基于多光谱影像的苜蓿地不同生育期土壤含盐量反演模型研究

Inversion Model of Soil Salinity at Different Fertility Stages in Alfalfa Fields Based on Multi-spectral Imagery

摘要:

为探究苜蓿地不同生育期不同深度的土壤含盐量快速反演模型，采集苜蓿地分枝期、现蕾期、初花期深度0~15cm、15~30cm、30~50cm土壤含盐量，基于无人机多光谱影像数据，提取采样点光谱反射率，在此基础上引入红边波段代替红波段与近红外波段计算光谱指数，采用皮尔逊相关系数法（Pearson correlation corfficient,PCCs）、灰色关联度（Gray relational analysis, GRA）分析法进行指数筛选，构建54个基于极端梯度提升（Extreme gradient boosting，XGBoost）算法、反向传播神经网络（Back propagation neural network，BPNN）和随机森林（Random forest，RF）的机器学习模型，确定苜蓿地不同生育期不同深度土层的土壤含盐量最佳反演模型。结果表明：XGBoost模型反演效果整体优于BPNN模型和RF模型，反演结果能真实反映不同生育期苜蓿地的土壤含盐量。从不同生育期反演来看，分枝期和初花期XGBoost模型反演效果优于其他模型，验证集决定系数（R2p）分别为0.835、0.709，均方根误差（RMSE）分别为0.042%、0.047%，平均绝对误差（MAE）分别为0.046%、0.037%；现蕾期RF模型反演效果优于其他模型，R2p为0.717，RMSE为0.034%，MAE为0.042%。从不同深度反演来看，0~15cm土层XGBoost模型反演效果优于其他模型，R2p为0.835，RMSE为0.053%，MAE为0.043%；15~30cm和30~50cm土层XGBoost和RF模型均优于BPNN模型，R2p分别为0.717、0.739，RMSE分别为0.034%、0.038%，MAE分别为0.042%、0.031%。分枝期为最佳反演生育期，0~15cm深度为最佳含盐量反演深度，且PCCs变量筛选方法与XGBoost机器学习算法的耦合模型精度最佳，建模集和验证集的R2分别为0.856、0.835，R2p/R2c为0.975，具有良好的鲁棒性。研究结果可为土壤含盐量的快速精确反演提供理论依据。

Abstract:

Soil salinization has always been an important factor restricting the sustainable development of agriculture in Northwest China. In order to explore the rapid inversion model of soil salinity at different depths in different growth stages of alfalfa land, soil salinity at the depths of 0~15cm,15~30cm and 30~50cm in the branching stage, budding stage and early flowering stage of alfalfa land was collected. Based on the multispectral image data of UAV, the spectral reflectance of sampling points was extracted. On this basis, the red band was introduced instead of the red band and the near-infrared band to calculate the spectral index. Pearson correlation corfficient (PCCs) and gray relational analysis (GRA) were used for index screening. A total of 54 machine learning models based on extreme gradient boosting (XGBoost) algorithm, back propagation neural network (BPNN) and random forest (RF) were constructed to determine the optimal inversion model of soil layers at different depths in different growth stages of alfalfa land. The results showed that the inversion effect of XGBoost model was better than that of BPNN model and RF model, and the inversion results could truly reflect the soil salt content of alfalfa field at different growth stages. According to the inversion of different growth stages, the inversion effect of XGBoost model in branching stage and early flowering stage was better than that of other models. The determination coefficient of validation set （R2p） was 0.835 and 0.709, respectively, the root mean square error （RMSE） was 0.042% and 0.047%, respectively, and the mean absolute error （MAE） was 0.046% and 0.037%, respectively. The inversion effect of RF model was better than that of other models, with R2p of 0.717, RMSE of 0.034% and MAE of 0.042%. From the perspective of different depths inversion, the inversion effect of XGBoost model in 0~15cm soil layer was better than that of other models. The R2p was 0.835, the RMSE was 0.053%, and MAE was 0.043%. The XGBoost and RF models were superior to the BPNN model in 15~30cm and 30~50cm soil layers, with R2p of 0.717 and 0.739, RMSE of 0.034% and 0.038%, and MAE of 0.042% and 0.031%, respectively.The branching period was the best inversion growth period, and the depth of 0~15cm was the best salinity inversion depth, and the coupling model of PCCs variable screening method and XGBoost machine learning algorithm had the best accuracy. The R2 of the modeling set and the verification set were 0.856 and 0.835, respectively, and R2p/R2c was 0.975, which had good robustness. The research results can provide a theoretical basis for rapid and accurate inversion of soil salinity.