Fine-Grained UAV Type Recognition via Multi-Modal Feature Fusion

Reliable fine-grained identification of Unmanned Aerial Vehicle (UAV) types in complex low-altitude airspace is a critical prerequisite for effective counter-UAV operations. The radar and optical modalities used in this work are linked only through shared class labels rather than sample-paired—so that all reported accuracies reflect class-level correspondence and not any temporal or geometric co-registration. Existing methods predominantly exploit a single sensing modality—narrowband micro-Doppler radar, wideband High-Resolution Range Profile (HRRP), or optical imagery—each of which exhibits fundamental limitations in distinguishing subtle inter-class differences under realistic operational conditions. This paper presents a comprehensive multi-modal UAV recognition framework that fuses narrowband micro-Doppler spectrograms, wideband HRRP scattering profiles, and optical images. We systematically analyze the underlying physical principles of each modality: the micro-Doppler modulation induced by rotor-blade rotation, the range-resolved scattering center distributions, and the shape-texture cues. To achieve robust fusion, three dedicated deep neural network encoders extract modality-specific features. These features are then aligned via a novel Cross-Attention Feature Alignment (CAFA) mechanism and adaptively weighted through Dynamic Confidence Gating (DCG). Extensive experiments on field-collected dual-band (S/Ku) narrowband and HRRP radar data together with publicly sourced optical imagery, along with a cross-domain (Ku/L) HRRP dataset, demonstrate that: (1) single-modal recognition achieves 81.8% (micro-Doppler data), 65.6% (HRRP data), and 73.0% (optical data) accuracy, respectively; and (2) the proposed multi-modal framework achieves 94.06% ± 1.18 accuracy (mean ± sd over four training seeds), a 12.3-percentage-point improvement over the best single-modal network (81.8%). Of that margin 10.4 points are already obtained by plain feature concatenation (92.16% ± 1.12) and 1.9 points by the proposed combination, a margin that exceeds the seed-to-seed spread of either model. The results confirm that the cross-attention mechanism learns class-level statistical correspondences among modality-specific features—associating semantically related scattering structures and optical regions of the same target class—rather than physically co-registering radar scattering centres with optical image pixels. The DCG module realises its weighting in log-odds space and is regularised by a batch-level load-balancing term that prevents the gate from collapsing onto a single branch; this is what makes the adaptive weighting effective: the learned gate beats a plain average of the same three branch posteriors by 1.9 points with all four seeds positive, and under HRRP degradation to 15 dB the fused framework is 1.9 percentage points more accurate than naive concatenation, rising to 2.76 points under optical additive noise of σ = 0.4 (four seeds).

Authors

Publication Details

Journal
Sensors
Published
2026-11-18
DOI
https://doi.org/10.3390/s26227287
Primary Topic
Advanced SAR Imaging Techniques
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Fine-Grained UAV Type Recognition via Multi-Modal Feature Fusion

Tian Xilan, Jun Zhang
1 links
Sensors
Advanced SAR Imaging Techniques
article

Fine-Grained UAV Type Recognition via Multi-Modal Feature Fusion

Tian Xilan, Jun Zhang
article en

Abstract

Reliable fine-grained identification of Unmanned Aerial Vehicle (UAV) types in complex low-altitude airspace is a critical prerequisite for effective counter-UAV operations. The radar and optical modalities used in this work are linked only through shared class labels rather than sample-paired—so that all reported accuracies reflect class-level correspondence and not any temporal or geometric co-registration. Existing methods predominantly exploit a single sensing modality—narrowband micro-Doppler radar, wideband High-Resolution Range Profile (HRRP), or optical imagery—each of which exhibits fundamental limitations in distinguishing subtle inter-class differences under realistic operational conditions. This paper presents a comprehensive multi-modal UAV recognition framework that fuses narrowband micro-Doppler spectrograms, wideband HRRP scattering profiles, and optical images. We systematically analyze the underlying physical principles of each modality: the micro-Doppler modulation induced by rotor-blade rotation, the range-resolved scattering center distributions, and the shape-texture cues. To achieve robust fusion, three dedicated deep neural network encoders extract modality-specific features. These features are then aligned via a novel Cross-Attention Feature Alignment (CAFA) mechanism and adaptively weighted through Dynamic Confidence Gating (DCG). Extensive experiments on field-collected dual-band (S/Ku) narrowband and HRRP radar data together with publicly sourced optical imagery, along with a cross-domain (Ku/L) HRRP dataset, demonstrate that: (1) single-modal recognition achieves 81.8% (micro-Doppler data), 65.6% (HRRP data), and 73.0% (optical data) accuracy, respectively; and (2) the proposed multi-modal framework achieves 94.06% ± 1.18 accuracy (mean ± sd over four training seeds), a 12.3-percentage-point improvement over the best single-modal network (81.8%). Of that margin 10.4 points are already obtained by plain feature concatenation (92.16% ± 1.12) and 1.9 points by the proposed combination, a margin that exceeds the seed-to-seed spread of either model. The results confirm that the cross-attention mechanism learns class-level statistical correspondences among modality-specific features—associating semantically related scattering structures and optical regions of the same target class—rather than physically co-registering radar scattering centres with optical image pixels. The DCG module realises its weighting in log-odds space and is regularised by a batch-level load-balancing term that prevents the gate from collapsing onto a single branch; this is what makes the adaptive weighting effective: the learned gate beats a plain average of the same three branch posteriors by 1.9 points with all four seeds positive, and under HRRP degradation to 15 dB the fused framework is 1.9 percentage points more accurate than naive concatenation, rising to 2.76 points under optical additive noise of σ = 0.4 (four seeds).

SensorsVol. 26(22)
National Key Research and Development Program of China
Advanced SAR Imaging Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon