不同粒径土壤有效磷光谱估测  

Estimation of available phosphorus content in soil with different particle sizes based on hyperspectral remote sensing

作　　者：梁智永 陈署晃 李娜 吕彩霞 赖宁 信会男 李永福 耿庆龙 LIANG Zhi-yong;CHEN Shu-huang;LI Na;LV Cai-xia;LAI Ning;XIN Hui-nan;LI Yong-fu;GENG Qing-long(Institute of Soil Fertilizer and Agricultural Water Saving,Xinjiang Academy of Agricultural Sciences,Urumqi Xinjiang 830054;College of Resources and Environment,Xinjiang Agricultural University,Urumqi Xinjiang 830052)

机构地区：[1]新疆农业科学院土壤肥料与农业节水研究所,新疆乌鲁木齐830054 [2]新疆农业大学资源与环境学院,新疆乌鲁木齐830052

出　　处：《中国土壤与肥料》2024年第12期230-239,共10页Soil and Fertilizer Sciences in China

基　　金：农业科技创新稳定支持专项(xjnkywdzc-2023002,xjnkywdzc-2023002-1,xjnkywdzc-2023007-3);新疆维吾尔自治区重大专项(2022A02011-2);国家农业重大科技项目(NK2022180804)。

摘　　要：采集新疆乌鲁木齐市110个风干的土壤样品,将筛选出的土壤样品分别过2.00、0.50、0.25 mm的筛,在室内进行反射率光谱数据的采集,对采集的光谱数据(R)与土壤养分含量数据剔除异常值,随后对光谱数据进行断点校正以及10种光谱反射率变换,筛选出与有效磷含量相关性最高的光谱变换为对数的一阶微分(lgR)′变换。在此基础上利用竞争性自适应加权算法(CARS)对经过(lgR)′处理后的数据进行特征波段的筛选,用偏最小二乘回归(PLSR)、BP神经网络、随机森林建立土壤有效磷含量预测的高光谱分析模型。模型评价指标采用决定系数、均方根误差、相对分析误差、平均绝对误差。结果显示:土壤原始光谱特征在各个波段与有效磷含量相关性都较差,不同形式的光谱数据变换均能够提高光谱反射率与有效磷含量的相关性;对比不同粒径处理的模型预测精度,过筛粒径越小对有效磷含量的估测精度越高,3种方法的最优拟合模型都是过0.25 mm筛的处理;对过2.00、0.50、0.25 mm筛的土壤样品有效磷含量分别采用3种方法建模,均以PLSR建立的模型预测精度最为突出,其模型决定系数、均方根误差、相对分析误差、平均绝对误差分别为0.76、10.96、2.07、8.69;0.77、12.52、2.00、9.74;0.77、10.90、2.13、8.81。因此,对过0.25 mm筛处理的土壤有效磷,利用CARS算法筛选出光谱特征波段结合PLSR,能够较好的估测乌鲁木齐市土壤中有效磷的含量,为室内高光谱估测土壤养分含量提供理论基础。110 air-dried soil samples were collected from Urumqi,Xinjiang.The selected soil samples were sieved at 2.00,0.50 and 0.25 mm,respectively.The reflectance spectral data were collected indoors,and the abnormal values were removed from the collected spectral data(R)and soil nutrient content data.Then the breakpoint correction and ten spectral reflectance transformations were carried out on the spectral data,and the log-first derivative(lgR)'with the highest correlation with the available phosphorus content was selected and transformed.On this basis,the competitive adaptive weighting algorithm(CARS)was used to screen the characteristic bands of the processed(lgR)′data,and the hyperspectral analysis model of soil available phosphorus content prediction was established by partial least square regression(PLSR),BP neural network and random forest.Coefficient of determination,root mean square error,relative analysis error and mean absolute error were used as evaluation indexes of the model.The results showed that the correlation between the original spectral characteristics and available phosphorus content was poor in each band,and the correlation between spectral reflectance and available phosphorus content could be improved by different spectral data transformation.Comparing the prediction accuracy of the models with different particle size treatments,the smaller the size of the sieve,the higher the estimation accuracy of the available phosphorus content.The optimal fitting models of the three methods were all over 0.25 mm sieve treatment.For the available phosphorus of 2.00,0.50 and 0.25 mm treatments,PLSR was used to establish the model with the most prominent prediction accuracy.The model determination coefficient,root mean square error,relative analysis error and average absolute error of the three methods were 0.76,10.96,2.07,8.69;0.77,12.52,2.00,9.74;0.77,10.90,2.13,8.81,respectively.Therefore,the characteristic bands selected by CARS algorithm and PLSR could better estimate the content of available phosphorus

关 键 词：土壤粒径 高光谱 估测模型 有效磷 CARS算法 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]