Large-scale wetland data automatic extraction method based on cloud computing

Abstract

The invention discloses a large-scale wetland data automatic extraction method based on cloud computing, and belongs to the technical field of remote sensing. Through a cloud computing platform, a regular hexagon segmentation mechanism is introduced, a simple non-iterative clustering algorithm and a random forest algorithm are taken as basic models, the relationship between pixels and the change mode of the region are autonomously learned, and the characteristics of wetland submerging dynamics, phenological characteristics, spectral textures, geographic space distribution and the like are comprehensively considered. And large-scale wetland data are automatically extracted. According to the invention, wetland extraction of large-scale remote sensing images is realized, and cloud computing characteristics and multi-source data integration capability of a cloud computing platform are fully utilized.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing technology and relates to a large-scale wetland data automatic extraction method based on cloud computing. Background Art

[0002] Wetlands are special ecosystems located at the junction of water and land, and they play an important role in biodiversity conservation, water resource protection and utilization, and disaster prevention. There are a large number of complex wetland connectivity networks in the subtropical region, which are characterized by large distribution scales and multiple types. With the impact of global climate change, the protection of wetlands has become increasingly important. Extracting large-scale wetland data and obtaining accurate wetland location and coverage data are important prerequisites for monitoring and protecting wetlands. Traditional survey methods have limitations in time and accuracy. Among them, the specificity of wetland regional morphology, the complexity of types, and the wide distribution range are the difficulties in wetland data extraction and the key points that need to be solved. Summary of the invention

[0003] In view of the problems in the prior art, the present invention provides a method for automatically extracting large-scale wetland data based on cloud computing, which relates to a method for automatically extracting large-scale regional wetland data through cloud computing in wetland remote sensing.

[0004] A large-scale wetland data automatic extraction method based on cloud computing, through the cloud computing platform, introduces the regular hexagon segmentation mechanism, takes the simple non-iterative clustering algorithm and the random forest algorithm as the basic model, autonomously learns the relationship between pixels and the change pattern of the region itself, comprehensively considers the wetland flooding dynamics, phenological characteristics, spectral texture and geographic spatial distribution characteristics, and automatically extracts large-scale wetland data, which includes the following steps: step 1, separation of inland offshore and coastal wetlands, step 2, water body wetland segmentation data extraction.

[0005] The advantages of the present invention are: using a cloud computing platform to extract and map large-scale wetland information, the implementation method is simple and effective, and the characteristics of cloud computing suitable for large-scale remote sensing image processing are fully utilized. Advanced SN IC algorithms and random forest algorithms are used to efficiently and accurately extract wetland data, and the wetland extraction effect is optimized by combining wetland phenological characteristics and morphological laws. The extracted wetland data can accurately distinguish different wetland type information, which meets the classification accuracy requirements of wetland research. The dynamic change process of the wetland can be intuitively displayed, and data such as wetland type and spatial distribution information can be quickly obtained. The analysis code and results can be quickly shared, which is convenient for cross-institutional and cross-regional wetland research and cooperation, and provides a scientific basis for the formulation of wetland ecosystem health assessment and restoration management strategies. It is of great significance to monitor wetland degradation and formulate wetland restoration and protection strategies.

[0006] The present invention realizes wetland extraction from large-scale remote sensing images in a simple and effective way, making full use of the cloud computing characteristics and multi-source data integration capabilities of the cloud computing platform. Combined with SN IC superpixel segmentation and random forest classifier, the phenological characteristics and morphological laws of wetlands are used to extract wetlands efficiently and accurately, and the classification results are adapted to the actual application needs of wetland protection and management. The final data extraction and mapping results not only meet the accuracy requirements of wetland scientific research, but also provide strong data support, providing an important reference for the protection, management and restoration of wetland ecosystems. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. As shown in the figure:

[0008] Figure 1 It is the overall technical flow chart of the present invention.

[0009] Figure 2 This is a diagram of tidal height data of the present invention.

