WO2024020744A1 - 基于高分辨率影像的生态图斑提取方法 - Google Patents

基于高分辨率影像的生态图斑提取方法 Download PDF

Info

Publication number
WO2024020744A1
WO2024020744A1 PCT/CN2022/107698 CN2022107698W WO2024020744A1 WO 2024020744 A1 WO2024020744 A1 WO 2024020744A1 CN 2022107698 W CN2022107698 W CN 2022107698W WO 2024020744 A1 WO2024020744 A1 WO 2024020744A1
Authority
WO
WIPO (PCT)
Prior art keywords
extraction
vector
image
ecological
network
Prior art date
Application number
PCT/CN2022/107698
Other languages
English (en)
French (fr)
Inventor
胡晓东
石含宁
夏列钢
张明杰
骆剑承
郜丽静
董文
雷一鸣
张乃祥
韩小妹
Original Assignee
苏州中科天启遥感科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 苏州中科天启遥感科技有限公司 filed Critical 苏州中科天启遥感科技有限公司
Priority to PCT/CN2022/107698 priority Critical patent/WO2024020744A1/zh
Publication of WO2024020744A1 publication Critical patent/WO2024020744A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks

Definitions

  • the present invention relates to the technical fields of ecology and remote sensing, and in particular to a "patch-level" ecosystem extraction method based on high-resolution remote sensing images.
  • An ecosystem is a unified whole composed of organisms and the environment within a certain space in nature.
  • the scope of ecosystems can be large or small, and the representation of ecosystems required at different scales is also inconsistent.
  • Most of the data are based on medium-resolution images, and the surface is divided into different ecosystem types through pixel-oriented classification. This method can quickly obtain ecosystem types under limited scale conditions and has a high accuracy.
  • the county scale due to the narrowing of the representation range, medium-resolution image representation information will also be greatly reduced, and the content represented in a single pixel is complex and diverse, resulting in impure ecosystem types.
  • High-resolution remote sensing images have richer spatial information characteristics than mid- and low-resolution images, and have more expressive content than mid-resolution images. If the pixel-oriented classification method is still used to classify ecosystem types, it will lead to fragmented and discontinuous patterns, which is not visually beautiful. High-resolution images can clearly identify each independent spatial geographical entity, and extracting specific geographical entities as objects can effectively reduce the fragmentation of map patches. At the same time, since the object represented by each patch is unique, its shape and boundary are highly consistent with the performance content on the image. Compared with the regular grid, its internal information is unique, and there are no multiple ecosystems in one patch. Mixed system types.
  • the purpose of the present invention is to provide an ecological patch extraction method based on high-resolution images, which can simplify complex land surfaces, concentrate elements with similar characteristics into several relatively independent sub-regions, and reduce the complexity of feature extraction. Difficulty.
  • embodiments of the present invention provide an ecological patch extraction method based on high-resolution images, which includes the following steps: obtaining high-resolution images, extracting road network, water system and terrain line data; Road network, water system and terrain data are used as extraction masks to split the image into multiple independent sub-task areas for extracting ecological patches in each region; based on the image characteristics in each sub-task area, the surface elements are classified Layer extraction, the surface elements include cultivated land, buildings, water surfaces, grasslands, woodlands and shrubs; and for the vector results of each land type extracted layer by layer, in order from large scale to small scale, the extracted ecological The patches are updated to the same vector file to form a full-element ecological patch vector, and the final ecosystem patch vector is obtained.
  • the main idea of the present invention is to deconstruct the complex land surface according to the hierarchical partitioning idea of geography.
  • the road network, water system and terrain data are used to divide the image into several independent sub-task areas, and then according to each sub-task Features of elements within the area are layered according to buildings, waters, cultivated land, forests and others. According to the different visual characteristics of each layer of elements, corresponding methods are selected to extract them in sequence. Finally, the data of each layer are merged to obtain the ecological environment of the complete region. Pattern vector data.
  • obtaining high-resolution images and extracting road network, water system and terrain line data include: obtaining high-resolution remote sensing images of the target area; and performing original processing of the high-resolution remote sensing images.
  • the data is subjected to image preprocessing, which includes orthorectification, radiometric calibration, atmospheric correction and cloud removal; based on the surface characteristics in the image, the main road network and river system are drawn; and the ridge lines and sums of the image areas are calculated. Valley lines and form a terrain network.
  • splitting the image into multiple independent sub-task areas includes: updating the road and terrain network to a unified layer file to form a complete partition control network; and use the partition control network to crop the image to form multiple sub-task areas with small data volumes.
  • the hierarchical extraction of surface elements based on the image characteristics in each sub-task area includes: sample production, model training and image prediction, raster vectorization and vector post-processing. processing, and vegetation extraction.
  • sample points of regular cultivated land, terraced fields, sloping cultivated land, construction areas and water surfaces are selected; according to different types of sample points, images of fixed sizes are cropped and Corresponding vector frame; perform feature drawing on the cropped sample file, and assign non-zero attribute information to the drawn vector; and convert the assigned sample into a binary image for training.
  • the sample is input into the corresponding convolutional neural network for training, and the corresponding cultivated land, building and water surface extraction models are obtained; and the images and extraction models of each sub-area are obtained Input it into the corresponding neural network to predict the image and obtain predicted intensity maps of farmland, buildings and water surfaces.
  • the predicted intensity map is vectorized.
  • the edge intensity map uses relevant tools to extract skeleton lines, connects the broken lines and then vectorizes them. After the vector lines are simplified and straightened, Convert it into a vector surface and obtain the pattern pattern; the surface predicted intensity map is converted into a vector surface result after expansion, hole filling and binarization processing. In vegetation extraction, forest, grassland, and shrub patches are obtained, and the vector results of vegetation extraction are obtained.
  • the vegetation extraction includes: using cultivated land, buildings and water surfaces as masks, and using terrain lines as constraints, performing multi-scale segmentation on the remaining image range; based on the segmentation results On the top, select the forests, shrubs and grasslands as classification samples, use the random forest method to identify all segmented patches, and obtain the forest, grassland, and shrub patches; and perform edge simplification and smoothing processing on the extraction results to obtain the final Vegetation extraction vector results.
