Traditional earthquake data processing often focuses on the application of single features or methods, overlooking the rich multi-scale information and interrelationships inherent in seismic wave signals. Furthermore, the high noise, nonlinearity, and complexity of seismic data present significant challenges for automated, efficient processing and accurate earthquake prediction. To address these issues, this paper proposes a novel method for earthquake identification and forecasting based on multi-scale features and contrastive learning, termed MCLEC (Multi-scale Contrastive Learning-based Method for Earthquake Classification). This method aims to deeply uncover hidden patterns within seismic wave data and enhance the classification capability for earthquake events. Firstly, during the feature extraction phase, we utilize autocorrelation functions, time–frequency analysis, spectral analysis, and wavelet transform across multiple dimensions to comprehensively capture the spatiotemporal characteristics of seismic waveforms, thereby fully extracting the complex information embedded in seismic signals. Secondly, in the modeling phase, we employ a contrastive learning mechanism to maximize the distance between samples of different categories while ensuring the compactness of samples within the same category in the feature space, significantly boosting the model's discriminative power. Finally, experiments on relevant datasets demonstrate that MCLEC outperforms baseline methods, showcasing its superior performance and validating its effectiveness in complex seismic data analysis.