基于随机森林模型的苹果叶片磷素含量高光谱估测

Hyperspectral estimation of phosphorus content for apple leaves based on the random forest model

【目的】针对传统化学方法测定苹果叶片磷素含量的不足,使用高光谱技术快速、准确和无损地估测苹果叶片磷素含量。【方法】以烟台栖霞市25个果园100株新梢旺长期苹果树叶片高光谱反射率和叶片磷素（phosphorus,P）含量为数据源,在分析其磷素含量与原始光谱反射率、原始光谱反射率的一阶微分、植被指数和高光谱特征参量相关性的基础上,筛选敏感波长,建立了基于高光谱数据的磷素含量随机森林模型。【结果】新梢旺长期苹果叶片磷素含量在绿光波段（507~590 nm）、红光波段（694~743 nm）和近红外短波波段（1 324~1 364 nm）呈显著负相关;基于植被指数RVI（542,1 094）、RVI（705,937）、DVI（556,712）、DVI（677,1 728）、NDVI（737,549）、DVI（FDR567,FDR1980）和DVI（FDR523,FDR1883）建立的随机森林回归模型有较好的估测效果,决定系数R2=0.923 6,均方根误差RMSE=0.015 8,相对误差RE=6.915%。【结论】光谱植被指数比较适合苹果磷素营养状况估测。

[ Objective] Qixia in Shandong provence, China is a widespread area designated for apple dis- tributing and planting, and it is honored as the township of apples due to its high yield. Apple growth is de- pendent on key nutrients such as N, P, K and so on, therefore, nutrient diagnosis and growth monitoring play a crucial role in precision agriculture. Among the key nutrients, phosphorus （P） is one of the neces- sary nutrient elements to sustain the growth of the crop, it participates directly and indirectly in many dif- ferent ways in the plant metabolism process, photosynthesis and the energy transfer process. The increas- ing use of fertilizer helps to raise the yield of the farm crop, but often generates a lot of fertilizer waste be- cause of the inappropriate amount in use and the absurdity method of applying fertilizer. Much worse, over inputting of chemical fertilization is the main source of water pollution, and this leads to a worsening eco- logical environment. Thus, the accurate diagnosis of crop nutrition is significant to the increase of crop yield and environmental protection. In order to rapidly, accurately and nondestructively evaluate the P- content of apple trees, hyperspectral remote sensing technology has emerged in response to the needs of the times. [Methods]Traditional methods of obtaining P-content of apple leaves, i.e., sulfuric acid hydro- gen peroxide and boiled-molybdenum antimony colorimetric methods, require grinding leaves and apply- ing chemical titrations, which would be harmful to the apple trees and also very time-consuming. Hyper- spectral remote sensing technology has the advantage of high spectral resolution, strong wavelength conti-nuity, and a large amount of spectral information, which is widely used in all kinds of plant diseases and insect pests＇ prevention, chlorophyll content prediction, crop yield forecast and crop nutrient elements monitoring etc. The diagnostic data were from 100 new shoots of the prosperous long-term ＇Red Fuji＇ ap- ple trees of 25 orchards in the Qixia area of Yantai city, which included P-content and hyperspectral re- flectance of the apple tree leaves. The apple leaves original spectral reflectivity data were captured by the ASD FieldSpec4 hyperspectral spectrometer, and the P-content were measured by the traditional chemi- cal ana!ysis in the laboratory. There was a certain degree of differences in the correlation analysis of the P- content with spectral data, in order to increase the accuracy of the P-content estimation for the apple trees, the data was transformed through the first derivative of the original spectral reflectivity, vegetation index and hyperspectral characteristic parameters, then the stepwise regression analysis method was used to select the sensitive wavelengths and spectral parameters and establishe the P-content estimation mod- els. According to the results of the systematic analysis and diagnosis, the random forests model method acts as a non-parametric regression technique, which analyzes the importance of each independent vari- able on the dependent variables and has a good adaptability for complex data. [Results ] Hyperspectral curves of the original reflectance for the apple leaves was distinctly differential. First, the apple leaf reflec- tance is low in the visible region （400-760 nm）, because of the pigment absorption, where there is a green peak （550 nm） that is a strong reflection edge of chlorophyll and a red valley （667 nm） that is a strong ab- sorption of chlorophyll; Second, the range spectral reflectivity has risen sharply in 680-800 nm region, where the strong absorption is steep and approximately close to linear form; Third, the strong reflection wave platform is presented in the near infrared region （800-1 300 nm）, which is related to the strong infra- red reflectance of the cavernous cavity. Finally, the reflectivity curve fluctuates widely in 1 300-2 500 nm, mainly associated with the strong absorption of water and carbon dioxide. By the correlation analysis, the P-content negatively correlated with the original spectral reflectivity （350-2 500 nm range） as a whole, and the correlation coefficient reached to a remarkable level （R=-0.648 5）, especially in the green light re- gion （507-590 nm）, red light region （694-743 nm） and near-infrared spectrum （1 324-1 364 nm）. In order to improve the correlation coefficients, the correlation analysis were carried out with the first derivative of original spectral reflectivity, vegetation index and hyperspectral characteristic parameters respectively. The remarkable wavelengths from the first derivative of the original spectral reflectivity were FDR523, FDR1879, FDR567, FDR1883, FDR701, FDR1980, FDR1876, and FDR2024. The remarkable wave- lengths from the vegetation index were DVI（556, 712）, RVI（705, 937）, DVI（677, 1 728）, RVI（542, 1 094）, DVI（FDR 567, FDR 1 980）, NDVI（9371 549）, and DVI（FDR 523, FDR 1 883）. The remarkable wave- lengths from the hyperspectral characteristic parameters were Db, Dy, Dr, Rr, Rg, SDb, SDy, SDr, SDg, Rg/Rr, （Rg-Rr）/（Rg＋Rr）, SDr/SDb, （SDr-SDb）/（SDr＋SDb）, and （SDr-SDy）/（SDr＋SDy）. The effect of the random forests model was good for estimation of the P-content of the apple trees, including the determina- tion coefficients between the measured value and estimated value which were greater than 0.82, the root mean square errors were below 0.016 4, the relative error were close to 6.915, which demonstrated that the random forests model estimation had a higher accuracy. By this analysis, the random forest model based on remarkable wavelengths from the vegetation index had the best estimation, including the determi- nation coefficient R2=0.923 6, root mean square error RMSE=0.015 8 and relative error RE=6.915 0%. The leaving data is simulated to validate the model and the results compared with the measured value and estimated value is in good agreement. By the verification analysis, the random forest model based on re- markable wavelengths from the vegetation index had the best estimation, including the determination coef- ficient R2=0.868 5, root mean square error RMSE=0.011 8 and relative error RE=5.85%. [Conclusion]Compared with the traditional measurement method, hyperspectral remote sensing technology does in- crease the accuracy and timeliness of the P-content measured value using the random forest model. The study demonstrated the importance of the non-linear random forest model for estimation of the P-content which is negatively correlated with the hyperspectral data. The analysis results will provide a theoretical guidance and technical support for diagnosis of the nutritional status of the orchards.