  • forming a full-element ecological patch vector from the hierarchically extracted vector results of each land type includes: merging the same type of patches in each sub-task area to obtain a complete Extract the result vectors of each ecosystem type in the region; update the extracted "forest-shrub-grassland-water surface-building-cultivated land” ecological map patches to the same vector file in the order of updating from large scale to small scale, and then Update roads and water systems into the vector file to form a full-element ecological patch vector; and perform a single-part to multi-part operation on the updated layer to eliminate small fragments and narrow patches to obtain the final A diagram of an ecosystem.
  • the sample is input into the corresponding convolutional neural network for training, and the corresponding cultivated land, building and water surface extraction model is obtained, including: cultivated land sample Input into the RCF network for training, building samples into the D-LinkNet network for training, water surface samples into the U-Net network for training, and obtain corresponding farmland, building and water surface extraction models.
  • the cultivated land samples are input into the RCF network for training, which includes: conducting an overall analysis of the cultivated land plots in the entire area, selecting cultivated land samples evenly within the image range, and drawing the cultivated land boundaries therein Information and converted into binary results as label data as training data for the edge monitoring network; by training boundary features, an extraction model for cultivated land boundary information is obtained, and the cropped images of each task area are predicted to obtain the boundary prediction of each task area.
  • Strength result map obtain a binary map of cultivated land prediction results through skeleton line extraction and broken line connection; and obtain cultivated land extraction vector results through raster vectorization, vector straightening, smoothing and other post-processing.
  • the building samples are input into the D-LinkNet network for training, which includes: analyzing the residences in each sub-task area, and extracting residences with similar characteristics as the same category Object, create building samples with different image characteristics, draw the building areas in the samples and use attribute values to distinguish them from the background; convert the drawn building area samples into binary label data based on attribute values; input the building binary label data into D -LinkNet network is trained to obtain the extraction model of residential areas; the network is used to predict the images of sub-regions, and the residential area prediction intensity map of the corresponding area is obtained; and the final built-up area is obtained through raster vectorization and vector post-processing operations.
  • the vector extraction result includes: analyzing the residences in each sub-task area, and extracting residences with similar characteristics as the same category Object, create building samples with different image characteristics, draw the building areas in the samples and use attribute values to distinguish them from the background; convert the drawn building area samples into binary label data based on attribute values; input the building binary label data into D -LinkNet network is
  • the water surface samples are input into the U-Net network for training, which includes: analyzing the characteristics of the river water surface, lake water surface, and pond water surface in the image, and selecting representative Rivers, lakes and ponds are used as samples, the water surface is drawn, and the difference from the background is marked according to the attributes; the drawn samples are converted into binary image label files according to the attributes, and input into the U-Net network Train the features to obtain a water surface extraction model; use the water surface extraction model to predict regional images and obtain the water surface prediction intensity map of the image; and use raster vectorization and vector post-processing tools to obtain the final water surface vector extraction result. .
  • the ecological patch extraction method based on high-resolution images has the following advantages:
  • the ecological patch extraction method based on high-resolution images of the present invention uses the partition idea to partition the data, which can simplify the complex surface and concentrate elements with similar characteristics into several relatively independent sub-regions, reducing the The difficulty of feature extraction.
  • generally high-resolution images have a large amount of data.
  • direct extraction is often time-consuming and laborious, or even impossible to extract, and the hardware requirements are relatively high.
  • partition processing an image with a large amount of data is divided into several small areas, which can effectively disperse the data amount of a single image and reduce the possibility of problems caused by the large amount of data during the extraction process.
  • multiple machines can also be used for multi-machine extraction. Tasks are processed in parallel, significantly shortening the extraction time.
  • the surface is divided into types with their own characteristics, and the targeted extraction method is selected based on the specific visual characteristics of each type. Compared with other direct extraction methods without partitioning and stratification, The effect of wrong mentions and missed mentions has been significantly improved.
  • Figure 1 is a simple flow chart of an ecological patch extraction method based on high-resolution images according to an embodiment of the present invention
  • Figure 2 is a detailed flow chart of an ecological patch extraction method based on high-resolution images according to an embodiment of the present invention
  • Figure 3 is an example of a superimposed vector pattern of a grassland ecosystem according to an embodiment of the present invention.
  • Figure 4 is an example of a superimposed vector pattern of a farmland ecosystem according to an embodiment of the present invention.
  • the ecological patch extraction method based on high-resolution images includes the following steps S1-S4.
  • step S1 high-resolution images are acquired and road network, water system and terrain line data are extracted.
  • the main road and water system data within the image area can be obtained from public data, and combined with the actual image features, GIS software can be used to supplement the missing roads and water systems to form the water system and road network data of the image area.
  • GIS software can be used to supplement the missing roads and water systems to form the water system and road network data of the image area.
  • DEM digital elevation models
  • step S1 may include: 1) Obtaining high-resolution remote sensing images of the target area, such as Google Image, GF2, QuickBird, etc.; 2) Performing orthorectification, radiometric calibration, atmospheric correction, and cloud removal on the original data and other image preprocessing; 3) Draw the main road network and river system based on the surface characteristics in the image; 4) Use DEM to calculate the ridge lines and valley lines of the image area, and form a terrain network; 5) Update the road and terrain network to a unified layer file to form a complete partition control network; 6) Use the partition control network to crop the image to form several sub-task areas with small data amounts.
  • the target area such as Google Image, GF2, QuickBird, etc.
  • step S1 may include: 1) Obtaining high-resolution remote sensing images of the target area, such as Google Image, GF2, QuickBird, etc.; 2) Performing orthorectification, radiometric calibration, atmospheric correction, and cloud removal on the original data and other image preprocessing; 3)
  • step S2 the road network, water system and terrain data are used as extraction masks to split the image into multiple independent sub-task areas (multiple blocks) for regional ecological map spot extraction.
