Model Of Cotton Waterlogging Stress Monitoring Based On Hyperspectral Remote Sensing

Cotton and its industrial products are closely related to people's lives and cotton is important strategic resource related to national economy and people's livelihood.Cotton is sown in spring and harvested in autumn.Summer is the key period for determining cotton yield and quality.During this period,the Yellow River Basin is hot and rainy,which is prone to floods.Cotton is sensitive to water and vulnerable to waterlogging,which affects cotton yield and quality.In view of the current situation of cotton production in the Yellow River Basin,this paper proposes to use cotton leaf hyper-spectral data training& testing machine learning algorithms,and deep learning algorithms to establish a classification model that can monitor the waterlogging stress of cotton based on leaves hyper-spectral data,providing an reference for cotton production rescue and output reassessment after flood disaster.This paper designed an experiment to simulate the situation of cotton under waterlogging stress.Hyper-spectral images& relative chlorophyll content of non-waterlogging cotton plants and waterlogging cotton plants at 2,4,6,8,and 10 days were collected using a hyper-spectral imaging camera and a chlorophyll measurement instrument.The cotton leaf area of the hyperspectral image was extracted through the mask file,and the dimensionality reduction was performed by the principal component analysis.The obtained information of the first principal component image accounted for over 91% of the total principal components.The cotton leaf spectral reflectance curve was extracted from the hyperspectral image,and various methods were used for smooth denoising.Comparing the cotton leaf reflectivity for different days of waterlogging,it was found that the spectral reflectance of cotton leaves changed with the number of waterlogging days.The spectral reflectance increases near 550 nm and 750-950 nm,and there were ?blue shift‘ phenomena for the position of red edge of the spectra.Analysis of the relationship between the SPAD and the number of waterlogging days found that the cotton leaf SPAD gradually decreased with the increase of waterlogging stress and was negatively correlated with the number of waterlogging days(r =-0.759).In order to construct the data set required for training and testing machine learning algorithms,18 waveband positions,3 curve areas,and 11 plant cover indices as spectral features were extracted from the cotton leaf spectral reflectance and first & second-order differential curves.8 kinds of statistics of the gray level co-occurrence matrix were considered as texture features extracted from the first principal component images,and cotton leaf SPAD values were considered as features as well.Support Vector Machine(SVM)and Random Forest(RF)algorithms are selected for training and testing.The results show that the overall classification accuracy of the SVM model for the waterlogging stress of cotton plants is 96.296%,and the overall classification accuracy of the RF model is 95.556%,both of them have a good classification effect on cotton leaf samples with non-waterlogging & 8?10 days of waterlogging treatment.GoogLeNet Inception-v3 and VGG-16 convolutional neural network were selected to construct models for monitoring waterlogging stress of cotton based on deep learning algorithms.The band random discard strategy and data augmentation methods were used to expand the cotton leaves hyperspectral images.The first principal component images after the dimensionality reduction of the hyperspectral images were used to establish a classification data set for training and testing the convolutional neural network.The results showed that Inception-v3 model is stable after 100 iterations,with a classification accuracy of 97.04% and a loss value of 0.0868;the VGG-16 model is stable after 300 iterations,with an accuracy of 97.33% and a loss value of 0.0817.The average classification accuracy of the deep learning model is 1.216% higher than that of the machine learning model,the samples are easy to obtain and handle,so that the deep learning model is more suitable as a model for monitoring the waterlogging stress of cotton.The monitoring model of the degree of waterlogging stress of cotton plants has certain reference value for the cotton production rescue strategies and the statistics of cotton losses after summer floods in cotton production areas of the Yellow River Basin.