Integrating Stride Attention and Cross-Modality Fusion for UAV-Based Detection of Drought, Pest, and Disease Stress in Croplands

Yan Li, Yaze Wu, W. Wang, Huiyu Jin, Xiaohan Wu, Jinyuan Liu, Chen Hu, Chunli LvAgronomy 2025 · DOI

TL;DR

This paper introduces a novel "stride-cross-attention" mechanism for UAV-based agricultural disaster detection, fusing multispectral and RGB imagery to accurately identify drought, pests, and diseases. The proposed system achieves state-of-the-art performance, outperforming existing models by a significant margin, and is designed for efficient deployment in real-world field conditions.

What Problem Does This Paper Address?

Ensuring global food security hinges on timely and accurate detection of agricultural disasters like drought, pests, and diseases. These events can devastate crop yields and impact ecosystems. Traditional monitoring methods often struggle with accuracy, robustness, and scalability, especially in complex environments. While deep learning has shown promise, existing models often face challenges with:

  • Complex Field Conditions: Ambiguous disaster boundaries, diverse land cover, and heterogeneous backgrounds make feature extraction difficult.
  • Multimodal Data Fusion: Effectively integrating information from different sensor types (like RGB and multispectral imagery) is challenging due to variations in resolution and spatial scale.
  • Computational Constraints: Deploying sophisticated models on resource-limited UAV platforms for real-time monitoring is a significant hurdle.

This paper aims to address these challenges by developing a robust, efficient, and deployable system for agricultural disaster detection using multimodal UAV imagery.

Methods

The core innovation of this paper is the "stride-cross-attention" mechanism, designed to enhance feature extraction and multimodal fusion.

  • Stride Attention: This mechanism introduces a sparse, stride-based connectivity for attention. Instead of every patch attending to every other patch (as in standard self-attention), patches interact with others at a fixed stride (here, stride=2). This reduces computational complexity while maintaining a large receptive field and focusing on critical local regions. It's visualized as a periodic pattern of active connections, making it more efficient and sparser than local attention.
  • Cross-Attention: This module facilitates semantic fusion between different modalities (RGB and multispectral imagery). It treats one modality (e.g., RGB) as the "query" and the other (e.g., multispectral) as the "key" and "value." This allows the model to dynamically align and extract complementary information, overcoming limitations where modalities are treated in isolation.
  • Dual-Branch Architecture: The model employs a dual-branch structure. One branch uses the stride attention for global modeling, while a residual branch preserves local structural information. These branches are fused through element-wise addition.
  • CNN Classifier: A lightweight, three-layer convolutional neural network (CNN) classifier is used for the final disaster recognition, integrated with the attention modules.
  • Data Preprocessing and Augmentation: The system includes robust preprocessing steps like geometric correction and denoising, along with extensive data augmentation (MixUp, CutMix, GridMask, RandAugment) to enhance sample quality and diversity.

Datasets

The study utilized a custom dataset collected from representative wheat and maize fields in Dalad Banner, Ordos City, Inner Mongolia.

  • Data Source: UAVs equipped with synchronized multispectral (blue, green, red, red edge, near-infrared) and high-resolution RGB sensors.
  • Disaster Categories: Six annotated categories were used: mild, moderate, and severe drought, pest infestation, disease stress, and normal conditions.
  • Data Collection: Field trials were conducted over several months, covering various crops, growth stages, and disaster types. Ground truth was collected through expert annotation, field surveys, and meteorological data.
  • Dataset Size: After annotation and augmentation, the dataset comprised 2100 RGB samples, 1600 multispectral samples, and 380 orthomosaic samples.

Key Insights and Ideas

  • Stride Attention for Efficiency: The stride attention mechanism is a clever way to reduce computational cost in attention models without sacrificing too much performance. By selectively connecting patches, it achieves a better balance between accuracy and speed, crucial for UAV applications.
  • Cross-Attention for True Fusion: Simply concatenating or summing multimodal data isn't enough. Cross-attention enables a dynamic, query-key-value interaction that truly aligns and fuses information from different sensor types, leading to a more comprehensive understanding of crop health.
  • Multimodal Advantage in Ambiguity: In scenarios where single-sensor data might be ambiguous (e.g., similar yellowing in RGB for drought and pests), combining multispectral data via cross-attention provides distinct spectral cues that help resolve these ambiguities.
  • Deployability is Key: The paper emphasizes not just model performance but also its practical deployment on edge devices (like NVIDIA Jetson AGX Xavier), incorporating techniques like quantization-aware training and TensorRT acceleration.