  • step S2 the road and terrain network can be updated into a unified layer file to form a complete partition control network; and the partition control network can be used to crop the image to form multiple sub-task areas with small data amounts.
  • step S3 based on the image features in each sub-task area, surface elements are extracted hierarchically.
  • the surface elements include cultivated land, buildings, water surfaces, grassland, woodland and shrubs.
  • Sample production According to the image characteristics in each sub-region and the type of cultivated land, sample points of regular cultivated land, terraced fields and sloping cultivated land, built-up areas and water surfaces are selected; the samples need to be evenly distributed and in moderate quantity. According to different types of sample points, images are cropped into fixed-size images and corresponding vector frames. Among them, regular farmland and terraced fields are line vector files, and sloping farmland, buildings and water surfaces are area vector files. Carry out feature drawing on the cropped sample file, draw edge feature information in the plot for regular farmland and terraced fields, draw the range of internal texture feature information for sloping farmland, buildings and water surfaces, and assign non-zero attribute information to the drawn vectors. Finally, the attribute-assigned samples are converted into binary images for training.
  • Model training and image prediction According to the type of sample, the sample is input into the corresponding convolutional neural network for training. Among them, cultivated land samples are input into the RCF network for training, building samples are input into the D-LinkNet network for training, and water surface samples are input into the U-Net network for training to obtain the corresponding cultivated land, building, and water surface extraction models. The images and extraction models of each sub-region are input into the corresponding neural network, the images are predicted, and the predicted intensity maps of cultivated land, buildings and water surfaces are obtained.
  • Raster vectorization and vector post-processing perform vectorization operations on the predicted intensity map.
  • the edge intensity map uses relevant tools to extract skeleton lines, connects the broken lines and then vectorizes them. After the vector lines are simplified and straightened, they are converted into vector surfaces to obtain the pattern spots; the surface predicted intensity map is expanded, hole filled and secondary. After value processing, it is converted into a vector surface result.
  • Vegetation extraction Use cultivated land, buildings and water surfaces as masks, and use terrain lines as constraints to perform multi-scale segmentation of the remaining image range.
  • the forests, shrubs, and grasslands were selected as classification samples by visual discrimination, and the random forest method was used to identify all segmented patches to obtain the forest, grassland, and shrub patches. Perform edge simplification and smoothing on the extraction results to obtain the final vegetation extraction vector result.
  • the cultivated land extracted here is targeted at planting plots with obvious cultivated land boundary characteristics, and cultivated land plots can be clearly identified in different image phases.
  • the RCF edge monitoring network has strong advantages in extracting fine boundary information. There are many and rich internal boundary features in cultivated land plots, and the edge monitoring network can be better used to extract them. Its extraction steps follow the basic steps of deep learning: "sample production - model training - image prediction - raster vectorization - vector post-processing".
  • Buildings are mainly distributed in places where people gather, and most of them appear in clusters. Buildings in the city are extracted according to the specific definition of urban ecosystem in the specification. Although the building has clear shape characteristics, in the image, because the materials of the roof and the ground are relatively similar, some architectural forms are easily damaged, and the specific shape of the building cannot be clearly determined. Similar materials cause the internal features in the building area to be relatively consistent, so internal texture features are used as the extraction conditions. Since the buildings have different shapes and colors, and the ground and building characteristics between the buildings are not very different, Therefore, internal texture is used as a feature for extraction, and the extraction process follows the basic steps of "sample drawing - label production - sample training - image prediction - vector extraction".
  • D-LinkNet has achieved satisfactory results in road extraction. According to tests, it can more accurately obtain the location and morphological information of the required extraction target in the case of artificial surface and complex internal texture features.
  • the water surface of rivers, lakes, ponds and the surrounding non-water surface have obvious transition characteristics, and the boundaries of the water surface can be visually determined with the naked eye.
  • the surface of the non-water surface portion is different, and the boundary characteristics will also be significantly different.
  • the features inside different water surfaces are similar and homogeneous, and are simpler than the boundaries, so the internal texture of the water surface is used as the extraction condition.
  • the extraction steps follow "sample drawing - label production - label training - image prediction - vector extraction".
  • the U-Net network is a classic image segmentation network that can quickly obtain position and morphological information for internally homogeneous water surfaces. Convert the drawn samples into binary image label files according to their attributes, and input them into the U-Net network to train the features and obtain the water surface extraction model. This model is used to predict regional images, and the predicted water surface intensity map of the image is obtained. Use raster vectorization and vector post-processing tools to obtain the final water surface vector extraction results.
  • step S4 for the vector results of each land type extracted hierarchically, the water surface, buildings, and cultivated land are sequentially updated on the forest sketch map spots in order from large scale to small scale using relevant tools.
  • the roads and water systems used as control data are updated to the ecosystem patch results, and the small fragments and narrow patches that exist are processed to obtain the final ecosystem patch vector map.
  • the same type of patches in each sub-task area are merged to obtain the extraction result vector of each ecosystem type in the complete area.
  • the extracted ecological map patches are updated to the same vector file, and then the roads and water systems are updated to the above files.
  • a full-element ecological map patch vector is formed. Convert the updated layer from a single part to a multi-part operation to eliminate the small and narrow spots and obtain the final ecosystem map.
  • Figures 3 and 4 show examples of superimposed grassland ecosystem vector patterns and superimposed farmland ecosystem vector pattern patterns according to an embodiment of the present invention.
  • the ecological map spot extraction method based on high-resolution images of the present invention uses the partition idea to partition the data, which can simplify the complex surface and concentrate elements with similar characteristics into several relatively independent sub-regions, reducing the complexity of ground objects. Difficulty of extraction.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Image Processing (AREA)
  • Processing Or Creating Images (AREA)

