Automatic recognition of wheat growth stages with a lightweight multimodal data fusion network

Abstract

Accurate growth stage recognition is vital for optimising crop inputs and improving yield efficiency. However, single-modal methods often fail to distinguish phenologically adjacent stages due to canopy similarity, spectral saturation, or structural ambiguity, leading to mistimed agronomic actions and resource loss. Moreover, the high computational demands of complex deep neural networks hinder practical implementation. To address this challenge, this paper proposes a lightweight multimodal data fusion network for wheat growth stage recognition. Specifically, the RGB images, multispectral (MS) data, and digital surface model (DSM) acquired by Unmanned Aerial Vehicles (UAV), along with derived spectral vegetation index (VI), are used to capture multimodal canopy features, including colour, spectral reflectance, and spatial structure. Furthermore, the approach utilises MobileNetV3-Small, a lightweight convolutional neural network, as the backbone to construct the multimodal data fusion framework. This framework enables efficient feature extraction and integration, achieving precise and rapid wheat growth stage recognition with minimal computational overhead. The results demonstrate that, compared to single-modal models, the proposed multimodal fusion model significantly enhances growth stage recognition accuracy, achieving an accuracy of 99.57 %, a precision of 99.58 %, a recall of 99.57 %, and an F1 score of 99.57 %. Notably, it improves stage differentiation in critical transitions such as Booting to Heading, reducing field misclassification risks and supporting quick decision-making. Comparative analysis with MobileNetV3-Large, ResNet-18, MNASNet, EfficientNet-B0, and ConvNeXt-Tiny demonstrates that MobileNetV3-Small offers the best trade-off between accuracy and resource efficiency, with only 1.53 M parameters and an inference time of 6.03 ms on RTX 4090 and 25.11 ms on Jetson Orin NX. This efficiency enables real-time deployment on resource-constrained edge devices, such as onboard UAV processors or in-field embedded systems. Overall, this study effectively overcomes the challenges of recognising adjacent growth stages and computational constraints, offering a robust theoretical foundation and an efficient, accurate solution for wheat growth stage recognition.

Plant phenotyping relevance

UAV画像・マルチスペクトル・DSMを用いて小麦の生育段階という植物状態を推定する融合手法を開発・評価しており、表現型取得・抽出が研究の中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.