Comparative Performance Analysis of Machine Learning Kernels for FM Spectrum Monitoring: A Grid Search vs. Random Search Optimization Study

Abstract Effective spectrum monitoring is critical for detecting unlicensed FM broadcasts and managing interference in congested frequency bands. Support Vector Machines (SVMs) have emerged as promising classifiers for this domain, yet kernel selection and hyperparameter optimization remain underexplored for real-world regulatory applications. This study presents a comprehensive comparative analysis of four SVM kernels (RBF, Linear, Polynomial, and Sigmoid) optimized using GridSearchCV and RandomizedSearchCV on a three-year FM broadcast dataset (2021–2023) from Nigeria. The dataset comprises 3169 pre-processed transmission records, consisting of assigned frequency, band occupancy, and multiplex (MPX) features derived from Nigeria's National Broadcasting Commission monitoring operations. Results demonstrate that the RBF kernel achieves superior performance across all metrics (cross-validated accuracy: 99.96%, precision: 99.38%, recall: 99.63%, F1-score: 99.50% on the training set; held-out test-set evaluation yields 99.53% accuracy with 100% recall and 0.53% false alarm rate), significantly outperforming Linear (accuracy: 96.2%), Polynomial (accuracy: 95.8%), and Sigmoid (cross-validated accuracy: 87.3%) kernels. Critically, RandomizedSearchCV achieves equivalent detection performance to GridSearchCV while reducing computational time by approximately 42–48%, making it more suitable for real-time regulatory monitoring. The optimized model exhibits zero false negatives and 0.53% false alarm rate, with successful cross-regional validation across three additional Nigerian regions demonstrating less than 2% performance degradation. Feature importance analysis reveals that band occupancy and MPX signal contribute 73% of classification power, validating regulatory monitoring priorities. The findings provide actionable insights for spectrum regulatory agencies seeking to deploy machine learning-based monitoring systems with an optimal balance between accuracy and computational efficiency, achieving 87% cost reduction compared to manual monitoring approaches.