Abstract

本发明公开了一种基于高分辨率影像的生态图斑提取方法，其包括如下步骤：获取高分辨率影像，提取路网、水系及地形线数据；以所述路网、水系和地形数据作为提取掩膜，将影像拆分成多个独立的子任务区，用于分区域的生态图斑提取；根据各个子任务区内的影像特征，对地表要素进行分层提取；以及对分层提取出的每种地类的矢量结果，按照由大尺度到小尺度的顺序，将提取的生态图斑更新到同一个矢量文件上，形成全要素的生态图斑矢量，得到最终的生态系统图斑矢量一张图。根据本发明的生态图斑提取方法，可将复杂地表进行简化，将相似特征的要素集中在若干个相对独立的子区域内，降低了地物提取的难度。

Description

基于高分辨率影像的生态图斑提取方法 技术领域
本发明是关于生态学与遥感技术领域，特别是关于一种基于高分辨率遥感影像的“图斑级”生态系统提取方法。
背景技术
生态系统是在自然界的一定空间内，生物与环境构成的统一整体。生态系统的范围可大可小，在不同的尺度上所需要的生态系统的表现形式也是不一致的。目前，在国家尺度、省域尺度和市域尺度，大多以中分辨率影像作为数据基础，通过面向像素的分类方式将地表划分成不同的生态系统类型。这种方式能够快速获取限定尺度条件下的生态系统类型，并且具有较高的准确率。然而，在县域尺度上，由于表现范围的缩小，中分辨率影像表现信息也会大量减少，单个像素内表现的内容复杂多样，从而导致生态系统类型不纯净。高分辨率遥感影像在空间上相比于中低分辨率的影像具有更加丰富的空间信息特征，表现内容也比中分辨率影像多。如果依然采用面向像素的分类方法进行生态系统类型划分，将会导致图斑破碎不连续的情况，在视觉上也不美观。高分影像可以清晰辨别每一个独立的空间地理实体，以具体的地理实体作为对象进行提取可以有效减少图斑破碎的情况。同时，由于每一个图斑表示的对象是唯一的，其形态和边界与影像上的表现内容高度吻合，相比于规则格网，其内部的信息唯一，不存在一个图斑内有多个生态系统类型混合的情况。目前，利用高分辨率遥感影像进行地表类型提取的方式有很多，从传统的面向对象的分类方法和像素级的分类方法，都取得了较好的结果。上述提取方法大多只是针对地表的某一类型进行提取，如建筑的提取、耕地地块的提取。当需要进行多种要素同时提取时，依靠其中某一种方法提取的效果往往无法让人满意，会出现大量的错误提取和漏提的情况。
如何通过高分影像进行生态系统类型的准确提取成为难点。地球作为一个复杂的巨系统，地表类型丰富多样，复杂条件下的地表要素依靠一种方法进行提取是不现实的。如何将复杂地表分解，提高提取结果的准确性，是本领域迄待解决的难题。
发明内容
本发明的目的在于提供一种基于高分辨率影像的生态图斑提取方法，其能够将复杂地表进行简化，将相似特征的要素集中在若干个相对独立的子区域内，降低了地物提取的难度。
为实现上述目的，本发明的实施例提供了一种基于高分辨率影像的生态图斑提取方法，其包括如下步骤：获取高分辨率影像，提取路网、水系及地形线数据；以所述路网、水系和地形数据作为提取掩膜，将影像拆分成多个独立的子任务区，用于分区域的生态图斑提取；根据各个子任务区内的影像特征，对地表要素进行分层提取，所述地表要素包括耕地、建筑、水面、草地、林地和灌丛；以及对分层提取出的每种地类的矢量结果，按照由大尺度到小尺度的顺序，将提取的生态图斑更新到同一个矢量文件上，形成全要素的生态图斑矢量，得到最终的生态系统图斑矢量一张图。
换言之，本发明的主要构思是根据地理学的分层分区思想，将复杂地表按照进行解构，首先利用道路网、水系和地形数据将影像分为若干个相互独立的子任务区，然后根据各个子任务区内部的要素特征，按照建筑、水域、耕地、森林和其他进行分层，根据每一层要素的不同视觉特征，选择对应方法依次进行提取，最后对各层数据进行合并，获得完整区域的生态图斑矢量数据。
在本发明的一个或多个实施方式中，获取高分辨率影像，提取路网、水系及地形线数据，包括：获取目标区域的高分辨率遥感影像；对所述高分辨率遥感影像的原始数据进行图像预处理，所述图像预处理包括正射校正、辐射定标、大气校正和去云；根据影像中的地表特征，绘制其中的主干路网河水系；和计算影像区域的山脊线和山谷线，并构成地形网。
在本发明的一个或多个实施方式中，所述以所述路网、水系和地形数据作为提取掩膜，将影像拆分成多个独立的子任务区，包括：将道路与地形网更新至统一图层文件中，形成完整的分区控制网；和利用分区控制网对影像进行裁剪，形成多个数据量小的子任务区。
在本发明的一个或多个实施方式中，所述根据各个子任务区内的影像特征，对地表要素进行分层提取，包括：样本制作、模型训练与影像预测、栅格矢量化及矢量后处理、和植被提取。
在样本制作中，根据各个子任务区内的影像特征和耕地的类型，选择规则耕地、梯田、坡耕地、建筑区和水面的样本点；按照不同类型的样本点，裁剪成固定大小的影像和对应的矢量框；对裁剪后的样本文件进行特征绘制，并对绘制的矢量赋予非零属性信息；以及将赋予属性后的样本转换成二值图，用于训练。
在模型训练与影像预测中，根据所述样本的类型，将样本输入到对应的卷积神经网络中进行训练，获得对应的耕地、建筑和水面提取模型；以及将各个子区域的影像和提取模型输入到对应的神经网络中，对影像进行预测，获得耕地、建筑和水面的预测强度图。
在栅格矢量化及矢量后处理中，对所述预测强度图进行矢量化操作，其中边缘强度图利用相关工具提取骨架线，对断线进行连接后矢量化，在矢量线简化、拉直后转换成矢量面，获得图斑；面预测强度图则在进行膨胀、孔洞填充及二值化处理后转换为矢量面结果。在植被提取中，获取森林、草地、灌丛图斑，并获得植被提取的矢量结果。
在本发明的一个或多个实施方式中，所述植被提取包括：以耕地、建筑和水面作为掩膜，并以地形线作为约束条件，对剩余影像范围进行多尺度分割；在分割结果的基础上，选取其中的森林、灌丛和草地作为分类样本，采用随机森林方法对所有分割图斑进行判别，获得森林、草地、灌丛图斑；和对提取结果进行边缘简化、平滑处理，获得最终的植被提取矢量结果。
在本发明的一个或多个实施方式中，对分层提取出的每种地类的矢量结果形成全要素的生态图斑矢量包括：对各子任务区的相同类型图斑进行合并，获得完整区域的各生态系统类型提取结果矢量；按照从大尺度到小尺度的更新顺序，将提取的“森林-灌丛-草地-水面-建筑-耕地”生态图斑更新到同一个矢量文件上，然后将道路、水系更新到所述矢量文件中，形成全要素的生态图斑矢量；和对更新后的图层进行单部件转多部件操作，消除其中的细小碎斑和狭长图斑，得到最终的生态系统一张图。
