real-world deployments are often affected by dust accumulation, leading to sunlight obstruction and reduced energy conversion efficiency. This paper proposes a transmittance-driven multi-task perception method for PV dust analysis. First, a physically interpretable dust synthesis mechanism is developed based on an optical blending model, enabling controllable generation of labeled data with pixel-wise transmittance maps and dust masks. Then, an improved multi-task network, termed photovoltaic dust perception UNet (PDP-UNet), is designed with a shared residual squeeze-and-excitation encoder and task-specific branches to jointly perform dust segmentation, transmittance regression, and dust type classification. Experiments show that the proposed method outperforms several mainstream multi-task baselines in Intersection over Union and transmittance error, and qualitative results on real PV images further demonstrate its applicability in practical inspection scenarios using unmanned aerial vehicles (UAVs) or fixed cameras.