Research On Populus Euphratica Forest Recognition Based On Deep Learning

Accurately obtaining information on Populous ekphratic in remote sensing images plays a key role in environmental monitoring and protection of Populus euphratica.Influenced by the development of remote sensing technology,high spatial resolution remote sensing images not only provide more complex spectral and spatial texture information for the identification of Populus euphratica,but also make the relationship between Populus euphratica cells in the neighborhood window more closely.If the traditional machine learning or pixel-by-pixel recognition method is used to identify and describe the Populus euphratica in the image,it will be difficult to meet the daily application requirements.Therefore,how to use the relevant feature extraction technology and deep learning classification recognition framework to accurately describe the information of Populus euphratica in remote sensing images has become a research hotspot.The main innovations of the paper:(1)In order to enhance the spatial relationship between adjacent pixels,a spatial neighborhood constraint algorithm is designed to constrain the Populus pixels in the neighborhood window.Although remote sensing images are composed of several pixel points,they do not exist independently,but there is some association in space.(2)In order to improve the description ability of each type of feature on Populus euphratica,a linear principal component analysis method is used to express various features.The fusion of multi-type features can not only improve the characterizing ability of the characters in the image,but also describe the content of the image from different directions.(3)Introduce word embedding technology and end-to-end model to describe the semantic information of Populus euphratica in remote sensing image,and use SoftMax classification function to realize the accurate classification of Populus euphratica information.Compared with the simple deep learning model,the algorithm uses the deep-learning end-to-end model’s encoding-decoding framework to mine the semantic information and avoid the loss of detailed information.The experimental results show that the proposed method is above 95% in both Quick Bird and UAV image data.