| With the development of aerospace technology,remote sensing images have become an important data source for human rapid earth observation.The accuracy evaluation of remote sensing image classification results is an important guarantee before remote sensing image application decision-making.Sampling inspection is one of the commonly used methods for product quality evaluation and data accuracy evaluation.The existing sampling inspection method is more suitable for the accuracy inspection of "independent and homogeneous" products,and does not consider the spatial correlation and spatial heterogeneity of remote sensing image classification results.Because of the characteristics of nature,the existing sampling inspection methods are directly used to evaluate the accuracy of remote sensing image classification results,and there are problems such as high sample information redundancy and long evaluation process.Therefore,this paper aims to evaluate the accuracy of remote sensing image classification results,and uses the aggregation index to propose a multi-layer and multi-granular spatial sampling method for accuracy evaluation of remote sensing image classification results.The specific research contents include:(1)Using the spatial heterogeneity of remote sensing images,the accuracy evaluation method of remote sensing image classification results with multi-layer spatial sampling is proposed to realize the rapid accuracy evaluation of remote sensing image classification results in a large rangeAiming at the problems of high information redundancy and long evaluation time in the accuracy evaluation of remote sensing image classification results,the aggregation index is used to quantify the spatial heterogeneity of remote sensing images,and a multi-layer spatial sampling method for accuracy evaluation of remote sensing image classification results is proposed.Including: 1)The selection process of sample points from the whole to the zoning space: use the aggregation index to quantify the spatial heterogeneity of the remote sensing image,and realize the spatial zoning of the remote sensing image;then calculate the weight index of each zoning space,and allocate the samples in each zoning space Quantity;lay out sample points in each division space.2)The accuracy evaluation process from the block space to the whole: compare the accuracy of the sample point information in each zoning space with the reference data,and mark the consistency as "1" and the inconsistency as "0";derive each zoning space according to the evaluation statistics The accuracy of classification;the accuracy of the classification results of the entire remote sensing image is derived from the classification accuracy of each division space.(2)Aiming at the spatial correlation of remote sensing images,the accuracy evaluation method of remote sensing classification results based on multi-granular spatial sampling is proposed to realize the rapid accuracy evaluation of high-resolution remote sensing images.The types of features in remote sensing images have multi-scale features,that is,the same features have different spatial morphological features on images with different resolutions.Therefore,the fixed-size sampling unit cannot take into account the spatial morphological characteristics of the ground objects,and it is easy to cause misjudgment,especially for the accuracy evaluation of the classification results of high-resolution remote sensing images.Taking into account the spatial correlation of remote sensing images,this paper proposes a multi-granular spatial sampling high-resolution remote sensing image classification accuracy evaluation method,including: 1)Multi-granular sampling unit determination: using connected component labeling method to calculate the number of connected regions in each region,Use it as the grouping basis of the k-means algorithm;cluster out the area of the smallest patch in each region,and determine the size of the sampling unit in each zoning space;2)The determination of sample size and the layout of sample points: Calculate the sample size of each zoning space The size,with multigranularity patches as the sampling unit,allocate sample points in each division space;3)Accuracy evaluation of remote sensing image classification results: compare the accuracy of the sample point information in each division space with the reference data,and use the area dominance method The accuracy evaluation of the feature information of the multigranularity sampling unit is marked as "1" for consistency and "0" for inconsistency;the classification accuracy of each division space is derived based on the evaluation results;the entire remote sensing image is derived from the classification accuracy of each division space The accuracy of the classification results.The high-resolution remote sensing image of the Urban dataset is used as the experimental data,and five types of ground object classification maps in the same area are selected as the verification data.The feasibility analysis of method(1)is carried out,and the experimental results show that this method improves the representativeness of sample points by quantifying the spatial heterogeneity of remote sensing images,that is,in areas with a high degree of aggregation of ground features,the probability of sample points High;In areas with low ground object types,the sampling probability of sample points is low;this method reduces the information redundancy of sample points,and better solves the problem of rapid accuracy evaluation of large-scale remote sensing image classification results.At the same time,the feasibility analysis of method(2)is carried out.The experimental results show that this method designs multi-granularity sample units by quantifying the spatial correlation of remote sensing images,that is,the classification results of remote sensing images of different features have different sampling results.Unit,the method takes into account the diversity of the spatial morphology of the remote sensing image,increases the representativeness of the sample,and improves the accuracy of the accuracy evaluation of the classification results of the high-resolution remote sensing image. |
