| 摘 要: 为实现拔节期玉米土壤水分的有效估测,基于无人机(UAV)多光谱数据构建了三分支深度学习回归模型LN_Net。通过灰度板校正、图像裁剪与配准等预处理手段,对可见光与多光谱波段进行融合,构建了具有8个通道的复合图像数据集。并借鉴 RepViT模型轻量化的思路,设计了多分支特征提取与融合结构,在可见光和红边、近红外波段独立提取光谱信息,并通过多层感知机(MLP)进行特征整合。多分支结构模型性能较单分支结构模型有所提升,尤其在红边与近红外波段增强时,模型性能最优。实验结果表明,所提模型在验证集上可达 RMSE=0.020、NRMSE=0.111、R2=0.641,在无需显式提取光谱指数的情况下,能较准确地预测土壤含水量。该方法为拔节期玉米的水分管理与精准灌溉提供了一条可行途径。 |
| 关键词: 拔节期玉米 无人机多光谱数据 深度学习回归 RepVit 土壤水分预测 |
|
中图分类号: TP391.7
文献标识码: A
|
|
| Soil Moisture Prediction for Maize at the Jointing Stage Based on UAV Imagery and Deep Learning |
|
LIU Jiayu, NIE Zhigang
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
liujy@st.gsau.edu.cn; niezg@gsau.edu.cn
|
| Abstract: To achieve effective estimation of soil moisture in maize during the jointing stage, this paper constructs a three-branch deep learning regression model named LN_Net based on UAV multispectral data. Through preprocessing methods such as grayscale panel correction, image cropping, and registration, visible and multispectral bands were fused to construct a dataset of composite images with 8 channels. Drawing on the lightweight design concept of the RepViT model, a mult-i branch feature extraction and fusion structure was designed. This structure independently extracts spectral information from the visible, red-edge, and nea-r infrared bands and integrates the features using a MultiLayer Perceptron(MLP).The performance of the mult-i branch model improved compared to the single-branch model, with optimal performance observed particularly when the red-edge and nea-r infrared bands were enhanced. The experimental results indicate that the proposed model can achieve RMSE=0.020, NRMSE=0.111, and R2=0.641 on the validation set, enabling relatively accurate prediction of soil water content without the explicit extraction of spectral indices. This method provides a feasible approach for water management and precision irrigation of maize during the jointing stage. |
| Keywords: maize at jointing stage UAV multispectral data deep learning regression RepViT soil moisture prediction |
