Harnessing wind energy is critically important for the trajectory of future energy development. To maximize wind energy capture, wind farms are frequently deployed in high-latitude regions, a practice that unfortunately elevates the susceptibility to icing phenomena. Conventional methodologies for blade icing detection typically rely on manual inspections or external sensor-based apparatus, yet these approaches are constrained by the variability of human judgment and introduce supplementary expenditures. Model-centric paradigms are heavily predicated on established domain expertise and are susceptible to interpretive inaccuracies. Data-driven strategies proffer promising alternatives but necessitate substantial volumes of annotated training data, which are not commonly accessible. Furthermore, datasets amassed for icing detection are inherently class-imbalanced, given that wind turbines predominantly operate under non-icing conditions. To surmount these impediments, this paper introduces a novel Deep Class-Imbalanced Semi-Supervised (DCISS) model tailored for the estimation of blade icing conditions. DCISS synergistically integrates class-imbalanced and semi-supervised learning (SSL) principles, employing a prototypical network architecture capable of feature re-balancing and quantifying the similarities between labeled and unlabeled instances. Moreover, a channel calibration attention module is proposed to augment the model’s capacity to discern salient features from raw sensor data. The efficacy of the proposed model has been rigorously evaluated utilizing blade icing datasets procured from three distinct wind turbines. In comparative analyses against both classical anomaly detection techniques and state-of-the-art SSL algorithms, DCISS exhibits pronounced advantages in terms of detection accuracy. When benchmarked against five diverse class-imbalanced loss functions, the proposed DCISS demonstrates robust competitiveness. The generalization capability and practical applicability of the DCISS model are further substantiated through an online estimation use case.