[0010] Figure 3 This is the second graph of the tidal height data of the present invention.

[0011] Figure 4 It is a river flooding frequency diagram of the river wetland of the present invention during the hydrological period.

[0012] Figure 5 It is a lake flooding frequency diagram of the river wetland of the present invention during the hydrological cycle.

[0013] Figure 6 This is the river wetland extraction result diagram of the present invention.

[0014] Figure 7 This is the result of the Yellow River Basin wetland data extraction of the present invention. DETAILED DESCRIPTION

[0015] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] Example 1: Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 As shown in the figure, a large-scale wetland data automatic extraction method based on cloud computing is used to extract large-scale wetland data, which makes up for the current problem of low efficiency and poor precision of large-scale wetland data extraction, provides technical support for wetland monitoring and protection, and provides more accurate basic data for wetland ecosystem research.

[0017] The invention mainly includes the separation of inland coastal wetlands and the segmentation and extraction of water body wetlands, such as Figure 1 As shown:

[0018] The first step is data preparation, which involves calling the Sentinel 2 satellite remote sensing dataset, the Shuttle Radar Topography Mission (SRTM) elevation dataset (1 arc second), and the Global Advanced Land Observing Satellite (Globa LOS) terrain dataset on the GEE cloud platform.

[0019] The second step is to conduct dynamic tidal monitoring in offshore and coastal areas, build a coastline buffer zone, monitor the dynamic changes of tides based on the full time series NDVI index, conduct seawater detection, form high tide images and low tide images, perform difference calculations, and separate the land and sea boundaries.

[0020] The third step is to perform multi-feature mask processing on non-wet areas in the land area, in which the slope mask, mountain shadow mask and bare land mask are differentially calculated to form a non-wetland impermeable surface mask, covering all non-wet areas to the greatest extent, eliminating the interference of non-wet areas on wetland classification, and completing the separation of inland coastal wetlands.

[0021] The fourth step is phenological classification, which involves identifying vegetation phenological characteristics and water body phenological characteristics respectively, and then using the random forest algorithm to call the training samples for classification to form a phenological image, and forming the phenological frequencies of different wetland types based on difference analysis; the fifth step is feature segmentation, which involves inputting NDWI and training samples based on the wetland shape characteristics, calling the SN IC segmentation algorithm, calculating the flooding frequency of different wetlands through the flooding frequency index, and completing the extraction of water body wetland segmentation information. Finally, according to the above wetland classification steps, the information extraction of five types of wetlands, including swamp wetlands, offshore and coastal wetlands, lake wetlands, river wetlands, and artificial wetlands, is completed.

[0022] The specific steps are as follows:

[0023] Step 1: Separation of inland offshore and coastal wetlands.

[0024] Step 1.1 Tidal dynamic monitoring.

[0025] Since wetlands are located at the junction of land and water, with the ebb and flow of the tide, inland wetlands and nearshore and coastal wetland areas will alternate at high and low tides.

[0026] Tidal flats are defined as mudflats and sandy beaches that are exposed at low tide but submerged at high tide. This definition is used to distinguish inland and offshore from coastal wetlands. Seawater pixels at high tide are detected as sea surface, and mudflat pixels or sandy beaches at low tide are detected as tidal flats. According to the tidal changes, a coastline buffer zone is constructed, buffering the coastline from the inland and sea surface by 2 km respectively, expanding the detection range of the tidal flat.

[0027] In the GEE cloud platform, all Sentinel 2 images in the study area and the study period are used to establish the statistical relationship between the spectral index and the tidal height. The normalized water index NDWI of each part is calculated from the Sentinel 2 images of the entire tidal cycle to obtain the NDWI value of the entire time series. The maximum sea surface range is obtained with the maximum NDWI value, and the minimum sea surface range is obtained with the minimum NDWI value. The time point image is used to detect seawater. The maximum and minimum water surfaces of the images that can be obtained in the area in the year are filtered out through high tide images and low tide images, respectively, and the tidal flat wetlands in the offshore and coastal wetlands are divided by difference calculation. The land and sea boundaries are separated.

