Monitoring Models Of Physiological Parameters Of Corn And Farmland Soil Information Based On Hyper-spectral Reflectance

Precision agriculture, an important approach to agricultural sustainable development, is an industry combined high and new technology with agricultural production. The accurate and real-time acquisition for crop growth and ecological environment information is the prerequisite and basis for the implementation of precision agriculture, and is one of the key technologies in the development of modern agriculture. Hyper-spectral remote sensing, characterized as multi-band and high-resolution, can used to detect slight differences of ground objects. These advantages make hyper-spectral remote sensing as new technology for quantifying surface information for precision agriculture. Therefore, the establishment of monitoring model on physicochemical parameters of crop and farmland soil information based on hyper-spectral remote sensing technology can strengthen the monitoring ability, and improve the monitoring precision and accuracy.In this thesis, to establish suitable models to monitor physiological parameters of corn and farmland soil information, hyper-spectral data were collected in field at different grown stages of corn. Based on analyzing these hyper-spectral data, the most sensitive wavebands and characteristic parameters to physiological parameters and farmland soil information were obtained. Thereafter, the hyper-spectral remote sensing monitoring models based on the sensitive wavebands and characteristic parameters were constructed by using simple statistical regression(SSR), partial least square regression(PLSR) and artificial neural network(ANN). The results of this study will provide theoretical basis and technical support for dynamic monitoring the growth of corn and scientific fertilization management in field. The main results are as follows:(1) The absorption peak at 550 nm increased as anthocyanin content increase in leaves, and anthocyanin concentration is most correlated to SVC(Spectra Vista Corporation) reflectance spectra at 548 nm whereas it is most correlated to SOC(Surface Optics Corporation) reflectance spectra at 540.73 nm. The normalized index and ratio index among the two-band index constructed by SVC and SOC spectra had high correlation with leaf anthocyanin content. The ANN model, based on SVC spectral feature index, could be used to monitor anthocyanin concentration in corn leaves with high accuracy and robustness, because the determination coefficient(R2) were 0.776 and 0.759 for training and validation, respectively, and the Root Mean Square Error(RMSE) and Residual Prediction Deviation(RPD) were 0.111 and 2.041, respectively for validation. Another ANN model, established based on SOC spectral characteristic index, was the optimum model to monitor anthocyanin content in corn leaves, because the R2 were 0.875 and 0.851, respectively for training and validation, and the RMSE and RPD were 0.087 and 2.604, respectively for validation. The results suggested that 1) models established based on SOC spectral parameters were more accurate than those using SVC spectral parameters; 2) models established based on characteristic indices formed much better than the models using vegetation indices; 3) models using characteristic index combined with ANN was the best to monitor anthocyanin concentration for corn leaf.(2) The sensitive bands to SPAD value of corn leaf were different among different growth stages. The vegetation indices, i.e., difference index 2(D2), green NDVI(GNDVI), improved soil adjusted vegetation index(MSAVI), normalised difference vegetation index(NDVI), optimised soil-adjusted vegetation index(OSAVI), modified second soil-adjusted vegetation index(OSAVI2), the ratio of modified second chlorophyll absorption ratio index and modified second soil-adjusted vegetation index(TCARI2/OSAVI2), modified second chlorophyll absorption ratio index(TCARI2), transformed chlorophyll absorption ratio index(TCARI) and the hyperspectral characteristic parameter, i.e., the ratio of the sum of first order differential in red edge and the sum of first order differential in blue edge(SDr/SDb), green peak skewness(Sg), and red valley(Ro) are significantly related to SPAD values at the four growth stages. The best models to monitor SPAD values were ANN models based on hyperspectral characteristic parameters(for 6-8 leaf stage and 10-12 leaf stage), vegetation index(for flowering-silking stage), reflectance spectra(for milky stage and Filling stage stage), because the corresponding R2 were 0.845, 0.880, 0.806, 0.763, and 0.785, respectively in model training, whereas it was 0.820, 0.919, 0.822, 0.814, and 0.760, respectively in validation and the RMSEs(RPD) were 0.677(2.358), 0.454(3.455), 0.846(2.374), 0.818(2.319), and 0.774(2.078), respectively. All models for 10-12 leaf stage can efficiently monitor corn leaf SPAD values.