融合无人机光谱信息与纹理特征的大豆土壤含水率估测模型研究

Estimation Model of Soybean Soil Moisture Content Based on UAV Spectral Information and Texture Features

摘要:

及时获取大田作物根区土壤含水率（Soil moisture content，SMC）对于实现精准灌溉至关重要。本研究采用无人机多光谱技术，通过连续2年（2021—2022年）田间试验，采集了大豆开花期不同土壤深度的SMC数据以及相应的无人机多光谱图像，建立了与作物参数具有较强相关性的植被指数及冠层纹理特征。通过分析植被指数和纹理特征与各深度土层SMC的相关性，分别筛选出与各深度土层SMC相关系数达显著相关（P<0.05）的参数作为模型的输入变量（组合1：植被指数；组合2：纹理特征；组合3：植被指数结合纹理特征），分别利用支持向量机（Support vector machine，SVM）、梯度提升模型（Extreme gradient boosting，XGBoost）和梯度提升决策树（Gradient boosting decision tree，GDBT）对各深度土层SMC进行建模。结果表明，与20~40 cm和40~60 cm土层深度相比，植被指数和纹理特征在0~20 cm土层深度中与SMC表现出更高的相关性。XGBoost模型为SMC估算的最佳建模方法，特别是对于0~20 cm土层深度。该深度估计模型验证集决定系数为0.881，均方根误差为0.7%，平均相对误差为3.758%。本研究结果为大豆根区SMC无人机多光谱监测提供了基础，为水分胁迫条件下作物生长的快速评估提供了参考。

Abstract:

Timely acquisition of soil moisture content (SMC) in the root zone of field crops is crucial for achieving precision irrigation. Drone-based multispectral technology and conducted field experiments over two consecutive years (2021—2022) were used to collect SMC data at different soil depths during the soybean flowering stage, as well as corresponding multispectral images from the drone. Vegetation indices and canopy texture features, which are highly correlated with crop parameters, were established. By analyzing the correlation between vegetation indices, texture features, and SMC at various soil depths, parameters with significant correlation coefficients (P<0.05) were selected as input variables for the model (Combination 1: vegetation indices;Combination 2: texture features;Combination 3: vegetation indices combined with texture features). Support vector machine (SVM), extreme gradient boosting (XGBoost), and gradient boosting decision tree (GDBT) models were used to model SMC at different soil depths. The results indicated that compared with soil depths of 20~40 cm and 40~60 cm, vegetation indices and texture features exhibited higher correlations with SMC at the 0~20 cm soil depth. The XGBoost model was found to be the best modeling method for SMC estimation, particularly for the 0~20 cm soil depth. For this depth, the validation set of the estimation model had a determination coefficient of 0.881, a root mean square error of 0.7%, and a mean relative error of 3.758%. The research result can provide a foundation for drone-based multispectral monitoring of SMC in the soybean root zone and offer a reference for rapid assessment of crop growth under water stress conditions.