Hyperspectral Retrieval Of Grape Leaf Area Index Based On Feature Selection And Long-short Term Memory Neural Network

Leaf Area Index is defined as the total area of one-sided green leaf per unit ground surface area.It is a key parameter of crop canopy structure,which determines many biophysical processes of crops,ecological processes such as vegetation photosynthesis,respiration,transpiration and carbon cycle.Dynamic information of crops,biomass and yield are closely related to LAI,LAI is also a key input parameter of crop growth model,biogeochemical model and land surface process model.Remote sensing technology has the advantages of wide monitoring range,low cost and nondestructive monitoring,which has been widely used in crop physical and chemical parameter retrieval.Remote sensing monitoring of grape LAI can enhance the objectivity and accuracy of quantitative evaluation of grape growth,and provide scientific basis for timely understanding and mastering grape growth dynamic information and taking scientific management measures(such as water and fertilizer management).LAI remote sensing retrieval is a hot spot of quantitative remote sensing research,and the appearance of hyperspectral remote sensing greatly improves the precision of quantitative retrieval of vegetation physical and chemical parameters.Hyperspectral remote sensing can obtain information in a very narrow spectral band,and its rich spectral information makes it possible to estimate LAI of crops without loss on a large scale.Although hyperspectral data can provide a lot of effec Abstracttive information,there are still many challenges in the process of LAI quantitative retrieval.The problems of LAI retrieval include: Firstly,the traditional machine learning algorithm is widely used in LAI retrieval,but it is a shallow structure algorithm,which has limitations on the processing of complex regression problems and its generalization ability is restricted to some extent.Secondly,crop LAI retrieval mainly uses a single index,and the abundant information of hyperspectral data is lost,the compound information of multiple indexes is not fully utilized.Thirdly,hyperspectral data exist redundancy phenomena,which will consume a lot of computing time,destroy the stability of the model and reduce the LAI retrieval accuracy.Aiming at the above problems in the LAI retrieval process of hyperspectral data,this paper carried out research work based on the simulated spectral data of grape canopy by PROSAIL model,field measured spectral data of grape canopy in jingyang,shaanxi province and simulated GF-5 satellite data.Firstly,the multi-attribute feature space was constructed by the spectral reflectance,vegetation index and key spectral position of vegetation.Secondly,GA,RF,UVE and CARS feature selection algorithms were used to construct a subset of multi-attribute feature space.Finally,deep learning long short-term memory neural network(LSTM)was applied to retrieve LAI.For both simulated and measured data sets,LSTM neural network LAI retrieval can achieve better results.The main research results and conclusions are as follows:1)In this study,an algorithm based on feature selection and LSTM neural network was proposed to retrieve grape LAI.LSTM neural network can simulate the interaction of complex nonlinear features and extract the abstract characteristics of higher levels.The algorithm combining feature selection and LSTM is superior to the classical GA-PLSR machine learning algorithm in both simulated data and field data.GA-PLSR algorithm’s LAI retrieval accuracy R2 in the field measured data,the simulated data of PROSAIL model,and the simulated gf-5 data were 0.9898 0.7279 and 0.7961,respectively;and the retrieval accuracy of the feature selection LSTM algorithm is 0.9994 0.9803 and 0.9808,respectively.1)The multi-attribute feature space was constructed by the spectral reflectance,vegetation index and spectral position,and the accuracy is better than that of single index,such as vegetation index LAI retrieval.The construction of feature space is crucial for LAI retrieval,and multi-attribute feature space can comprehensively consider composite information of multi-source indexes.The common interval of feature selection results of simulated data and measured data,which indicate that the near-infrared band reflectance,the vegetation index(RDVI and MTVI1)and the red-edge area have the best contribution to LAI retrieval accuracy.2)Feature selection algorithm can improve LAI retrieval accuracy of LSTM neural network for deep learning.Feature selection can remove redundant information in high-dimensional data,effectively reduce the dimension of data,and improve the accuracy of LAI estimation of the model.RF feature selection algorithm combined with LSTM neural network achieves optimal results in the field measured data,the simulated data of PROSAIL model,and the simulated GF-5 data.Without feature selection,LAI retrieval accuracy R2 of LSTM neural network were 0.9502 0.8357 and 0.8662,respectively;and R2 of the LSTM neural network combined with RF feature selection algorithm is 0.9994 0.9803 and 0.9808,respectively.