基于无人机多光谱遥感和机器学习的烟田土壤碱解氮估测

Prediction of Soil Alkaline Nitrogen in Tobacco Fields Based on UnmannedAerial Vehicle Multispectral Remote Sensing and Machine Learning

准确、快速、无损掌握烟田土壤碱解氮分布情况,可为烟草生长过程中合理施用氮肥提供重要依据。以湖北保康县和宣恩县境内的4块典型烟田为研究对象,获取无人机多光谱遥感影像和土壤碱解氮数据,基于传感器“双红边”波段优势,引入红边波段对传统光谱指数进行改进,运用随机森林(random forest,RF)、支持向量机(support vector machine,SVM)、反向传播神经网络(back propagation neural networks,BPNN)和极限梯度提升(extreme gradient boosting,XGBoost)4种机器学习算法,构建烟田土壤碱解氮含量的估测模型。结果显示:(1)以红边和红光两个波段组合为主改进的光谱指数与土壤碱解氮含量的灰色关联度值高于以近红外、红光、绿光和蓝光波段为主的传统光谱指数;(2)基于改进光谱指数和RF算法方法构建的估测模型精度最高,最高建模集和验证集的决定系数、均方根误差分别为0.833、0.784和0.846、0.951;(3)RF算法稳定性验证结果中R^(2)和RMSE分别为0.814和0.906,与单独建模时表现一致,表明RF算法的稳定性较高。该研究为武陵山区烟田土壤碱解氮含量的估测提供了一种新的方法。

During the growth process of tobacco,accurate and rapid acquisition of soil alkaline nitrogen(SAN)distribution in tobacco fields can provide an important basis for the reasonable application of nitrogen fertilizers.Firstly,four tobacco fields which are located in Baokang and Xuanen counties in Hubei Province,China,were selected in this study.Unmanned aerial vehicle(UAV)multispectral remote sensing images and SAN data from sampling points were obtained.Secondly,because the sensor contains two red edge bands,improvements have been made by introducing the red edge band to traditional spectral indices.Thirdly,four machine learning methods,including random forest(RF),support vector machine(SVM),back propagation neural networks(BPNN)and extreme gradient boosting(XGBoost),were used to construct models for estimating SAN content in tobacco fields.The results showed that:(1)The gray correlation shows that improved spectral indices are higher than traditional spectral indices in terms of correlation with SAN.(2)The combination of the improved spectral indices and RF algorithm is the best method to construct the prediction model.The coefficient of determination(R^(2))and root mean square error(RMSE)for the highest modeling and validation sets are 0.833,0.784 and 0.846,0.951,respectively.(3)The RF algorithm was validated on 30%of the overall sample,and the R^(2) and RMSE were 0.814 and 0.906,respectively,indicating that the stability of the RF algorithm was consistent with its performance when modeled alone.Overall,the prognosis is accurate.This study provides a new method for prediction of SAN content in tobacco fields of the Wuling mountain area.