【摘要】：精準農(nóng)業(yè)是一種高新技術(shù)與農(nóng)業(yè)生產(chǎn)相結(jié)合的產(chǎn)業(yè),是可持續(xù)農(nóng)業(yè)發(fā)展的重要途徑。高精度、及時的獲取農(nóng)作物長勢和生態(tài)環(huán)境信息是精準農(nóng)業(yè)實施的前提和基礎,也是現(xiàn)代農(nóng)業(yè)發(fā)展的關(guān)鍵技術(shù)之一。高光譜遙感具有波段多、間隔窄的特點,能構(gòu)成獨特的多維光譜空間,直接捕獲地物的微弱光譜差異信息。高光譜為遙感信息的定量應用開辟了新的領(lǐng)域,也為農(nóng)田信息獲取帶來了巨大的前景,并逐漸成為新興精準農(nóng)業(yè)最重要的技術(shù)手段之一。因此,應用高光譜遙感建立生理參數(shù)及農(nóng)田土壤信息監(jiān)測模型,可增強對作物生理參數(shù)及農(nóng)田土壤信息的監(jiān)測能力,提高作物生長及農(nóng)田土壤信息監(jiān)測的精度和準確性。本研究以田間試驗的玉米為研究對象,結(jié)合地面高光譜遙感技術(shù)與生理參數(shù)分析技術(shù),系統(tǒng)分析玉米不同生理參數(shù)及農(nóng)田土壤信息的高光譜特征,在相關(guān)分析的基礎上,提取特征波段、植被指數(shù)、高光譜特征參數(shù)、特征指數(shù),應用簡單統(tǒng)計回歸(SSR)、偏最小二乘回歸(PLSR)和人工神經(jīng)網(wǎng)絡(ANN)方法,建立玉米生理參數(shù)及農(nóng)田土壤信息的高光譜監(jiān)測模型,為動態(tài)監(jiān)測玉米生長狀況及科學的田間施肥管理提供理論依據(jù)和技術(shù)支持。主要研究結(jié)果如下:(1)隨著玉米葉片花青素含量增加,550 nm處吸收峰增大,SVC和SOC光譜與玉米葉片花青素含量最大相關(guān)波段分別為548 nm和540.73 nm。以SVC和SOC光譜構(gòu)建的兩波段歸一化指數(shù)和比值指數(shù)與葉片花青素含量相關(guān)性最高;基于SVC光譜特征指數(shù)建立的ANN模型,訓練和驗證R2分別為0.776和0.759,驗證RMSE為0.111,RPD值為2.041,預測精度較高、模型比較穩(wěn)定,能有效監(jiān)測玉米葉片花青素含量;基于SOC光譜特征指數(shù)建立的ANN模型是玉米葉片花青素含量監(jiān)測的最優(yōu)模型,訓練和驗證R2分別為0.875和0.851,驗證RMSE為0.087,RPD值為2.604。SOC光譜參數(shù)建立的模型擬合及驗證精度整體高于SVC光譜參數(shù)建立的模型,特征指數(shù)建立的模型優(yōu)于植被指數(shù)建立的模型;特征指數(shù)結(jié)合ANN方法是建立玉米葉片花青素含量監(jiān)測模型的最優(yōu)方法。(2)玉米不同生育期SPAD值的敏感波段有差異。植被指數(shù)D2、GNDVI、MSAVI、NDVI、OSAVI、OSAVI2、TCARI2/OSAVI2、TCARI2、TCARI和高光譜特征參數(shù)SDr/SDb、Sg、Ro均與玉米4個生育期葉片的SPAD值極顯著相關(guān),通用性較好。基于6-8葉期、10-12葉期的高光譜特征參數(shù),開花吐絲期的植被指數(shù),灌漿期、乳熟期的原始光譜建立的ANN模型訓練及驗證精度均較高,模型較穩(wěn)定,是各個生育期玉米葉片SPAD值監(jiān)測的最優(yōu)模型。訓練R2分別為0.845、0.880、0.806、0.763、0.785,經(jīng)獨立樣本驗證,R2分別為0.820、0.919、0.822、0.814、0.760,RMSE分別為0.677、0.454、0.746、0.818、0.774,RPD值分別為2.358、3.455、2.374、2.319、2.078。10-12葉期,以3種方法建立的模型均能對玉米葉片SPAD值進行有效監(jiān)測。(3)不同生育期玉米光譜與生物量的相關(guān)性差異較大,植被指數(shù)GI、GNDVI、MSAVI、MTCI、NDVI、NDVI3、OSAVI、SR、OSAVI2、TCARI2、TCARI2/OSAVI2、MCARI2、DDn、SPVI、TVI、RTVI均在2個生育期與玉米生物量極顯著相關(guān);高光譜特征參數(shù)Rg、SRg、SDg在3個生育期與玉米生物量極顯著相關(guān),通用性較好。6-8葉期以原始光譜、10-12葉期以植被指數(shù)、開花吐絲期以一階微分光譜建立的ANN模型,訓練R2分別為0.908、0.938、0.800,驗證R2分別為0.918、0.939、0.762,RMSE分別為0.086 kg·m-2、0.123 kg·m-2、0.400 kg·m-2,RPD值分別為3.507、4.051、2.051,訓練和驗證結(jié)果均較好,是監(jiān)測各生育期玉米生物量的最優(yōu)模型。6-8葉期和10-12葉期的監(jiān)測模型精度高于開花吐絲期;乳熟期建立的模型不能進行生物量有效監(jiān)測。(4)850-1790 nm和1960-2400 nm范圍,隨著玉米植株含水量增加波段深度增大,不同生育期玉米植株含水量與光譜的相關(guān)性差異較大;FD730-1330和新建光譜指數(shù)FDD(725,925)、FDD(725,1140)、FDD(725,1330)與玉米不同生育期植株含水量相關(guān)性較好,通用性較強。6-8葉期、10-12葉期、開花吐絲期,基于一階微分光譜建立的ANN模型,經(jīng)獨立樣本驗證,預測值與實測值之間的R2分別為0.858、0.877、0.804,RMSE為0.359%、0.479%、0.819%,RPD值為2.654、2.850、2.261,模型的預測精度較高,穩(wěn)定性較好,是進行各生育期玉米植株含水量監(jiān)測的最優(yōu)模型。灌漿期和乳熟期建立的模型預測效果不理想,有待進一步研究。(5)隨土壤含水量增加光譜反射率下降,1400、1900 nm附近的水分吸收谷朝長波方向偏移。與土壤含水量相關(guān)性最大的光譜位于570、1430、1950 nm,相關(guān)性最大的吸收特征參數(shù)是最大吸收深度(D)、吸收總面積(A)、吸收峰右面積(RA)、吸收峰左面積(LA)。基于C1950、D1900、RA1900建立的一元線性模型和A1900、A1400建立的對數(shù)模型是預測土壤含水量的最優(yōu)模型,擬合R2位于0.927-0.943之間,驗證R2位于0.936-0.96之間,RMSE位于1.299-1.773%之間,RPD值位于3.538-4.885之間。(6)不同全氮含量的土壤光譜差異較大;堿解氮含量增大到一定值時,反射率之間的差異變小;與土壤氮含量相關(guān)性最好的兩波段光譜指數(shù)是差值指數(shù)。以PLSR和ANN方法建立的全氮含量監(jiān)測模型預測效果較好。其中,基于一階微分光譜建立的ANN模型,訓練和驗證R2分別為0.886和0.880,RMSE為0.0077%和0.0086%,RPD值為2.971和2.846,訓練和驗證結(jié)果較好,模型最穩(wěn)定,是監(jiān)測土壤全氮含量的最優(yōu)模型。基于CB+CS+CI建立的ANN模型,訓練R2為0.757,驗證R2為0.758,驗證RMSE為2.1262 mg·kg-1,RPD值為2.033,是監(jiān)測土壤堿解氮含量的最優(yōu)模型。