在本发明的一个或多个实施方式中，所述根据所述样本的类型，将样本输入到对应的卷积神经网络中进行训练，获得对应的耕地、建筑和水面提取模型，包括：耕地样本输入到RCF网络中进行训练，建筑样本输入到D-LinkNet网络中进行训练，水面样本输入到U-Net网络中进行训练，获得对应的耕地、建筑和水面提取模型。
在本发明的一个或多个实施方式中，所述耕地样本输入到RCF网络中进行训练，包括：对全域的耕地地块进行总体分析，在影像范围内均匀选择耕地样本，绘制其中的耕地边界信息，并转换为二值结果作为标签数据作为边缘监测网络的训练数据；通过训练边界特征，获得耕地边界信息的提取模型，对经过裁剪的各任务区影像进行预测，获得各任务区的边界预测强度结果图；通过骨架线提取和断线连接，获得耕地预测结果的二值图；和通过栅格矢量化和矢量拉直、平滑等后处理获得耕地提取矢量结果。
在本发明的一个或多个实施方式中，所述建筑样本输入到D-LinkNet网络中进行训练，包括：对各子任务区中的居住地进行分析，将相似特征的居住地作为同一类提取对象，制作不同影像特征的建筑样本，绘制样本中的建筑区域并利用属性值与背景进行区分；将绘制的建筑区样本根据属性值转换为二值标签数据；将建筑物二值标签数据输入D-LinkNet网络进行训练，获得居住地的提取模型；利用该网络对分区域的影像进行预测，获得对应区域的居住地预测强度图；和通过栅格矢量化和矢量后处理操作获得最终的建筑区的矢量提取结果。
在本发明的一个或多个实施方式中，所述水面样本输入到U-Net网络中进行训练，包括：对影像中的河流水面、湖泊水面、坑塘水面的特征进行分析，选择具有代表性的河流、湖泊和坑塘作为样本，对其中的水面进行绘制，并根据属性标注出与背景的区别；将绘制好的样本根据属性转换为二值图标签文件，并输入到U-Net网络中对特征进行训练，获得水面提取模型；利用所述水面提取模型对分区域的影像进行预测，获得影像的水面预测强度图；和采用栅格矢量化和矢量后处理工具获得最终的水面矢量提取结果。
与现有技术相比，根据本发明实施方式的基于高分辨率影像的生态图斑提取方法，具有如下优点：
(1)本发明的基于高分辨率影像的生态图斑提取方法，采用分区思想对数据进行分区，可将复杂地表进行简化，将相似特征的要素集中在若干个相对独立的子区域内，降低了地物提取的难度。同时，一般高分影像的数据量较大，当提取面积较大时，直接进行提取往往费时费力，甚至出现无法提取的状态，并且对硬件的要求也比较高。通过分区处理后，将较大数据量的影像分成若干个小区域后，可有效分散单一影像的数据量，降低提取过程中由于数据量大造成问题的可能性，同时，也可进行多机多任务并行处理，显著缩短 提取的时间。
(2)利用分层思想，将地表划分成具有各自特征的类型，根据每一种类型具体的视觉特征而选择的针对性提取方法，和其他未进行分区分层、直接提取的方法相比，错提、漏提的效果有明显的改善。
(3)矢量后处理的加入，使最终获得的结果具有更加美观的形态；图层更新后为专题图的制作以及后续的应用提供完整的数据支持。
附图说明
图1是根据本发明一实施方式的基于高分辨率影像的生态图斑提取方法的简易流程图；
图2是根据本发明一实施方式的基于高分辨率影像的生态图斑提取方法的详细流程图；
图3是根据本发明一实施方式的叠加的草地生态系统矢量图斑样例；
图4是根据本发明一实施方式的叠加的农田生态系统矢量图斑样例。
具体实施方式
下面结合附图，对本发明的具体实施方式进行详细描述，但应当理解本发明的保护范围并不受具体实施方式的限制。
除非另有其它明确表示，否则在整个说明书和权利要求书中，术语“包括”或其变换如“包含”或“包括有”等等将被理解为包括所陈述的元件或组成部分，而并未排除其它元件或其它组成部分。
如图1至图2所示，根据本发明优选实施方式的基于高分辨率影像的生态图斑提取方法，包括如下步骤S1-S4。
在步骤S1中，获取高分辨率影像，提取路网、水系及地形线数据。
具体而言，可以从公开数据中获取影像区域范围内的主干道路和水系数据，并结合实际的影像特征，利用GIS件补充绘制缺失的道路和水系，形成影像区域的水系和道路路网数据。利用数字高程模型(DEM)对影像范围内的地形线进行提取，根据路网、水系和地 形数据，合成影像范围内的控制网数据，用于后续的影像拆分。
在一实施方式中，步骤S1可以包括：1)获取目标区域高分辨率遥感影像，如谷歌影像、GF2、QuickBird等；2)对原始数据进行正射校正、辐射定标、大气校正、去云等图像预处理；3)根据影像中的地表特征，绘制其中的主干路网河水系；4)利用DEM计算影像区域的山脊线、山谷线，并构成地形网；5)将道路与地形网更新至统一图层文件中，形成完整的分区控制网；6)利用分区控制网对影像进行裁剪，形成若干数据量小的子任务区。
在步骤S2中，以所述路网、水系和地形数据作为提取掩膜，将影像拆分成多个独立的子任务区(多个block)，用于分区域的生态图斑提取。
具体而言，步骤S2中，可以将道路与地形网更新至统一图层文件中，形成完整的分区控制网；并利用分区控制网对影像进行裁剪，形成多个数据量小的子任务区。
在步骤S3中，根据各个子任务区内的影像特征，对地表要素进行分层提取，所述地表要素包括耕地、建筑、水面、草地、林地和灌丛。
具体而言，本发明的生态图斑提取方法中，按照提取的难易程度，将地表要素划分成众多不同的类型，依次进行提取。该分层提取主要包括以下步骤：
(1)样本制作：根据各个子区域内的影像特征，根据耕地的类型，选择规则耕地、梯田和坡耕地、建筑区和水面的样本点；样本需均匀分布且数量适中。按照不同类型的样本点，裁剪成固定大小的影像和对应的矢量框。其中，规则耕地、梯田为线矢量文件，坡耕地、建筑和水面为面矢量文件。对裁剪后的样本文件进行特征绘制，规则耕地和梯田绘制地块中的边缘特征信息，坡耕地、建筑和水面绘制内部纹理特征信息的范围，并对绘制的矢量赋予非零属性信息。最后将赋予属性后的样本转换成二值图，用于训练。
(2)模型训练与影像预测：根据样本的类型，将样本输入到对应的卷积神经网络中进行训练。其中，耕地样本输入到RCF网络中进行训练，建筑样本输入到D-LinkNet网络中进行训练，水面样本输入到U-Net网络中进行训练，获得对应的耕地、建筑和水面提取模型。将各个子区域的影像和提取模型输入到对应的神经网络中，对影像进行预测，获得耕地、建筑和水面的预测强度图。
