Land Cover Product Updates Using Time Series Images

Land cover is a combination of various material types and their natural attributes and characteristics on the land of the earth.It is the most obvious and most commonly used indicator for characterizing the land and corresponding human or natural processes.The land covering product is a concrete manifestation of various types of features,the use of remote sensing image data to generate land covering products is the main method of making land covering products.At present,there are two main methods for producing land cover: one is to separately classify each period of images to obtain land cover products separately;Change detection is performed between images,and only the changed parts are classified to obtain change patches and then incrementally updated.The second method can increase the speed of product generation,reduce errors,and ensure consistency between products.The second method can increase the speed of product generation and reduce errors.The accuracy of the second production of land covering products mainly depends on four parts: the accuracy of the reference product,the accuracy of change detection,the classification accuracy of the changed area,and the accuracy of the update process based on the reference product.In the current production process of land cover products,there are mainly the following problems: the current product resolution is not high,the accuracy of the existing classification method is not high,the product update speed is slow,And the problem of sliver polygons caused by errors in geometric correction and boundary fit errors caused by fuzzy judgment in change detection during the process of extracting change patches from time-series images.In view of the above problems,the research is mainly from two aspects of improving the classification accuracy of the changing area and the accuracy of the update process based on the benchmark product.First,the superpixel collaborative segmentation algorithm is used to extract the change patches from the image;Then,combined with the local spectral signature evidence the expected category change evidence,and D-S evidence fusion algorithm,the evidence decision classification rules are designed to classify the remote sensing image of the changed area;finally,the concept of sliver polygons is introduced to adjust the boundary of the incremental partial remote sensing image classification results,improve the boundary fit,and improve the accuracy of the product during the update process.The study takes the domestic Gaofen-1 wide field of view time series remote sensing images in the northwestern region of China as an example to conduct experimental exploration.The overall classification accuracy of the incremental part of the two phases of remote sensing images obtained by using classification decision rules has reached 93.7% and 91.6 %.Through the processing of sliver polygons,the sliver polygons in the two-phase images only accounted for 0.96% and 0.98% of the total area.Although they accounted for a relatively small amount,they did adjust the boundary of the feature type of the products covered by the land.The overall accuracy of the resulting updated two-stage land cover products was 85.2% and 86.3%.The innovations in this article include:(1)Design evidence decision classification rules,classify remote sensing images of the changed area by super-pixel cooperative segmentation,and improve the classification accuracy of incremental remote sensing images.(2)Introduce the concept of sliver polygons,adjust the boundaries of the classification results through the sliver polygons,improve the boundary fit between the target period products and the reference land cover products,and improve the continuity of the changes in land cover products.