Classification And Inversion Of Grassland Features Based On UAV Hyperspectral Remote Sensing

Grassland degradation is one of the most serious ecological challenges facing our country and the world.Over the past half century,more than 30%of the world’s land has been affected by desertification to varying degrees,and more than 90% of our country’s natural grasslands are in the process of desertification to varying degrees.Therefore,it is imperative to monitor grassland ecology and curb grassland degradation.The high-precision identification and inversion of degraded grassland features is the basis of grassland ecological monitoring.The national standard "Natural Grassland Degradation,Desertification,and Salinization Grading Indicators" uses plant community coverage,community plant composition structure,grassland degradation indicator plant species proportion,rat hole area,etc.as monitoring indicators for natural grassland degradation and desertification.Therefore,rapid and high-precision dynamic monitoring of degraded grassland plant communities and plant species is an important basis for assessing the degree of grassland degradation and grassland resource management.Advantages of UAV hyperspectral in grassland survey and research.Traditional manual surveys are time-consuming and laborious,and cannot meet the needs of large-scale grassland surveys.Although satellite remote sensing has broken through the limitations of the spatial scale of manual surveys,it is limited by spatial resolution and spectral resolution,and can only achieve statistical research on large-scale vegetation coverage.For vegetation subtypes or small-scale community structure characterization There are obvious deficiencies.In recent years,the UAV hyperspectral remote sensing platform can provide richer fine-grained features of research objects by collecting hyperspectral remote sensing images with centimeter-level spatial resolution,and providing a hardware foundation for grassland degradation monitoring.Advantages of deep learning in remote sensing data analysis.Traditional analysis methods mainly use characteristic bands to establish spectral indices,and accomplish tasks such as image classification through threshold segmentation.This method has the limitations of cumbersome processes,heavy tasks,and high requirements for prior knowledge.Compared with the spectral index analysis method,traditional machine learning has greatly improved the accuracy of the model,but it still has the disadvantage of requiring handcrafted features.In recent years,deep learning has rapidly become the frontier technology of remote sensing image analysis due to its advantages of automatic feature extraction,high precision and high efficiency.Deep learning automatically obtains spectral information,spatial information,and spectralspatial joint information from remote sensing images through convolution operations,avoiding the tedious manual design of features,improving accuracy,and providing a method for grassland degradation monitoring.This study takes the desert grassland in central Inner Mongolia as the research object,and collects the hyperspectral remote sensing data of the desert grassland through the UAV hyperspectral remote sensing platform.Aiming at the problem that the spectral difference between desert grassland features is small and the discrimination degree is not significant.In this study,the vegetation index is established by looking for characteristic bands,the distribution range of each object is counted,and the segmentation threshold is determined to complete the classification.After accuracy verification,the overall classification accuracy is 82.01%,and the Kappa coefficient is 0.79.The results show that the threshold classification method based on vegetation index meets the accuracy requirements of remote sensing research,and the algorithm is simple and intuitive.Aiming at the problems of insufficient extraction of small target features and large interference of mixed pixels,the recognition accuracy of degraded indicator species is low.This study uses deep learning methods to create a DISCNN model,and obtains the best DIS-CNN model by optimizing the model structure and model parameters,and conducting performance comparisons with the basic model and other classic algorithms.The results show that the overall accuracy of the DIS-CNN model proposed in this thesis is 98.78%,the average accuracy is 96.91%,and the Kappa coefficient is 0.92.In addition,compared with other models,it is proved that DIS-CNN has obvious advantages in identifying and classifying indicator species of desert steppe degradation.Aiming at the problem of low recognition accuracy of prairie rat holes caused by small samples,high redundancy and nonlinear mixed pixels.This study builds the GM-CBAM model based on algorithm ideas such as multiscale three-dimensional convolution,dual-branch feature fusion,and CBAM attention mechanism.Through the optimization of parameters such as Spatial Size and Learning Rate,different combinations of channel attention modules and spatial attention modules The best GM-CBAM model is obtained sequentially.The results show that the overall accuracy and Kappa coefficient of the optimized GM-CBAM are 99.35% and 98.90%,respectively,which are3.96% and 3.35% higher than the basic model.It provides technical support for the subsequent rapid interpretation of prairie rat holes and the evaluation of rodent damage levels.Aiming at the problem of insufficient ability of CNN algorithm to capture global information.This study proposes the Grassland Former model.In the Grassland Former model,the spatial-spectral convolutional feature extraction module(SSCFE)and the encoder cross-layer dynamic fusion module(CLDF)were redesigned,and the best model was obtained through model parameter optimization.The overall classification accuracy was 98.84%,with an average accuracy is 97.47% and the Kappa coefficient is 0.98.In addition,the superiority of the Grassland Former model is proved by comparing it with other classic models.Aiming at the problem of low inversion accuracy caused by mixed pixel and fault data in the inversion of grassland features.In this study,a DGIHCNN inversion model was constructed.The best inversion model was obtained through model parameter optimization.The overall accuracy of the optimized DGIH-CNN model in training reached 99.30%,with a Kappa coefficient of0.98,and the overall accuracy of inversion testing reached 99.02%,with a Kappa coefficient of 0.96.In addition,the superiority of the DGIH-CNN model in the inversion of desert grassland features is proved by comparing the performance with the other 5 classic models.In this study,based on the UAV hyperspectral remote sensing system,the hyperspectral remote sensing data of the desert grassland in central Inner Mongolia were collected,and the vegetation index threshold method,3 deep learning classification models and 1 deep learning inversion model were used to realize the desert grassland features.High-precision identification and inversion is a beneficial attempt to classify and invert desert grassland features based on low-altitude remote sensing information.It provides methods and models for high-precision grassland ecological monitoring,and provides data for the study of the degradation and succession of dominant species in desert grasslands.It also provides a decision-making reference for the ecological restoration of degraded grasslands.