Identification And Classification Of Grassland Degradation Indicator Based On UAV Hyperspectral Remote Sensing

The identification and statistics of grassland degradation indicator features is an important fundamental part of grassland ecological monitoring.This steps is also an important quantitative parameter for grassland degradation degree and an important basis for the development of grassland ecological restoration programmes.Over the last half century,the earth's ecosystem has been facing serious challenges due to human activities and climate change.More than 34% of the terrestrial ecosystem in the world is already in different stages of desertification.Grassland ecosystem is the most vulnerable landscape to desertification due to its simple community structure.Nowadays grassland ecosystem has become the major source of desertification and one of the main sources of dust storms,which has seriously affected and even threatened the development and survival of local people.Due to degraded grasslands are widely distributed and of varying degrees of degradation,the key to restoring grassland ecology is to tailor precise measures to local conditions.And efficient grassland ecological information statistics and accurate surveys of the degree of degradation are essential prerequisites.Plant community cover,the relative proportion of degradation indicator species in grassland and vegetation composition in vegetation communities are all necessary monitoring items according to the national standard "Grading Indicators of Natural Grassland Degradation,Desertification and Salinization".However,traditional artificial field survey consumes a lot of manpower and cannot realize regional dynamic monitoring.The quality of satellite remote sensing imaging is limited by vulnerability to weather,long revisit cycles and low spatial resolution,making it difficult to achieve accurate statistics on small features.The lack of statistical and monitoring tools for grassland ecological information restricts the process of ecological restoration.So th ere is an urgent need for a quick,non-destructive,high-precision,large-area monitoring tool and the method for identification and classification of grassland degradation indicators.The identification and classification of degradation grassland indicator ground object based on UAV hyperspectral remote sensing not only lays the foundation for real-time monitoring and evaluation of grassland degradation,but also makes quantitative research on grassland degradation based on remote sensing possible.Which will provide a new means and model for grassland degradation monitoring and research to achieve efficient,real-time,accurate and quantitative research,contributing to the prevention and control of grassland degradation in China.In this study,a low-altitude unmanned aerial vehicle(UAV)hyperspectral remote sensing platform was established,which has the advantages of map integration,high spatial resolution,high spectral resolution,mobility and flexibility,providing a hardware basis for grassland ecological information statistics and monitoring.Remote sensing images of desertification grasslands in Inner Mongolia were collected for two consecutive plant growth periods to obtain base data on grassland degradation ground indicator features in the study area.Due to the limitations of data acquisition and classification methods,currently,the UAV-based remote sensing feature identification is still at the level of identification and classification of a single feature over a large area.And desertificati on grassland with features of smaller plants,high similarity between populations,fragmented patchy distribution,difficult to identification and classification,puts forward higher requirements on the quality of acquired data and classification methods,and cannot be solved with the existing data acquisition and classification methods.Only by making a breakthrough can classification accuracy be achieved and the research be of practical application value.This study focuses on the problem of high similarity of degraded grassland features,which are difficult to classify.Combining the characteristics of grassland degradation indicator ground objects and the characteristics of UAV hyperspectral data,a high-precision recognition and classification model between communities and populations of grassland degradation indicator features is established based on traditional image classification methods and deep learning classification methods respectively,providing a model basis for grassland ecological informatio n statistics and monitoring.In this study,CSI and CDI indices were proposed and classification rules were established.Which solved the problem of high similarity of grassland species and weak spectral differences that make them difficult to classify,and achieved high precision classification among construction species,dominant species and companion species in grassland degradation indicator ground object.The overall accuracy of the classification was evaluated to be 93.12% with a kappa coefficient of 0.91.Moreover,a combination of feature band extraction,PSIR index,spectral scaling and threshold statistics were used to establish the classification rules for plant communities in grassland degradation indicator ground object with bare soil and withered herbage,achieving high precision classification among communities in grassland degradation indicator ground object.The degree of generalization of deep learning classification models varies.At present,there are few studies on classification models for low altitude UAV hyperspectral high-dimensional and high spatial resolution data,and even fewer studies on classification models for features with fine vegetation in desertified grasslands,random distribution and the presence of a large number of mixed image pixels.In this study,a lightweight DGC-3D-CNN model was established and a UAV hyperspectral grassland degradation indicator ground object community dataset was produced to evaluate the model classification performance.The study showed that the DGC-3D-CNN model made full use of the combined spectral-spatial information of the UAV hyperspectral data and showed well classification potential for grassland degradation indicator features in the community dataset.By optimizing the nine parameters of the model,the efficiency and accuracy of the model are further improved,and the number of convolutional kernels,Batch size,and Spatial size of the DGC-3D-CNN model are found to have a greater impact on the classification accuracy of the community dataset,while a lighter network structure and smaller convolutional kernel size had a better performance in classifying the grassland degradation small indicator ground object.The model achieved high accuracy in classifying degraded indicator features between communities.In this study,a lightweight GDIF-3D-CNN model was established to deal with the problem of high similarity between grassland plant populations and the difficulty of classification.Two types of UAV hyperspectral grassland degradation population datasets,namely the population pure pixel dataset and the population mixed pixel dataset,were producedto evaluate the classification performance of the model.The study showed that the GDIF-3D-CNN model had good classification potential for population building and dominant species,with the classification accuracy of single class exceeding 91%.The efficiency and accuracy of the model were further improved by optimising the eight parameters of the model.And the overall accuracy of the classification of the population pure pixel dataset reached 94.815% with an efficiency improvement of 76.2%.The overall accuracy of the classification of the population all pixel datasets reached 92.625%with an efficiency improvement of 76.5%.The overall accuracy of the model was92.625% and the efficiency was improved by 76.5%,achieving a high accuracy classification between populations of degraded indicator ground object.using deep learning methods.The two vegetation indexes with classification rules and two depth classification models established in this study have achieved high precision classification of inter-community and inter-population of grassland degradation indicator ground object in UAV hyperspectral data.Which is a meaningful exploration on classification research of desertification grassland ground object based on remote sensing information,laying the model foundation for real-time,efficient and high-precision ecological information monitoring and statistics of desertification grassland,and providing reference for the precise implementation of ecological restoration measures for desertification grassland.