Mixture of Receptive Fields Fusion for Accurate Rice Panicle Segmentation and Yield Prediction in UAV-Based Remote Sensing

Abstract

Unmanned aerial vehicle (UAV)-based remote sensing enables high-throughput rice phenotyping, yet precise panicle segmentation remains challenging. Conventional plot-level vegetation indices (VIs) or panicle counting proxies inherently discard pixel-level structural information, rendering them incapable of organ-level morphological analysis. While ultrahigh-resolution (UHR) cameras now enable fine-grained segmentation, three interrelated bottlenecks persist: extreme scale heterogeneity across genotypes induces receptive field mismatch, dense canopy occlusion coupled with spectral similarity between withered leaves and panicles causes severe foreground–background confusion, and gigapixel-scale orthomosaics for thousands of microplots impose prohibitive computational burdens. To address these, we propose mixture of receptive fields fusion (MiRF), where dynamic preview convolution (DyPConv) adaptively selects optimal dilation rates to capture critical local contextual information across diverse genotypic panicle morphologies while compressing computation to only K active branches via hard path pruning; frequency-aware refinement (FAR) leverages Laplacian high-frequency components with learnable gating to explicitly magnify textural disparities between panicles and cluttered backgrounds of withered leaves; and efficient cross-scale feature aggregator (ECFA) employs globally shared bidirectional fusion to reduce computational burdens when processing gigapixel-scale orthomosaics while preserving multiscale context. Experiments on two public datasets achieve state-of-the-art intersection over union (IoU)-Panicles (+3.48% and +3.11%) and $F1$ -scores (+2.83% and +2.31%) with significantly reduced complexity. Cross-season validation on 610 varieties spanning 2024–2025 achieves R2 exceeding 0.826 for yield prediction, supporting MiRF’s generalization across diverse genotypes. Edge deployment experiments demonstrate its superiority in real-time processing.