(7)光譜反射率隨土壤磷含量增加而減小,當土壤磷含量增大到一定值時,土壤光譜反射率之間的差異變小�；跉w一化微分、CB+CS和CB+CS+CI建立的ANN模型,可以對土壤有效磷含量進行準確預測,其中,CB+CS+CI建立的ANN模型預測效果最好,訓練和驗證R2分別為0.806和0.811,驗證RMSE為2.691 mg·kg-1,RPD值為2.216;PLSR和ANN方法建立的模型精度較低,不能進行土壤全磷含量的有效監(jiān)測。(8)土壤全鉀含量較高時,對土壤光譜反射率影響較大;土壤速效鉀含量對土壤光譜影響較小,變化規(guī)律不明顯。以PLSR和ANN方法建立的模型精度均較高,能對土壤全鉀含量進行準確預測。其中,基于波段深度微分建立的ANN模型是監(jiān)測土壤全鉀含量的最優(yōu)模型,訓練和驗證R2分別為0.967和0.971,驗證RMSE分別為0.033%和0.030%,RPD值分別為5.416和5.957�；跉w一化微分光譜建立的ANN模型的訓練和驗證R2大于0.83,驗證RMSE為14.457 mg·kg-1,RPD值為2.591,是土壤速效鉀含量預測的最優(yōu)模型。土壤全鉀含量的預測精度高于速效鉀含量,微分變換可以提高模型的預測精度。
[Abstract]:Precision agriculture is an industry combining high and new technology with agricultural production. It is an important way for sustainable agricultural development. High precision, timely acquisition of crop growth and ecological environment information is the premise and basis of precision agriculture implementation, and is also one of the key technologies of modern agricultural development. Hyperspectral remote sensing has many bands and narrow spacing. It can form a unique multi-dimensional spectrum space and directly capture the weak spectral difference information of the ground objects. Hyperspectral has opened up a new field for the quantitative application of remote sensing information. It also brings great prospects for the acquisition of farmland information, and has gradually become one of the most important technical means of new precision agriculture. Therefore, the application of hyperspectral remote sensing is established. The physiological parameters and farmland soil information monitoring model can enhance the monitoring ability of crop physiological parameters and farmland soil information, improve the accuracy and accuracy of crop growth and farmland soil information monitoring. Based on the analysis of different physiological parameters of maize and the hyperspectral characteristics of farmland soil information, based on the correlation analysis, the characteristic bands, vegetation index, hyperspectral characteristic parameters, characteristic index, simple statistical regression (SSR), partial least squares regression (PLSR) and artificial neural network (ANN) were used to establish maize physiological parameters and farmland soil information. The hyperspectral monitoring model provides theoretical basis and technical support for dynamic monitoring of maize growth and scientific field fertilization management. The main results are as follows: (1) with the increase of anthocyanin content in maize leaves, the absorption peaks at 550 nm are increased, and the bands of the SVC and SOC spectra and the maximum content of leaf anthocyanins are 548 nm and 540, respectively. The two band normalization index and ratio index of.73 nm. based on SVC and SOC spectra have the highest correlation with the content of leaf anthocyanins. The ANN model based on the SVC spectrum characteristic index is 0.776 and 0.759, respectively. The RMSE is 0.111, the RPD value is 2.041, the prediction accuracy is higher, the model is more stable, and the maize leaf can be effectively monitored. Anthocyanin content; ANN model based on SOC spectral characteristic index is the best model for monitoring anthocyanin content in maize leaves, training and verification R2 are 0.875 and 0.851, RMSE is 