Key Results

The proposed model achieved impressive results:

  • Overall Performance:
    • Accuracy: 93.2%
    • F1 Score: 92.7%
    • Precision: 93.5%
    • Recall: 92.4%
  • Comparison to Baselines: The proposed method significantly outperformed mainstream models like ResNet50, EfficientNet-B0, and ViT, showing improvements of over 8% in accuracy.
  • Ablation Studies:
    • Cross-Attention: Removing the cross-attention module led to a substantial performance drop (e.g., accuracy decreased from 93.2% to 84.6%), highlighting its critical role in multimodal fusion.
  • Stride Setting: A stride of 2 was found to offer the best balance between detection accuracy and inference speed (FPS), achieving 61.8 FPS with only a minor drop in accuracy compared to stride=1.
  • Edge Deployment: The system achieved 8-10 frames per second on an NVIDIA Jetson Xavier NX, with a preprocessing time of 0.12 seconds per image.

Limitations and Future Work

  • Dataset Specificity: The dataset was collected from a specific region (Ordos, Inner Mongolia) and within a limited seasonal window. The generalization ability of the model to other regions or environmental conditions requires further validation.
  • Limited Modalities: The study primarily used RGB and multispectral imagery. Future work could explore integrating other data sources like meteorological data, soil sensor readings, and agricultural machinery operation records.
  • Edge Deployment Enhancements: While the system is designed for edge deployment, computationally intensive tasks like image stitching and large-scale spatial context modeling are still executed on the server side. Future work aims to integrate these into the onboard inference pipeline.
  • Domain Adaptation: The authors suggest extending the system to more diverse agricultural scenarios and exploring domain adaptation techniques to enhance robustness and transferability.

Summary updated

Abstract

Timely and accurate detection of agricultural disasters is crucial for ensuring food security and enhancing post-disaster response efficiency. This paper proposes a deployable UAV-based multimodal agricultural disaster detection framework that integrates multispectral and RGB imagery to simultaneously capture the spectral responses and spatial structural features of affected crop regions. To this end, we design an innovative stride–cross-attention mechanism, in which stride attention is utilized for efficient spatial feature extraction, while cross-attention facilitates semantic fusion between heterogeneous modalities. The experimental data were collected from representative wheat and maize fields in Inner Mongolia, using UAVs equipped with synchronized multispectral (red, green, blue, red edge, near-infrared) and high-resolution RGB sensors. Through a combination of image preprocessing, geometric correction, and various augmentation strategies (e.g., MixUp, CutMix, GridMask, RandAugment), the quality and diversity of the training samples were significantly enhanced. The model trained on the constructed dataset achieved an accuracy of 93.2%, an F1 score of 92.7%, a precision of 93.5%, and a recall of 92.4%, substantially outperforming mainstream models such as ResNet50, EfficientNet-B0, and ViT across multiple evaluation metrics. Ablation studies further validated the critical role of the stride attention and cross-attention modules in performance improvement. This study demonstrates that the integration of lightweight attention mechanisms with multimodal UAV remote sensing imagery enables efficient, accurate, and scalable agricultural disaster detection under complex field conditions.
BibTeX
@inproceedings{li2025integrating,
  author = {Yan Li and Yaze Wu and W. Wang and Huiyu Jin and Xiaohan Wu and Jinyuan Liu and Chen Hu and Chunli Lv},
  title = {{Integrating Stride Attention and Cross-Modality Fusion for UAV-Based Detection of Drought, Pest, and Disease Stress in Croplands}},
  booktitle = {Agronomy},
  year = {2025},
  doi = {10.3390/agronomy15051199},
}
Figure