Crop Classification Of Northern Xinjiang Based On High Resolution Imagery

The planting structure and area of crops are important data sources for predicting agricultural productivity and population carrying capacity.It is of great significance to obtain crop planting information on a regional scale in a timely and accurate manner.With the rapid development of commercial small satellite and UAV remote sensing technology,high-resolution remote sensing images have been widely used in the agricultural field.As the spatial resolution improves,the detailed information of ground features is gradually highlighted,such as geometric structure and texture features.At the same time,the spectral information of the pixels gradually weakens.In the image,the variance within the same type of feature increases,and the difference between the different types of features decreases.Pixel spectral information is relatively lacking,and traditional pixel spectral processing methods are not fully adapted to high-resolution image information extraction.Most crops are planted in fields.Interplanting methods are basically non-existent.The extracted crop planting information should also be a block-like area.However,high-resolution images are generally affected by phenomena such as "identical foreign matter" and "identical foreign matter",and crop growth is also different,which will show different degrees of difference in space.Therefore,this paper explores and studies from three aspects of high-resolution image classification features,object-oriented crop field segmentation,and high-precision classification algorithms.Main tasks as follows:(1)Construction of crop classification features in high-resolution images.Visible light images contain red,green,and blue-gray information and more image structure information.It is difficult to distinguish crop types based on color information alone.From the perspective of pattern recognition,the selection or extraction of representative features for specific classification objects or uses is the key to improving accuracy.Therefore,in combination with the main types of crops in the study area,from the aspects of color,spectral index,texture,and morphology,the characteristics suitable for the classification of crops were dug and selected.After verification,H-CLP,H-Ent,I-Cor,I-CLP,I-Ent,S-CLP,and I-Var features can be used well for crop classification of high-resolution images and can be combined with object-oriented Analysis techniques to further improve classification accuracy.(2)Object-oriented multi-scale segmentation of farmland images.Most object-oriented methods implement the "segmentation-classification" mode.Image segmentation is the basis of subsequent classification and plays a key role in the entire image classification process.Spatial data has two basic characteristics of space and attributes.Scale parameters often differ due to the richness of the features or the degree of spatial aggregation.In order to reduce the impact of non-vegetation information on crop extraction,non-vegetated areas are first identified and eliminated.The non-vegetation in the image is eliminated by the binary classification method using multiple vegetation index principal component components,and the accuracy reaches 93.39%.Then,a layered processing strategy is adopted,and the theoretical optimal spatial scales for local objects in different areas using local estimation methods are 55,95,and 140.Furthermore,object-level information extraction and synthesis is performed on multiple segmentation levels to achieve multi-scale segmentation and multi-level classification and synthesis,so as to improve the degree of refinement of geographic object-oriented image processing and analysis.(3)Improved object-oriented FCN classification.The research on deep learning classification of remote sensing images has now entered the object-oriented CNN(Object-CNN),which has significantly improved the classification accuracy of the algorithm.This paper expands on the model of full convolutional neural network,improves the image features of Chapter 3 and the multi-scale object-oriented segmentation of Chapter 4 were improved.The convolution kernel structure was adjusted and the final prediction output was optimized using CRFs.A remote-sensing image that can automatically identify high-scoring images was constructed.Improved FCN model for crop types.By comparing and analyzing with the traditional method(the accuracy of ENVI SVM is 83.94%),the recognition effect of farmland crops obtained by combining the multi-scale FCN model established in this paper is the best,and the highest accuracy rate is 90.31%.At the same time,the model has certain universality,can be extended to identify other crop types,and achieves better accuracy.