0.087, RPD value is set up by 2.604.SOC spectral parameters and the accuracy is higher than the model established by SVC spectral parameters, and the characteristic index is established. The model is superior to the model established by the vegetation index. The characteristic index combined with ANN method is the best method to establish the maize leaf anthocyanin content monitoring model. (2) the sensitive bands of SPAD values in different growth stages of maize are different. The vegetation index D2, GNDVI, MSAVI, NDVI, OSAVI, OSAVI2, TCARI2/OSAVI2, TCARI2, TCARI, and hyperspectral characteristic parameters SDr/SDb The SPAD value of the leaves of 4 growing periods of maize was very significant, and the generality was better. Based on the 6-8 leaf stage, the hyperspectral characteristic parameters of the 10-12 leaf period, the vegetation index of the flowering stage, the grain filling period and the original spectrum of the milk ripening period were higher, the model type was more stable, and the SPAD value monitoring of the leaves of each growth period was the monitoring of the SPAD value of the leaves of the maize. The optimal model. The training R2 is 0.845,0.880,0.806,0.763,0.785 respectively. The independent samples are verified that R2 is 0.820,0.919,0.822,0.814,0.760, RMSE is 0.677,0.454,0.746,0.818,0.774 and RPD value is 2.358,3.455,2.374,2.319,2.078.10-12 leaf period respectively. The model established by 3 methods can effectively monitor the SPAD value of maize leaves. (3) the correlation between the spectral and biomass of Maize at different growth stages was quite different, the vegetation index GI, GNDVI, MSAVI, MTCI, NDVI, NDVI3, OSAVI, SR, OSAVI2, TCARI2, TCARI2/OSAVI2 were all significantly related to the maize biomass at the 2 growth stages, and the hyperspectral characteristic parameters were very significant in 3 growth periods and maize biomass. It is related that the.6-8 leaf phase is better for the.6-8 leaf phase, with the vegetation index and the ANN model of the first order differential spectrum in the flowering stage, and the training R2 is 0.908,0.938,0.800 respectively. The R2 is 0.918,0.939,0.762, and the RMSE is 0.086 kg. M-2,0.123 kg. M-2,0.400 kg. The optimal model of maize biomass in each growth period,.6-8 leaf stage and 10-12 leaf stage, was higher than flowering. (4) the range of 850-1790 nm and 1960-2400 nm, with the increase of the depth of maize plant water content and the increase of different growth bands. The correlation between the water content and the spectrum of the maize plants was great. FD730-1330 and the new spectral index FDD (725925), FDD (7251140), FDD (7251330) were well correlated with the water content of the plants at different growth stages. The general.6-8 leaf stage, the 10-12 leaf stage, the flowering spit period, and the ANN model based on the first order differential spectrum were independent samples. It is proved that the R2 between the predicted value and the measured value is 0.858,0.877,0.804, RMSE is 0.359%, 0.479%, 0.819%, and RPD is 2.654,2.850,2.261. The prediction accuracy of the model is higher and the stability is better. It is the optimal model for monitoring the water content of maize plants at various growth stages. (5) (5) with the increase of soil moisture