Abstract: To address the issues that L-band radar echoes are susceptible to strong clutter interference when detecting "low, slow, and small (LSS)" targets such as unmanned aerial vehicles (UAVs), and conventional high-precision classification models struggle to meet the requirements of low-cost edge deployment, this paper proposes a radar signal processing and classification framework based on an improved Rime Optimization Algorithm optimized Variational Mode Decomposition (CC-RIME-VMD) combined with low-complexity machine learning. Traditional Empirical Mode Decomposition (EMD) is prone to generating distortion spikes under strong clutter, while blindly presetting VMD parameters easily leads to micro-motion energy leakage. This paper introduces Circle chaotic mapping and a Cauchy mutation long-tail perturbation mechanism to improve the RIME algorithm, and adaptively obtains the optimal VMD parameters for signal purification by taking the minimum envelope entropy as the objective function. On this basis, three-dimensional pure physical fingerprint features—micro-Doppler bandwidth, logarithmic true energy, and spectral centroid—are extracted and cascaded with a Random Forest classifier for target recognition. Experimental results demonstrate that the proposed method effectively filters out environmental noise, increases the effective feature amplitude to 0.0546 (a 162.5% improvement compared with traditional VMD), and reduces the envelope entropy to 8.3978, completely eliminating the artificial low entropy phenomenon inherent in traditional algorithms. Under 5-fold cross-validation, this method significantly improves the average classification accuracy for four types of UAVs from 72.4% (EMD) and 77.0% (VMD) to 90.1%. While effectively reducing computational costs, this framework maintains a high recognition accuracy, providing an effective solution for the engineering deployment of low-power edge early warning radar systems.To address the issues that L-band radar echoes are susceptible to strong clutter interference when detecting "low, slow, and small" (LSS) targets such as unmanned aerial vehicles (UAVs), and that conventional high-precision classification models struggle to meet the requirements of low-cost edge deployment, a radar signal processing and classification framework based on an improved RIME optimization algorithm (CC-RIME) optimizing Variational Mode Decomposition (VMD) combined with low-complexity machine learning is proposed.Traditional Empirical Mode Decomposition (EMD) is prone to generating distortion spikes under strong clutter, while blindly presetting VMD parameters easily leads to micro-motion energy leakage. This paper introduces Circle chaotic mapping and a Cauchy mutation long-tail perturbation mechanism to improve the RIME algorithm. By utilizing the minimum envelope entropy as the objective function, the optimal VMD parameters are adaptively obtained for signal purification. On this basis, three-dimensional pure physical fingerprint features—micro-Doppler bandwidth, logarithmic true energy, and spectral centroid—are extracted and cascaded with a Random Forest classifier for target recognition.Experimental results demonstrate that the proposed method effectively filters out environmental noise, increases the effective feature amplitude to 0.0546 (a 162.5% increase compared to traditional VMD), and reduces the envelope entropy to 8.3978, completely eliminating the "artificial low entropy" phenomenon inherent in traditional algorithms. Under 5-fold cross-validation, this method achieves a significant leap in the average classification accuracy for multiple UAV types, rising from 72.4% (EMD) and 77.0% (VMD) to 90.1%. While effectively reducing computational costs, this framework maintains a high recognition accuracy, providing an effective solution for the engineering deployment of low-power edge early warning radar systems.