(3) The correlation between biomass and corn reflectance spectra was different at each growth stages. The vegetation indices including greenness index(GI), GNDVI, MSAVI, MERIS terrestrial chlorophyll index(MTCI), NDVI, NDVI3, OSAVI, simple ratio index(SR), OSAVI2, TCARI2, TCARI2/OSAVI2, improved modified chlorophyll absorption ratio index(MCARI2), new double difference index(DDn), spectral polygon vegetation index(SPVI), triangular vegetation index(TVI), and red edge triangular vegetation index(RTVI) were significantly correlated with corn biomass at two growth stages. The hyperspectral characteristic parameters, including green peek reflectance(Rg), the sum of green reflectance(SRg), and the sum of first order differential in blue peek(SDg) were significantly correlated with corn biomass at three growth stages. These parameters had good generality when used to estimate the biomass. The results suggest that these indexes can be used to estimate the biomass. The ANN models based on reflectance spectra(for 6-8 leaf stage), vegetation index(for 10-12 leaf stage), and first derivative spectra(for milky stage and Filling stage stage) were optimal to monitor biomass, because the corresponding R2 were 0.908, 0.938, and 0.800, respectively in model training, whereas it was 0.918, 0.939, and 0.762, respectively in validation and the RMSEs(RPD) were 0.086 kg·m-2(3.507), 0.123 kg·m-2(4.051), and 0.400 kg·m-2(2.051), respectively. The prediction accuracy of monitoring models at 6-8 leaf stage and 10-12 leaf stage were higher than those at flowering-silking stage, but the model at milky stage could not efficient monitor the biomass.(4) The band depth of canopy spectral reflectance at 850-1790 nm and 1960-2400 nm are increasing with the increase of corn plant moisture content. The correlations between plant moisture content and corn reflectance spectra were different among different growth stages. The published index(FD730-1330) and the new hyperspectral indices(FDD(725,925), FDD(725, 1140), FDD(725, 1330)) had good correlation with corn plant moisture content, and had good generality. The ANN monitoring models based on the first derivative spectra have good validation accuracy. So they were the optimal models used to monitor corn plant moisture content at 6-8 leaf stage, 10-12 leaf stage, and flowering-silking stage. R2 of training models were 0.858, 0.877, and 0.804, and RMSE were 0.359%, 0.479%, 0.819%, and RPD were 2.654, 2.850, 2.261, respectively. The prediction accuracy of monitoring models at Filling stage and milky stage was not ideal, so the monitoring models needed to be improved.(5) Spectral reflectance decreased as soil moisture increase; meanwhile, the absorption located at near 1400 and 1900 nm shifted to longer waveband. The maximum correlation bands located at 570, 1430, and 1950 nm. The spectral absorption characteristic parameters of the most correlate coefficients with soil moisture were maximum absorption depth(D) and absorption area(A), absorption peak right area(RA), and absorption peak left area(LA). The optimal prediction models of soil moisture content were the linear models using C1950, D1900 and RA1900 as the independent variable and logarithm models using A1900 and A1400 as the independent variable. The R2 of fitting and validation were range from 0.927 to 0.943, 0.936 to 0.96, respectively, and RMSE range from 1.299% to 1.773%, and RPD range from 3.538 to 4.885.(6) Soil spectra showed different with different total nitrogen(TN). The differences of soil spectral became small when available nitrogen(AN) content increased to a certain value. The difference index could best reflect the correlation between soil characteristic indices and TN. The monitoring models constructed by PLSR and ANN method had good estimation effect. The ANN monitoring model based on the first derivative spectra had good training and validation accuracy, was the optimal model to monitor TN. The training and validation R2 were 0.886 and 0.880, RMSE were 0.0077% and 0.0086%, RPD were 2.971 and 2.846, respectively. The ANN model based on CB+CS+CI was the optimal model for estimating the AN, training R2 was 0.757; validation R2 was 0.758, RMSE and RPD were 2.1262 mg·kg-1 and 2.033, respectively.(7) The spectral reflectance decreased with the increase of soil phosphorus content. Soil spectral differences became small when soil phosphorus content increased to a certain value. The ANN models based on the first derivative spectra, CB+CS, CB+CS+CI could accurately estimate available phosphorus(AP). The best effective estimation model was the ANN model based on CB + CS + CI, R2 of training and validation were 0.806 and 0.811, RMSE was 2.691 mg·kg-1, RPD was 2.216. The models established using PLSR and ANN method could not be used to monitor total phosphorus(TP).(8) The soil spectral were obvious affected by high total potassium(TK) content, but small affected by available potassium(AK) content and not regular. The monitoring models constructed by PLSR and ANN method had high precision and could predict TK accurately. The ANN model established based on the first derivative of band depth was the best monitoring model for estimating TK. R2 based on training and validation of ANN were 0.967 and 0.971, RMSE and RPD of validation were 0.033%, 0.030% and 5.416, 5.957, respectively. The ANN model using normalized differential spectrum as independent variable was the best prediction model for AK. R2 based on training and validation of ANN model were bigger than 0.83, and RMSE of validation was 14.457 mg·kg-1, and RPD was 2.591. Prediction accuracy of TK using visible and near-infrared spectral technique was higher than that of AK and differential transform could improve the prediction precision of the model.