Deep neural network-driven adaptive fusion recognition method for radar-optical heterogeneous remote sensing imagery

Abstract The fusion of radar and optical remote sensing imagery presents significant challenges due to fundamental differences in imaging mechanisms and feature representations. This paper proposes an adaptive fusion recognition framework based on deep neural networks for heterogeneous radar-optical imagery. The framework comprises three core components: a heterogeneous feature extraction and alignment module that projects SAR and optical features into a shared semantic space through contrastive learning, a cross-modal adaptive fusion mechanism employing dual-attention architecture with dynamic weight generation to enable context-sensitive modality integration, and an end-to-end recognition network that jointly optimizes all processing stages. To probe the framework from multiple angles, we ran a comprehensive experimental campaign on the SEN1-2 benchmark, which spans head-to-head comparisons against ten representative fusion baselines, ablations isolating each module as well as the GLCM texture branch, a sensitivity sweep over the contrastive temperature parameter, stress tests under additive speckle noise and single-modality missing scenarios, and an efficiency audit reporting parameter count and inference latency. The proposed method reaches 91.38% overall accuracy with a Kappa coefficient of 0.892 and macro F1 score of 90.45%, surpassing competitive baselines while holding inference cost to 11.4 ms per sample. Ablation studies confirm the effectiveness of each component, with feature alignment contributing 3.45% accuracy improvement and the complete adaptive fusion mechanism providing additional 5.46% gains over baseline concatenation. The dynamic weighting strategy exhibits robust performance under degraded input conditions by automatically adjusting modality contributions based on estimated reliability.