In addressing the issue of weak target detection at sea, most existing feature detection methods are designed for scenarios with low sea states and small grazing angles. Under high sea states and large grazing angles, variations in scattering mechanisms lead to changes in feature characteristics, resulting in performance degradation when these methods are applied directly due to scene mismatch. To address this, this paper employs four quantitative metrics—mean feature value, coefficient of variation, Bhattacharyya distance, and Spearman correlation coefficient—to analyze the centrality, variability, separability, and correlations of nine features in the time, frequency, and time-frequency domains under varying sea states and grazing angles. The study reveals that, with increasing sea state and changing grazing angles, the separability of time-frequency features, especially the time-frequency ridge accumulation, declines more gradually than other features, and feature correlations generally weaken. These findings provide a reference for joint feature detection in complex scenarios. To optimize feature application, the Spearman correlation coefficient matrix was transformed into a generalized distance matrix, and spectral clustering was used to group features with strong correlations. Feature selection was then performed from the clusters based on mean feature value, coefficient of variation, and Bhattacharyya distance, yielding an optimal feature set for the current scenario. Validation on the SDRDSP dataset under sea states 4–5 showed that the proposed method achieved an average detection probability 10.64% higher than existing methods. Further validation on the Yantai angle airborne test dataset, with grazing angles ranging from 62° to 82°, showed an average detection probability increase of 10.07% over existing methods.