High-throughput end-to-end aphid honeydew excretion behavior recognition method based on rapid adaptive motion-feature fusion

Abstract

Introduction: Aphids are significant agricultural pests and vectors of plant viruses. Their Honeydew Excretion(HE) behavior holds critical importance for investigating feeding activities and evaluating plant resistance levels. Addressing the challenges of suboptimal efficiency, inadequate real-time capability, and cumbersome operational procedures inherent in conventional manual and chemical detection methodologies, this research introduces an end-to-end multi-target behavior detection framework. This framework integrates spatiotemporal motion features with deep learning architectures to enhance detection accuracy and operational efficacy.

Methods: This study established the first fine-grained dataset encompassing aphid Crawling Locomotion(CL), Leg Flicking(LF), and HE behaviors, offering standardized samples for algorithm training. A rapid adaptive motion feature fusion algorithm was developed to accurately extract high-granularity spatiotemporal motion features. Simultaneously, the RT-DETR detection model underwent deep optimization: a spline-based adaptive nonlinear activation function was introduced, and the Kolmogorov-Arnold network was integrated into the deep feature stage of the ResNet50 backbone network to form the RK50 module. These modifications enhanced the model's capability to capture complex spatial relationships and subtle features.

Results and discussion: Experimental results demonstrated that the proposed framework achieved an average precision of 85.9%. Compared with the model excluding the RK50 module, the mAP50 improved by 2.9%, and its performance in detecting small-target honeydew significantly surpassed mainstream algorithms. This study presents an innovative solution for automated monitoring of aphids' fine-grained behaviors and provides a reference for insect behavior recognition research. The datasets, codes, and model weights were made available on GitHub (https://github.com/kuieless/RAMF-Aphid-Honeydew-Excretion-Behavior-Recognition).

Keywords: Kolmogorov-Arnold networks; RT-DETR-RK50; aphid behavior recognition; honeydew excretion detection; rapid adaptive motion feature fusion.

Figure 1

Dataset images of different population densities and light intensities.

Figure 2

Flowchart of the RAMF framework.

Figure 3

RAMF processing workflow and visualization diagram.

Figure 4

Aphid behavior classification: (a) crawling locomotion; (b) Leg Flicking; (c) Aphid Honeydew Excretion.

Figure 5

Architecture diagram of RT-DETR-RK50 network.

Figure 6

Network architecture diagrams of normal blocks and RK50 blocks.

Figure 7

Network model diagram of KAN conv.

Figure 8

Visual representation of motion feature extraction across different temporal windows.

Figure 9

Feature activation heatmap comparison across RT-DETR-RK50 variants: (a) R50, (b) RK50-1, (c) RK50-2 (Ours), (d) RK50-3.

Figure 10

Confusion matrix comparison of detection models across different Aphid behavior categories.

Figure 11

Comparison of training loss curves across different model architectures.

Figure 12

Detection results based on the model with the highest mAP50 metric.

Figure 13

Grad-CAM++ visualization results under varying lighting conditions and Aphid densities: (a) Different lighting only, (b) Different density only, (c) HE target detection effect under different lighting and density.

Figure 14

Stage-wise honeydew excreting detection pipeline and results.