Experimental analysis on the use of RADAR images for compressive strength assessment in concrete slabs using machine learning algorithms

This study explores the feasibility of using satellite-borne radar imagery to non-destructively estimate the compressive strength of concrete slabs through machine learning classification. Traditional compressive strength testing methods are often invasive and impractical for in-service structures, prompting the need for scalable, image-based alternatives. In this experiment, high-resolution Sentinel-1 Ground Range Detected (GRD) radar imagery was analysed to extract image features—specifically backscatter intensity, lineament density, and texture metrics derived from gray-level co-occurrence matrices (GLCM). These features were used to train and validate three supervised machine learning models, being (i) Support Vector Machine (SVM), (ii) Random Forest (RF), and (iii) k-Nearest Neighbour (k-NN), against compressive strength labels obtained via in-situ Ultrasonic Pulse Velocity (UPV) testing. Among the classifiers, SVM achieved the highest accuracy (91.7%) and F1-score (92.5%) on the validation set, and 100% accuracy on the test set, demonstrating a robust relationship between radar-derived features and physical concrete integrity. This approach presents a suitable, non-contact alternative to conventional destructive testing methods, with potential for real-time structural health monitoring in complex urban environments.