Related Experiment Video
Updated: Sep 9, 2026
Computer Vision-Based Biomass Estimation for Invasive Plants
Published on: February 9, 2024
MFK-Net: a computationally efficient Mamba-Fourier-KAN hybrid architecture for UAV-based crop classification
1School of Computer Science and Technology, Jilin University, Changchun, China.
Introduction:
Accurate crop classification from unmanned aerial vehicle (UAV) imagery is essential for precision agriculture and food security monitoring. Existing approaches face challenges in balancing computational efficiency with classification accuracy, particularly when dealing with fine-grained crop categories and complex aerial backgrounds.
Methods:
This paper proposes MFK-Net, a computationally efficient hybrid architecture that integrates state-space models (Mamba), Fourier frequency-domain enhancement, and Kolmogorov-Arnold Networks (KAN) for efficient UAV crop classification. The proposed framework employs a streamlined Mamba backbone to capture long-range spatial dependencies with linear complexity, a learnable Fourier filtering module to enhance textural and boundary features in the frequency domain, and a KAN-based classification head to improve nonlinear decision boundaries.
Results:
Evaluated on the publicly available UAV-CM benchmark dataset comprising 7,666 images across 9 crop and 2 background categories, MFKNet achieves 93.67% accuracy, outperforming ResNet50 (87.34%), ViT-Small (89.42%), and SwinTiny (91.56%) while maintaining competitive efficiency with 31.33M parameters and 3.78G FLOPs. Extensive ablation studies demonstrate the complementary contributions of each component. Cross-dataset evaluation on the PlantDoc plant disease detection dataset demonstrates transfer learning capability with 95.09% macro-averaged accuracy. Furthermore, cross-dataset evaluation on the Martell Forest UAV tree species dataset confirms robust generalization to temperate forest vegetation with 89.45% macro-averaged accuracy, outperforming ResNet50 (83.15%) and SwinTiny (87.68%) on this distinct geographic domain and vegetation type.
Discussion:
Five-fold cross-validation on UAV-CM yields 93.41% ± 0.31% mean accuracy, confirming result stability. Comprehensive sensitivity analyses of KAN hyperparameters and Fourier filter configurations provide practical guidance for deployment.