(3)栅格矢量化及矢量后处理：对预测强度图进行矢量化操作。其中边缘强度图利用相关工具提取骨架线，对断线进行连接后矢量化，在矢量线简化、拉直后转换成矢量面，获得图斑；面预测强度图则在进行膨胀、孔洞填充及二值化处理后转换为矢量面结果。
(4)植被提取：以耕地、建筑和水面作为掩膜，并以地形线作为约束条件，对剩余影像范围进行多尺度分割。在分割结果的基础上，采用目视判别的方式选取其中的森林、灌丛和草地作为分类样本，采用随机森林方法对所有分割图斑进行判别，获得森林、草地、灌丛图斑。对提取结果进行边缘简化、平滑处理，获得最终的植被提取矢量结果。
下面按照从小到大的尺度，对子区域影像中依次提取耕地、建筑、水面、草地、林地、灌丛进行详细介绍。
耕地地块提取
此处提取的耕地以具有明显耕地边界特征的种植地块为对象，并且在不同的影像时相上，均能够清晰辨别的耕地地块。RCF边缘监测网络在细小的边界信息的提取上具有较强的优势，耕地地块内部边界特征多且丰富，可以较好的使用边缘监测网络对其进行提取。其提取步骤遵循深度学习的基本步骤：“样本制作-模型训练-影像预测-栅格矢量化-矢量后处理”。
首先对全域的耕地地块进行总体分析，在影像范围内均匀选择耕地样本，利用GIS软件绘制其中的耕地边界信息，并转换为二值结果作为标签数据作为边缘监测网络的训练数据。通过训练边界特征，获得耕地边界信息的提取模型，对经过裁剪的各任务区影像进行预测，获得各任务区的边界预测强度结果图，在此基础上，通过骨架线提取和断线连接，获得耕地预测结果的二值图，最后通过栅格矢量化和矢量拉直、平滑等后处理获得耕地提取矢量结果。
建筑
建筑主要分布在人口聚集的地方，大多成片出现。根据规范中的城镇生态系统的具体定义，对城市中的建筑物进行提取。尽管建筑物具有明确的形状特征，但在影像上由于屋顶与地面的材质较为相似，因此部分建筑形态容易受到破坏，无法明确建筑物的具体形状。而相近的材质，造成建筑区内的内部特征较为一致，因此采用内部纹理特征作为提取条件， 由于建筑物具有不同的形状和颜色，且各建筑之间的地面与建筑的特征差异并不大，因此采用内部纹理作为特征进行提取，提取流程遵循“样本绘制-标签制作-样本训练-影像预测-矢量提取”基本步骤。
首先对各子任务区中的居住地进行分析，将相似特征的居住地作为同一类提取对象，制作不同影像特征的建筑样本，绘制样本中的建筑区域并利用属性值与背景进行区分。将绘制的建筑区样本根据属性值转换为二值标签数据。D-LinkNet在道路提取上获得令人满意的效果，根据测试，其在人工地表及复杂内部纹理特征的情况下，能够较为准确获取所需提取目标的位置及形态信息。将建筑物二值标签数据输入D-LinkNet网络进行训练，获得居住地的提取模型；利用该网络对分区域的影像进行预测，获得对应区域的居住地预测强度图；最后通过栅格矢量化和矢量后处理操作获得最终的建筑区的矢量提取结果。
水面
在高分遥感影像上，河流水面、湖泊水面、坑塘水面与其周围非水面地表具有明显的过渡特征，可通过肉眼直观判别水面的边界。但非水面部分的地表各不相同，边界特征也会存在明显差异。不同水面内部的特征相似且均质，相比边界较为简单，因此采用水面内部纹理作为提取条件。提取步骤遵循“样本绘制-标签制作-标签训练-影像预测-矢量提取”。
首先对影像中的河流水面、湖泊水面、坑塘水面的特征进行分析，选择具有代表性的河流、湖泊和坑塘作为样本，对其中的水面进行绘制，并根据属性标注出与背景的区别。U-Net网络为经典的图像分割网络，对内部较均质的水面能够快速获取其位置及形态信息。将绘制好的样本根据属性转换为二值图标签文件，并输入到U-Net网络中对特征进行训练，获得水面提取模型。利用该模型对分区域的影像进行预测，获得影像的水面预测强度图。采用栅格矢量化和矢量后处理工具获得最终的水面矢量提取结果。
灌林草
在完成上述地表类型的提取之后，以提取结果作为掩膜，对影像中剩余区域进行多尺度分割，获得初步的灌丛、草地、湿地生态系统图斑。在图斑的基础上，根据森林、灌丛、草地在植被、土壤等特性上的差异，构建决策树分类体系，进一步确定分割图斑的生态系统类型。
林地与草地作为自然地表，与其他半自然和人工地表存在显著差异。在耕地、建筑、水面提取完成后，将上述要素作为掩膜，对掩膜之外的未提取区域进行多尺度分割，将自然植被分割成更小尺度图斑，结合归一化植被指数(NDVI)、植被覆盖度(FVI)和土壤侵蚀性参数，采用随机森林分类方法对分割后的图斑进行分类，获得森林、灌丛和草地的分类结果。
根据遥感影像上森林、灌丛、草地在光谱特征上的差异，对应以上三种植被类型，分别选择具有典型代表性的区域作为分类样本，并利用DEM、植被指数作为辅助因子构建随机森林分类器，对分割的图斑进行分类，得到分类结果的置信度，对于低置信度的图斑，对照影像上的特征，修正该图斑的属性，并添加到份分类样本中进行迭代分类，不断优化灌丛、森林、草地的分类结果。
在步骤S4中，对分层提取出的每种地类的矢量结果，按照由大尺度到小尺度的顺序，将水面、建筑、耕地使用相关工具按顺序依次更新到林草图斑上，最后将作为控制数据的道路和水系更新到生态系统图斑结果上，并对其中存在的细小碎斑和狭长图斑进行处理，得到最终的生态系统图斑矢量一张图。
具体而言，对各子任务区的相同类型图斑进行合并，获得完整区域的各生态系统类型提取结果矢量。按照从大尺度到小尺度的更新顺序(森林-灌丛-草地-水面-建筑-耕地)，将提取的生态图斑更新到同一个矢量文件上，然后将道路、水系更新到上述文件中，形成全要素的生态图斑矢量。对更新后的图层进行单部件转多部件操作，消除其中的细小碎斑和狭长图斑，得到最终的生态系统一张图。
图3和图4显示了根据本发明一实施方式的叠加的草地生态系统矢量图斑样例和叠加的农田生态系统矢量图斑样例。
本发明的基于高分辨率影像的生态图斑提取方法，采用分区思想对数据进行分区，可将复杂地表进行简化，将相似特征的要素集中在若干个相对独立的子区域内，降低了地物提取的难度。
前述对本发明的具体示例性实施方案的描述是为了说明和例证的目的。这些描述并非想将本发明限定为所公开的精确形式，并且很显然，根据上述教导，可以进行很多改变和 变化。对示例性实施例进行选择和描述的目的在于解释本发明的特定原理及其实际应用，从而使得本领域的技术人员能够实现并利用本发明的各种不同的示例性实施方案以及各种不同的选择和改变。本发明的范围意在由权利要求书及其等同形式所限定。

