Remote sensing, especially high-resolution remote sensing, has been increasingly applied in emergency management due to its advantages, including instantaneous imaging, wide coverage, dynamic updates, minimal constraints from ground conditions, long-term monitoring capabilities, and comprehensive information acquisition. It makes its application in emergency management highly significant. Based on the typical real-world case of the MH370 remote sensing image search and rescue, this study addresses challenges such as the vast search area, severe background interference, diverse target information, and weak feature representation. By leveraging an application-oriented spaceborne remote sensing sample database, we propose an adaptive Gaussian kernel support vector machine (SVM) search technique for typical maritime accident targets that integrates multiple features. This method first extracts multiple features to form feature vectors and dynamically adjusts the kernel parameters of each sample point based on local data density. It enables automatic adaptation, thereby improving classification accuracy and model generalization while addressing the overfitting and underfitting issues caused by fixed kernel parameters in traditional SVMs. Additionally, the study employs multispectral cosine similarity to further assess the resemblance between floating objects and other targets on the sea surface, verifying the reliability of suspected areas where the missing aircraft might have been located.