Advancing Radar-Based AI: The Power of Preprocessing in Machine Learning Models

Abstract

Radar-classification-based systems are particularly critical in automotive safety and intelligent transportation systems, but environmental noise, signal overlap, and class imbalance substantially reduce performance productivity. This experiment measured the effect of Feature Selection and Synthetic Minority Oversampling Technique (SMOTE) on radar-classification-based systems through Random Forest (RF) and Support Vector Machine (SVM) models. A dataset of 45,000 spectrograms produced from radar were used in the analysis, with results on the data set showing SVM producing accuracies of 65.16% and RF producing 62.08%, both produced noisy and imbalanced data results. Post Feature Selection (top 10,000 features selected) and SMOTE (1:4 oversampling ratio), the SVM model produced an accuracy of 69.38% and RF model produced an accuracy of 63.20% while demonstrating ‘more robustness and generalizability’, all of which provide visual support that preprocessing the radar spectrograms has a crucial role in improving classification accuracy.