Claims (10)

  1. 一种基于高分辨率影像的生态图斑提取方法，其特征在于，包括如下步骤：
    S1：获取高分辨率影像，提取路网、水系及地形线数据；
    S2：以所述路网、水系和地形数据作为提取掩膜，将影像拆分成多个独立的子任务区，用于分区域的生态图斑提取；
    S3：根据各个子任务区内的影像特征，对地表要素进行分层提取，所述地表要素包括耕地、建筑、水面、草地、林地和灌丛；以及
    S4：对分层提取出的每种地类的矢量结果，按照由大尺度到小尺度的顺序，将提取的生态图斑更新到同一个矢量文件上，形成全要素的生态图斑矢量，得到最终的生态系统图斑矢量一张图。
  2. 如权利要求1所述的基于高分辨率影像的生态图斑提取方法，其特征在于，所述步骤S1包括：
    获取目标区域的高分辨率遥感影像；
    对所述高分辨率遥感影像的原始数据进行图像预处理，所述图像预处理包括正射校正、辐射定标、大气校正和去云；
    根据影像中的地表特征，绘制其中的主干路网河水系；和
    计算影像区域的山脊线和山谷线，并构成地形网。
  3. 如权利要求1所述的基于高分辨率影像的生态图斑提取方法，其特征在于，所述步骤S2包括：
    将道路与地形网更新至统一图层文件中，形成完整的分区控制网；和
    利用分区控制网对影像进行裁剪，形成多个数据量小的子任务区。
  4. 如权利要求1所述的基于高分辨率影像的生态图斑提取方法，其特征在于，所述步骤S3包括：
    样本制作：根据各个子任务区内的影像特征和耕地的类型，选择规则耕地、梯田、坡耕地、建筑区和水面的样本点；按照不同类型的样本点，裁剪成固定大小的影像和对应的矢量框；对裁剪后的样本文件进行特征绘制，并对绘制的矢量赋予非零属性信息；以及将赋予属性后的样本转换成二值图，用于训练；
    模型训练与影像预测：根据所述样本的类型，将样本输入到对应的卷积神经网络中进行训练，获得对应的耕地、建筑和水面提取模型；以及将各个子区域的影像和提取模型输入到对应的神经网络中，对影像进行预测，获得耕地、建筑和水面的预测强度图；
    栅格矢量化及矢量后处理：对所述预测强度图进行矢量化操作，其中边缘强度图利用相关工具提取骨架线，对断线进行连接后矢量化，在矢量线简化、拉直后转换成矢量面，获得图斑；面预测强度图则在进行膨胀、孔洞填充及二值化处理后转换为矢量面结果；以及
    植被提取：获取森林、草地、灌丛图斑，并获得植被提取的矢量结果。
  5. 如权利要求4所述的基于高分辨率影像的生态图斑提取方法，其特征在于，所述植被提取包括：
    以耕地、建筑和水面作为掩膜，并以地形线作为约束条件，对剩余影像范围进行多尺度分割；
    在分割结果的基础上，选取其中的森林、灌丛和草地作为分类样本，采用随机森林方法对所有分割图斑进行判别，获得森林、草地、灌丛图斑；和
    对提取结果进行边缘简化、平滑处理，获得最终的植被提取矢量结果。
  6. 如权利要求1所述的基于高分辨率影像的生态图斑提取方法，其特征在于，所述步骤S4包括：
    对各子任务区的相同类型图斑进行合并，获得完整区域的各生态系统类型提取结果矢量；
    按照从大尺度到小尺度的更新顺序，将提取的“森林-灌丛-草地-水面-建筑-耕地”生态图斑更新到同一个矢量文件上，然后将道路、水系更新到所述矢量文件中，形成全要素 的生态图斑矢量；和
    对更新后的图层进行单部件转多部件操作，消除其中的细小碎斑和狭长图斑，得到最终的生态系统一张图。
  7. 如权利要求4所述的基于高分辨率影像的生态图斑提取方法，其特征在于，所述根据所述样本的类型，将样本输入到对应的卷积神经网络中进行训练，获得对应的耕地、建筑和水面提取模型，包括：耕地样本输入到RCF网络中进行训练，建筑样本输入到D-LinkNet网络中进行训练，水面样本输入到U-Net网络中进行训练，获得对应的耕地、建筑和水面提取模型。
  8. 如权利要求7所述的基于高分辨率影像的生态图斑提取方法，其特征在于，所述耕地样本输入到RCF网络中进行训练，包括：
    对全域的耕地地块进行总体分析，在影像范围内均匀选择耕地样本，绘制其中的耕地边界信息，并转换为二值结果作为标签数据作为边缘监测网络的训练数据；
    通过训练边界特征，获得耕地边界信息的提取模型，对经过裁剪的各任务区影像进行预测，获得各任务区的边界预测强度结果图；
    通过骨架线提取和断线连接，获得耕地预测结果的二值图；和
    通过栅格矢量化和矢量拉直、平滑等后处理获得耕地提取矢量结果。
  9. 如权利要求7所述的基于高分辨率影像的生态图斑提取方法，其特征在于，所述建筑样本输入到D-LinkNet网络中进行训练，包括：
    对各子任务区中的居住地进行分析，将相似特征的居住地作为同一类提取对象，制作不同影像特征的建筑样本，绘制样本中的建筑区域并利用属性值与背景进行区分；
    将绘制的建筑区样本根据属性值转换为二值标签数据；
    将建筑物二值标签数据输入D-LinkNet网络进行训练，获得居住地的提取模型；利用该网络对分区域的影像进行预测，获得对应区域的居住地预测强度图；和
    通过栅格矢量化和矢量后处理操作获得最终的建筑区的矢量提取结果。
  10. 如权利要求7所述的基于高分辨率影像的生态图斑提取方法，其特征在于，所述水面样本输入到U-Net网络中进行训练，包括：
    对影像中的河流水面、湖泊水面、坑塘水面的特征进行分析，选择具有代表性的河流、湖泊和坑塘作为样本，对其中的水面进行绘制，并根据属性标注出与背景的区别；
    将绘制好的样本根据属性转换为二值图标签文件，并输入到U-Net网络中对特征进行训练，获得水面提取模型；
    利用所述水面提取模型对分区域的影像进行预测，获得影像的水面预测强度图；和
    采用栅格矢量化和矢量后处理工具获得最终的水面矢量提取结果。
PCT/CN2022/107698 2022-07-25 2022-07-25 基于高分辨率影像的生态图斑提取方法 WO2024020744A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2022/107698 WO2024020744A1 (zh) 2022-07-25 2022-07-25 基于高分辨率影像的生态图斑提取方法

