Accurate Wheat Lodging Extraction from Multi-Channel UAV Images Using a Lightweight Network Model.

Abstract

The extraction of wheat lodging is of great significance to post-disaster agricultural production management, disaster assessment and insurance subsidies. At present, the recognition of lodging wheat in the actual complex field environment still has low accuracy and poor real-time performance. To overcome this gap, first, four-channel fusion images, including RGB and DSM (digital surface model), as well as RGB and ExG (excess green), were constructed based on the RGB image acquired from unmanned aerial vehicle (UAV). Second, a Mobile U-Net model that combined a lightweight neural network with a depthwise separable convolution and U-Net model was proposed. Finally, three data sets (RGB, RGB + DSM and RGB + ExG) were used to train, verify, test and evaluate the proposed model. The results of the experiment showed that the overall accuracy of lodging recognition based on RGB + DSM reached 88.99%, which is 11.8% higher than that of original RGB and 6.2% higher than that of RGB + ExG. In addition, our proposed model was superior to typical deep learning frameworks in terms of model parameters, processing speed and segmentation accuracy. The optimized Mobile U-Net model reached 9.49 million parameters, which was 27.3% and 33.3% faster than the FCN and U-Net models, respectively. Furthermore, for RGB + DSM wheat lodging extraction, the overall accuracy of Mobile U-Net was improved by 24.3% and 15.3% compared with FCN and U-Net, respectively. Therefore, the Mobile U-Net model using RGB + DSM could extract wheat lodging with higher accuracy, fewer parameters and stronger robustness.

Plant phenotyping relevance

UAV画像とDSM/ExGを用いてコムギ倒伏状態を抽出する軽量セグメンテーション手法を開発し、複数データセットおよび既存モデルと精度・速度比較しており、植物表現型取得法が中心である。

Code and data availability

The paper's UAV wheat lodging images, annotated datasets, and Mobile U-Net model are self-built and not publicly deposited; the data availability statement restricts access to permission from the authors' institute. The only public URL mentioned (Labelme) is a generic annotation tool, not a paper-specific asset.

No evidence-backed public reproduction asset is currently recorded.