| 13 Nov 2025
DTL-IceNet: A Dual-Task Learning Architecture with Multi-Scale Fusion Mechanisms for Enhanced Ice Detection on Transmission Lines
Abstract. Icing on transmission lines can significantly impact the stable operation of the power system. Deep learning-based ice image recognition is effective but remains vulnerable to background interference and noise, degrading accuracy. Moreover, when detecting ice thickness, the 2D nature of ice images introduces spatial limitations in representing the 3D ice state, which can lead to detection errors caused by a single viewpoint. To tackle the aforementioned challenges, this paper proposes DTL-IceNet (Dual-Task Learning Ice Detection Network), a transmission line icing detection network based on a dual-task learning framework, designed to accurately identify both the type and thickness of ice on overhead transmission lines. DTL-IceNet incorporates a multi-branch structured ice coating recognition module, ResSepNet (Residual & Depth-Separable Convolution Network), which segments the background and conductor areas to mitigate the influence of background noise. Additionally, a semantic segmentation module, MOMSA-SegNet (MobileOne & Multi-Scale Attention Segmentation Network) is designed to segment the ice-covered areas in both the main and side views of the image. The multi-scale attention mechanism is employed to extract spatial features from the raw icing image. When calculating ice thickness, the multi-scale fusion and correction optimization are adopted to enhance the algorithm. Experimental results show that compared with other models, the proposed method achieves an improvement of 4.17 % in icing type identification accuracy and a MAPE of 11.82 % in icing thickness detection. The application of this approach is crucial for reducing the hazards caused by ice coating on transmission lines and improving the stability of the power grid.
This paper proposes DTL-IceNet, a dual-task learning–based network for detecting icing on overhead transmission lines, designed to accurately identify both the type and the thickness of ice. The study demonstrates that, by segmenting the background and isolating the ice-covered regions, the method achieves high accuracy in distinguishing different icing types. Furthermore, based on the segmentation and classification outputs, the authors estimate ice thickness by incorporating meteorological data, and the reported results indicate strong performance, highlighting a promising direction for further research. In addition, the authors’ decision to release the dataset and source code is particularly commendable and will significantly benefit the research community.
Major comments
It is recommended to supplement the Summarize section or add a dedicated Discussion section. This section should include a more in-depth analysis of the reasons why the model performs well or fails under certain conditions, a discussion of the model’s limitations (for example, its performance under significant terrain variations or extreme weather conditions not represented in the dataset), and a more balanced interpretation of the results in the context of existing literature.
The manuscript presents two distinct results: (1) MOMSA-SegNet achieves the highest segmentation mIoU, and (2) the overall framework reports a thickness MAPE of 11.82%. To clarify the relationship between these two findings, additional controlled experiments would be helpful. In particular, demonstrating that, under the same test set and using the same thickness estimation procedure, the proposed segmentation model yields a consistently lower thickness estimation error compared with the other segmentation models discussed in the paper would provide more direct evidence of its contribution. Without such comparisons, the extent to which the segmentation module influences the final estimation accuracy remains uncertain.
Regarding environmental conditions, although the segmentation component includes descriptions of performance under different weather scenarios, the influence of these varying conditions on the final thickness estimation results is not examined. Expanding the Discussion section to include an evaluation of thickness estimation performance across different weather conditions would help provide a more complete understanding of the method’s behavior.
The manuscript claims to propose a comprehensive “dual-view” solution. While the approach of collecting real thickness data in controlled field experiments is understandable and commendable, the current experimental setup does not sufficiently validate the core contribution of the method, and instead highlights certain limitations. Specifically, the final performance evaluation is conducted on a restricted, single-view version of the system. This creates a substantial mismatch between the claimed capabilities and the empirical validation.We note that the authors provide “site conditions” as a rationale for this choice. However, this results in the core claim of the method—that leveraging multi-view structures from a single image enhances information capture—remaining unverified in thickness estimation experiments. By effectively omitting the higher-error components during validation, a critical question arises: does the reported thickness accuracy truly reflect the capability of the complete main-view and side-view system, or does it primarily represent performance in a simplified main-view scenario, which conveniently avoids the error propagation associated with the less accurate side-view segmentation?
On the other hand, achieving strong final results does not, by itself, validate the correctness or effectiveness of the front-end image segmentation plus area ratio approach. It primarily demonstrates the strength of the back-end correction module. Only with supplementary ablation experiments can it be convincingly shown that meteorological data and image information are complementary and both necessary, thereby substantiating the true value of the fusion framework.
Minor comments
Line 141 and Fig9 “glaz”->”glaze”
Table 1 Consider rephrase the table title.
Fig13 Consider rephrase the figure title.