CORRELATION-BASED FUSION OF FEATURES FOR SATELLITE IMAGE OBJECT CLASSIFICATION

Аннотация

This study analyzes feature fusion methods based on Convolutional Neural Networks (CNN) and Canonical Correlation Analysis to enhance the effectiveness of vehicle detection and classification in remote sensing images. The experimental investigations were conducted on the Cars Overhead With Context (COWC) dataset. Initially, fine-tuning was performed on ResNet18, DenseNet121, and VGG16 models using 20% of the data, and feature vectors were extracted from their output layers. Each model was evaluated individually, as well as through feature fusion using various statistical approaches–CCA (Canonical Correlation Analysis), DCCA (Discriminative Canonical Correlation Analysis), MCCA (Multiple Canonical Correlation Analysis), and DMCCA (Discriminative Multiple Canonical Correlation Analysis). The results demonstrated that among individual CNN models, DenseNet121 achieved the highest accuracy (≈87.45%). However, the features fused via DMCCA, when classified with an SVM, achieved the best overall performance with an accuracy of approximately 91.52%, along with superior precision, recall, and F1-score values