content, the spectral reflectance decreased and the water absorption Valley in the vicinity of 14001900 nm shifted toward the long wave direction. The maximum correlation of the correlation with soil moisture was located at 57014301950 nm. The maximum absorption characteristic parameters were the maximum absorption depth (D), the total absorption area (A), the absorption peak right area (RA), the absorption peak left Area (LA). The linear model based on C1950, D1900 and RA1900 and the logarithmic model established by A1900 and A1400 are the best models for predicting soil water content. The fitting of R2 is between 0.927-0.943 and R2 is between 0.936-0.96 and RMSE lies between 1.299-1.773% and between 1.299-1.773%. (6) the difference of soil spectral difference of different total nitrogen content The difference of the reflectance is smaller when the content of alkali nitrogen is increased to a certain value. The two band spectral index of the best correlation with the soil nitrogen content is the difference index. The total nitrogen content monitoring model based on PLSR and ANN is better. The ANN model based on the first order differential spectrum is trained and verified that R2 is 0, respectively. .886 and 0.880, RMSE are 0.0077% and 0.0086%, RPD values are 2.971 and 2.846. The results of training and verification are better. The model is the most stable model. The ANN model based on CB+CS+CI, training R2 is 0.757, and R2 is 0.758, RMSE is 2.1262 mg kg-1, RPD is 2.033, which is the best monitoring of soil alkaline nitrogen content. (7) (7) the spectral reflectance decreases with the increase of soil phosphorus content. When the soil phosphorus content increases to a certain value, the difference between soil spectral reflectance is smaller. Based on the normalization and differentiation, the ANN model established by CB+CS and CB+CS+CI can accurately predict the content of soil available phosphorus, of which the ANN model established by CB+CS+CI is the best. The training and verification of R2 were 0.806 and 0.811 respectively, which proved that RMSE was 2.691 mg. Kg-1, RPD value was 2.216; PLSR and ANN methods had low model precision. (8) the soil total potassium content was higher, the soil spectral reflectance was more affected; soil available potassium content had little influence on Soil spectrum. The precision of the model established by PLSR and ANN is high, and the total potassium content of soil can be accurately predicted. Among them, the ANN model based on the differential wave band depth differential is the best model for monitoring the total potassium content of soil. The training and verification of R2 are 0.967 and 0.971 respectively, and the RMSE is 0.033% and 0.030% respectively, and the RPD value is 5.416, respectively. The ANN model based on normalized differential spectrum (5.957.) is trained and verified that R2 is more than 0.83, and RMSE is 14.457 mg. Kg-1, and the RPD value is 2.591. It is the best prediction model of soil available potassium content. The prediction accuracy of soil total potassium content is higher than the content of available potassium. Differential transformation can improve the prediction accuracy of the model.
【學位授予單位】：西北農(nóng)林科技大學
【學位級別】：博士
【學位授予年份】：2016
【分類號】：S513
，

本文編號：2137559