基于Sentinel-2数据的干旱区典型绿洲植被叶绿素含量估算

以渭干河-库车河绿洲(渭-库绿洲)为研究区,采用在机器学习方面具有明显优势的随机森林回归算法,对绿洲内的4种典型植被(棉花、芦苇、杨树、大枣)叶片的叶绿素相对含量(soil and plant analyzer development,SPAD)考。进行估算和验证。首先基于"红边"处光谱信息丰富的哨兵2号(Sentinel-2)影像和由其衍生的一阶微分、二阶微分影像各提取23种对叶绿素敏感的宽波段光谱指数,加入3种影响植物生长的土壤参量(土壤含水量,土壤有机质,土壤电导率)作为影响叶片SPAD的特征变量,再根据以上特征变量对每种植被叶片各建立3种方案的SPAD估算模型,从而实现对绿洲内植被叶绿素的监测。结果表明:①影像经一阶微分再提取的植被指数相比原位光谱植被指数,在SPAD估测模型中起到了更重要的作用,在随机森林算法的重要性排序中位居前列;②4种植被叶片的SPAD估测模型都取得了不错的效果,芦苇叶片尤为显著,确定系数(R^2)达到了0.926;③分析对比3种方案下模型预测能力,方案3(包含土壤参量)的预测能力卓越〔2.143<相对百分比偏差(RPD)<2.692〕,其预测能力排序为:方案3>方案1>方案2,土壤属性和模型预测结果有较强的非线性相关。Sentinel-2数据具有理想的估算绿洲植被叶绿素含量的潜力,提供了一种高效、低成本、潜在高精度的方案来估算叶绿素含量,可为干旱区绿洲农业、生态系统实现更有效的保护和管理提供参考。

Estimation of Chlorophyll Content of Typical Oasis Vegetation in Arid Area Based on Sentinel-2 Data

The Ogan-Kuqa River Delta Oasis,a typical oasis in the arid zone in China,was taken as the study area. The method of Random Forest with a comparative advantage in machine learning was chosen to model and estimate the relative contents of chlorophyll (SPAD values) of leaves from four kinds of representative vegetation (cotton,reed,poplar and jujube). The 23 broadband spectral indices of vegetation,which are sensitive to chlorophyll content,were obtained based on the reflectance of original Sentinel-2 image with rich spectral information in the"red edge"bands. These vegetation indices were extracted again in the original band order on the firstly-derived Sentinel-2 image and secondly-derived Sentinel-2 image. Three soil parameters (soil moisture content,SMC;soil organic matter,SOM;electrical conductivity,EC) related to vegetation growth were all conducted as the characteristic variables affecting SPAD values. According to the characteristic variables above,three modelling schemes could be developed to monitor the SPAD values of vegetation leaves in oasis. The results showed that:① Vegetation indices obtained from the firstly-derived image played a more important role than the original vegetation indices in the SPAD estimation model.② It could be concluded that SPAD-RF regression model,based on the Sentinel-2 satellite image data,could be used to effectively monitor the SPAD values of leaves of the four vegetation types. Especially for the estimation model of SPAD of reed leaves,R^2 reached 0. 926.③ By analyzing and comparing the model prediction capability under the three schemes,the prediction capability of scheme 3 (including soil parameters) was excellent (2. 143 < relative percentage deviation (RPD)< 2. 692),and the prediction capability was ranked as scheme 3 > scheme 1 > scheme 2. There was a significant nonlinear correlation between the soil properties and the model prediction results. Holistically,Sentinel-2 data has great potential for predicting chlorophyll content of oasis vegetation. This study provided an efficient,low-cost,potentially high-precision solution to estimate SPAD.