| Desert grassland rodent pests are one of the important factors limiting the healthy development of grassland ecosystems.In recent years,Chinese grasslands have been gradually desertifying due to the combined effects of global climate change and human factors.The desertification of grasslands provides suitable habitats for rodents,which makes their populations increase greatly and eventually causes the occurrence of rodent infestation in desert grasslands.The proliferation of grassland rats not only aggravates the degradation process of desert grassland,but also carries a large number of viruses that threaten the health of humans and animals.Accurately grasping the spatial distribution of rodent populations in relation to vegetation and soil is an important prerequisite for the implementation of ecological prevention and control of rodent infestation in desert grassland,which is of great significance to the ecological protection of grassland and sustainable development of livestock industry.However,the traditional manual grassland rat pest information survey method has many problems such as a time-consuming,costly and long survey cycle,and it is difficult to realize real-time and regional dynamic monitoring.In addition,satellite remote sensing cannot meet the requirements of accurate identification of fine grassland ground objects due to the limitation of spatial resolution and the influence of clouds and fog.Therefore,in order to realize intelligent monitoring of grassland rodent infestation in desert grasslands,this thesis formed an unmanned aerial vehicle(UAV)hyperspectral remote sensing platform using a six-rotor UAV and a hyperspectral imager,and used the platform to collect hyperspectral remote sensing images of a variety of ground objects samples in Inner Mongolia desert grasslands.Subsequently,a model based on UAV hyperspectral remote sensing was established for extracting the information on rodent pests in the desert grassland through the spectral feature analysis method and deep learning method,which provides a model basis for the statistics and monitoring of rat pest information in the desert grassland.The main research contents and conclusions of this thesis are as follows:(1)To address the requirement that the existing vegetation indices cannot meet the identification of rat holes in desert grassland,a rat hole index identification model was established by the spectral feature analysis method.Based on the statistical analysis of a large number of thresholds,the optimal separability threshold interval of mouse holes was determined,and the accurate extraction of mouse holes in desert grassland was realized.The results showed that the overall accuracy and Kappa coefficient of the model could reach 97.00% and 0.9343,respectively,and proved the effectiveness of the proposed model by having higher recognition accuracy compared with various vegetation index models.(2)To address the blank of deep learning in UAV-based hyperspectral rodent infestation monitoring,the 3D-Dense Net network model is improved by introducing the idea of residuals and asymmetric convolution structure,and then proposing the deep learning model of three-dimensional deep dense residual network(3D-DDRNet).Firstly,by comparing the ability of deep learning and machine learning for grassland fine ground objects recognition and classification,the feasibility and effectiveness of deep learning for desert grassland rodent infestation monitoring ground objects recognition and classification are verified,and the 3D-DDRNet model is established on this basis.Secondly,by optimizing the four main parameters of the 3D-DDRNet model,the overall accuracy and Kappa coefficient of the optimized model were obtained as 96.68% and0.922,respectively.in addition,the overall classification accuracy of the 3D-DDRNet model was improved by 1.46% before compared with the original 3D-Dense Net model,and the training time and model parameters were reduced by 79.8% and 18.3%,realizing the high-precision recognition and classification of desert grassland rodent infestation monitoring ground objects based on UAV hyperspectral remote sensing.Finally,the effectiveness of the proposed model is further verified by comparing it with several deep learning models.(3)To address the problems of small sample learning and lightweighting of the model,a lightweight transformer attention network(TAN)deep learning model is proposed by improving the transformer model combined with a two-dimensional convolutional neural network.The model adopts a two-stage feature extraction structure,which effectively improves the classification performance of the model.In each stage,local features are firstly extracted by a fixed convolution kernel to enhance detailed texture features;secondly,the extracted local features are refined by a contour convolution(CC)module to enrich feature information at the edges of the feature map;finally,the global features are extracted by transformer attention(TA)module to improve the classification performance.The TA module is used to focus on the global pixels,thus suppressing the background information and enhancing the output of effective information.The results show that the overall accuracy of the TAN model is 97.71% and the number of parameters is only 0.17 M.By comparing with several deep learning models,the proposed model has a high classification performance and achieves high accuracy recognition and classification of desert grassland rodent pest monitoring ground objects based on small sample learning.(4)To address the problems of difficult classification of desert grassland vegetation species and difficult labeling of data samples,we propose a lightweight Neighborhood aggregation network(NANet)deep learning model by constructing the pixel-neighborhood features of hyperspectral image data.In addition,to reduce the redundant features of hyperspectral image data and improve computational efficiency and accuracy,a genetic algorithm and optimal index factor are used for band selection.The model first learns the features of hyperspectral image data by a two-layer neighborhood feature aggregation function,and then uses MLP for the final classification result output.By comparing with SVM and various deep learning models,the NANet model has the highest computational efficiency,the smallest model size,and the highest classification accuracy.The results showed that the overall accuracy of the NANet model was 93.47%,and the number of parameters was 0.15 M,which achieved the recognition and classification of desert grassland vegetation species.(5)To address the problem of low classification accuracy of desert grassland vegetation species,we designed the local feature enhancement(LFE)and global feature enhancement(GFE)modules and combined them with the convolutional block attention module(CBAM)to propose a lightweight local-global feature enhancement network(LGFEN)deep learning model.The designed LFE and GFE modules can effectively learn the local-global information of hyperspectral images and enhance the learning ability of the model for features.Finally,the CBAM is used to further fuse and refine the learned features to enhance the classification performance and stability of the model.By comparing with the latest hyperspectral image classification methods,the LGFEN model has a high classification performance.The results show that the overall accuracy of the LGFEN model is 98.61%,and the number of parameters is 0.18 M,which achieves the high-precision identification and classification of desert grassland vegetation species. |
PDF Full Text Request