Retrieval Model And Its Optimization For Estimating Soil Components With Hyperspectra In Lakeside Area

Hyper spectral remote sensing is an important direction of quantitative remote sensing, and soil remote sensing is one of the research hotspots of remote sensing applications. At present, the precision agriculture fertilizer and heavy metal pollution monitoring have become respectful to solve the contradictory between people and land, which may further ensure the grain security and the agricultural sustainable development. Rapid estimation of soil components, especially the fast, accurate access to the main soil nutrient and heavy metal content, has become necessary for the development of precision agriculture and soil environmental monitoring. So the application of hyperspectral technology for soil components has become a hot research direction.This study region is located in the Jianghan plain of Hubei province, Honghu city, China, and takes the farmland soil in lakeside area as the research object. There are many questions in estimating soil components with hyperspectra, just like the low precision and complexity of the PLSR calibration model with all spectral bands, the failings of model caused by difference of external factors (water content). In order to solve above problems, this paper used a variety of chemometrics methods to optimizing model. Statistical research process consisted of data acquisition and analysis in the study area, soil components and spectrum characteristics analysis, model calibration and optimization for estimating soil components with hyperspectra, and the distribution characteristics analysis of soil components. The main conclusions are as follows:(1) The statistical characteristics of soil components and spectrum were obtained in our study area.Based on the Mahalanobis distance we eliminated abnormal samples, and then analyzed soil components and spectral characteristics. The results show that:?in this area, the content of soil organic matter (OM) and total nitrogen (TN) is rich, which results in the higher fertility. From the point of view of heavy metals, the content of Fe and Zn is lower, while the content of Cu is on the high side.?Range of each soil component variation is relatively wide which helps to improve the robustness of the retrieval moded.?Soil OM and TN, heavy metals Cu and Zn, Fe and Cu content have a significant positive correlation, which may facilitate the calibration and validation of retrieval model for estimating soil components.?Pretreatments of smoothing, the first derivative, transmission to absorbance and standard normal variate can eliminate the interference of the system noise, baseline and background, and improve the ability of prediction soil components using spectral information.?With the correlation analysis of soil and spectra, the spectral characteristics of OM, TN, Fe, Cu and Zn are located near by660nm,660nm,850nm,400nm and2350nm.(2) Estamating the PLSR calibration models for soil components prediction with differenct pretreatments, and finding the problems and deficiencies in calibration models.The retrieval model was established for each component of dry soil and moist soil using PLSR method. The following conclusions have been made through the analysis of the model predictions:?the content of OM and TN can be effectively estimated in PLSR models using dry soil spectral. There is a good prediction accuracy of the retrieval model for OM, while rough for TN.?After rewetting, the accuracy of the prediction model has been significantly reduced. There is none but high and low value can be distinguished for OM using PLSR retrieval model and it can't even been predicted for TN. Therefore, the influence of external environment factors (such as water) should not be ignored in establishing a model.?The prediction accuracy of PLSR models in estimating the contents of Fe, Cu and Zn are obvious difference. For Fe, the model can approximate quantitative prediction, and for Cu, the model can discriminate between high and low values, while for Zn, the model is unsuccessful. Therefore, for trace elements in soil, the accuracy of PLSR models remains to be further improved.(3) In order to solve the problems that exist in the PLSR models, the CARS feature bands selection method and GLSW filtering algorithm are introduced into the PLSR inversion model. After optimization, the retrieval models are stronger in prediction ability, better in robustness and more compact in structure.There are some problems in retrieval model for estimating soil components with all bands, for example multicollinearity and complexity, etc. A solution to establish better correction model was put forward using CARS that could select the characteristic wave bands and eliminate irrelevant variables, so as to strengthen the ability in prediction and model robustness. With the traditional GA method comparing, the main conclusions are as follows:?with the band variable selection, PLSR prediction model got obvious promotion in estimated performance. As the stability of the model was enhanced, content of OM, TN, Fe, Cu and Zn in soil heavy metal could be effectively predicted.?The improvements of the retrieval model based on CARS or GA are obviously different. The improvements are more effective for Fe, Cu, Zn, while less for OM and TN.?The performance of the retrieval model based on CARS was better than model based on GA. The former could greatly reduce the band number involved in modeling. Relative to the hundreds of thousands of bands participated in the PLSR and GA-PLSR models, this model could effectively avoid the band interference and reduce the risk of overfitting.?Although GA-PLSR model could effectively choose the characteristics bands, it was often difficult to adjust to the optimal state because of the need to adjust the parameters. Also the efficiency of GA is not high in the process of actual operation which limited practical performance of the model.?With the analysis of important bands in PLSR model, the important bands of soil OM, TN, Fe, Cu and Zn are mainly distributed in the long wave band of visible light and near infrared, which is similar to the existing related research conclusion. Therefore, in high spectral estimation of soil components, CARS-PLSR could be used as the preferred method for nondestructive, accurate and fast estimation.The GLSW filtering methods was used to eliminate the interference of water content on soil OM and TN estimation and the traditional OSC method filtering methods was compared. The main conclusions are as follows:?OSC and GLSW could eliminate of moisture interference for soil OM and TN estimation, and the two filtering methods could effectively remove the influence of soil moisture.?The GLSW-PLSR models for estimating soil OM and TN are proved to be more accurate than OSC-PLSR.?With the GLSW filtering optimization, the retrieval model of OM and TN in soil could be transferred among different humidity conditions. As a result, the optimized GLSW-PLSR model can greatly improve the applicability of the retrieval model under different conditions. The results lay scientific foundation for retrieval models that have transferability from indoor to outdoor, and made it possible for estimating soil OM and TN quickly and efficiently using spectra of wet soil.(4)On the basis of retieval model and its optimization, we acquired the spatial distribution characteristics of soil OM, TN, Fe, Cu and Zn.On the basis of retrieval model and its optimization for estimating soil components with hyperspectra, we obtained a series of values of soil components. Based on the spatial autocorrelation of soil, spatial interpolation of soil components was carried out and the space distribution features of soil OM, TN, Fe, Cu and Zn were obtained. Then, the results were validated using the real soil components. On this basis, the spatial distribution characteristics of each soil component in this study area were analyzed. The results showed that:the spatial distribution characteristics of each soil component can be predicted and controlled efficiently using hyperspectral data space. On this basis, the spatial distribution characteristics of each soil component in this study area were analyzed. The results showed that:the spatial distribution characteristics of each soil component can be predicted and controlled efficiently using hyperspectral data space. The real-time information of the soil could be grasped quickly and accurately using hyperspectral technology. It offers basic data to support precision agriculture fertilization and environmental management which can accelerate the development of precision agriculture and environmental protection.