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2022/107698 WO2024020744A1 (zh) 2022-07-25 2022-07-25 基于高分辨率影像的生态图斑提取方法

Publications (1)

Publication Number Publication Date
WO2024020744A1 true WO2024020744A1 (zh) 2024-02-01

Family

ID=89704890

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/107698 WO2024020744A1 (zh) 2022-07-25 2022-07-25 基于高分辨率影像的生态图斑提取方法

Country Status (1)

Country Link
WO (1) WO2024020744A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117953430A (zh) * 2024-03-15 2024-04-30 湖南省第二测绘院 一种通信铁塔视频实时监测耕地破坏的方法和系统

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108898096A (zh) * 2018-06-27 2018-11-27 重庆交通大学 一种面向高分影像的信息快速精确提取方法
US20200065968A1 (en) * 2018-08-24 2020-02-27 Ordnance Survey Limited Joint Deep Learning for Land Cover and Land Use Classification
CN114663771A (zh) * 2022-04-14 2022-06-24 中国农业科学院农业资源与农业区划研究所 一种基于分区分层理论的山区耕地智能化提取方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108898096A (zh) * 2018-06-27 2018-11-27 重庆交通大学 一种面向高分影像的信息快速精确提取方法
US20200065968A1 (en) * 2018-08-24 2020-02-27 Ordnance Survey Limited Joint Deep Learning for Land Cover and Land Use Classification
CN114663771A (zh) * 2022-04-14 2022-06-24 中国农业科学院农业资源与农业区划研究所 一种基于分区分层理论的山区耕地智能化提取方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
JIANCHENG LUO, TIANJUN W U, ZHIFENG W U, ZHOU YA'NAN, LIJING GAO, YINGWEI SUN, WEI W U, YINGPIN YANG, XIAODONG H U, XIN ZHANG, ZHA: "Methods of Intelligent Computation and Pattern Mining based on Geo-parcels", JOURNAL OF GEO-INFORMATION SCIENCE, vol. 22, no. 1, 25 January 2020 (2020-01-25), pages 57 - 75, XP093133970, DOI: 10.12082/dqxxkx.2020.190462 *
张海军 (ZHANG, HAIJUN): "全球地表覆盖分类实验研究-以陕西省为例 (Non-official translation: Experimental Study on Global Land Cover Classification-Taking Shaanxi Province as an Example)", 中国优秀硕士学位论文全文数据库 基础科学辑 (BASIC SCIENCES, CHINA MASTER'S THESES FULL-TEXT DATABASE), no. 04, 15 April 2014 (2014-04-15), ISSN: 1674-0246 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117953430A (zh) * 2024-03-15 2024-04-30 湖南省第二测绘院 一种通信铁塔视频实时监测耕地破坏的方法和系统

Similar Documents

Publication Publication Date Title
CN111986099B (zh) 基于融合残差修正的卷积神经网络的耕地监测方法及系统
Gamanya et al. Object-oriented change detection for the city of Harare, Zimbabwe
CN108830870B (zh) 基于多尺度结构学习的卫星影像高精度农田边界提取方法
Morton et al. Final report for LCM2007-the new UK land cover map. Countryside survey technical report no 11/07
Trier et al. Automatic detection of mound structures in airborne laser scanning data
KR102053582B1 (ko) 딥러닝 기반 영상패턴학습을 이용한 토지피복 분류 방법
CN111028255A (zh) 基于先验信息与深度学习的农田区域预筛选方法及装置
Li et al. Breaking the resolution barrier: A low-to-high network for large-scale high-resolution land-cover mapping using low-resolution labels
CN113223042B (zh) 一种遥感影像深度学习样本智能采集方法及设备
CN115187866A (zh) 基于高分辨率影像的生态图斑提取方法
US20210027055A1 (en) Methods and Systems for Identifying Topographic Features
CN104239890A (zh) 高分1号卫星的沿岸陆地土地覆被信息自动提取的方法
Dehvari et al. Comparison of object-based and pixel based infrared airborne image classification methods using DEM thematic layer
WO2024020744A1 (zh) 基于高分辨率影像的生态图斑提取方法
Cai et al. Improving agricultural field parcel delineation with a dual branch spatiotemporal fusion network by integrating multimodal satellite data
CN117475314B (zh) 一种地质灾害隐患立体识别方法、系统及介质
O’Neil-Dunne et al. Incorporating contextual information into object-based image analysis workflows
CN116415318B (zh) 一种基于数学形态学的内流区湖泊水文连通性建模方法
Tejeswari et al. Building footprint extraction from space-borne imagery using deep neural networks
CN115223000B (zh) 面向耕地资源监测的深度学习样本制作方法
CN111461992A (zh) 基于深度学习的超高分辨率遥感影像茶叶分布自动提取方法及装置
Forghani et al. Object-based classification of multi-sensor optical imagery to generate terrain surface roughness information for input to wind risk simulation
Li et al. Multi-scale object-oriented building extraction method of Tai'an city from high resolution image
Lehrbass et al. Techniques for object-based classification of urban tree cover from high-resolution multispectral imagery
Cao et al. An object-oriented approach to map wetland vegetation: a case study of yellow river delta

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22952211

Country of ref document: EP

Kind code of ref document: A1