[0028] Step 1.2 Multi-feature mask of non-wetland area.

[0029] For non-wetland areas, we first selected areas with a slope greater than 5° through the slope mask. Then we used the threshold segmentation DEM and the normalized vegetation index NDVI to perform a mountain shadow mask to remove the mountain forest vegetation coverage area and mountain shadows. Finally, we used the normalized building index NDB I, NDVI, and NDWI in the filter to perform a bare ground mask, and calculated the impervious surface mask by difference.

[0030] At this point, the process of separating inland offshore and coastal wetlands is completed, and further water wetland segmentation and data extraction are carried out based on this.

[0031] Step 2: Extract water body wetland segmentation data.

[0032] Step 2.1 Phenological classification.

[0033] Because the swamp surface is in a wet state for a long time, it has unique vegetation and soil formation processes. There are characteristic differences between different phenological periods, so different types of wetlands can be distinguished by phenological classification.

[0034] After masking out all classified wetland extents and non-wetland extents containing disturbance data, the RF algorithm was used to construct a phenological classifier to draw mixed wetland maps of vegetation and water bodies.

[0035] The conditions for identifying phenological characteristics are:

[0036] (1) Enhanced Vegetation Index (EVI): EVI can improve the spectral detection sensitivity of chlorophyll concentration and biomass, and significantly compensate for the effects of atmosphere, water, shadow and soil background using aerosol drag coefficients (C1, C2), canopy background adjustment (L) and gain factor (G).

[0037] (2) The ratio vegetation index (RVI) uses the near-infrared band and the red band to directly infer vegetation growth data from the reflectance of vegetation leaves.

[0038] (3) Capture the phenological dynamics of wetland vegetation through NDVI, NDWI, EVI, and RVI indices, and identify the phenological characteristics of vegetation and water bodies.

[0039] In GEE, the time series of Sentinel 2 images, EVI, NDVI, NDWI, and RVI were connected to the raster collection, and three types of training samples were redistributed, including lake wetlands, swamp wetlands, and non-wetlands. The phenological image was obtained using the random forest algorithm. Finally, the phenological frequency was obtained by calculating the difference between the NDVI, NDWI, EVI, and RVI indices at different phenological periods.

[0040] Step 2.2 Feature segmentation.

[0041] Since different types of water surfaces have unique shape characteristics, wetland types can be distinguished to a certain extent based on their shapes. Among them, artificial water bodies (ponds, etc.) are usually approximately rectangular, reservoirs usually have complex geometric shapes, seasonal river wetlands have longer rivers than other types of rivers, and seasonal lake wetlands have natural polygonal boundaries.

[0042] The shape features and NDWI are combined to train the training samples through the SN IC algorithm. K initial centroids are generated in the image plane using a regular grid, and then K corresponding elements are created. Each element includes spatial position, CI ELab color, pixel label, and the distance from the pixel centroid to the candidate pixel. The algorithm quantifies the distance metric through a 5-dimensional feature vector to achieve local clustering of image pixels.

[0043] The flooding frequency (IF) is the frequency at which a single pixel is identified as a water body within a year. Different water bodies in each wetland type have similar spectral characteristics. The spatial resolution of the green and near-infrared bands used to calculate NDWI using Sentinel 2 images is 10m, which is conducive to better distinguishing more detailed wetland types such as aquaculture ponds and canals. The IF calculation formula is:

[0044] IF=f_water / f(1)

[0045] Where f_water is the number of times it is identified as a water body, and f is the number of valid observations of the pixel.

[0046] Through the calculation of the above process in the GEE cloud platform, we can segment and extract data of five major types of wetlands: lake wetlands, river wetlands, swamp wetlands, offshore and coastal wetlands, and artificial wetlands.

[0047] This completes the entire process of automatic extraction of large-scale wetland data based on cloud computing.

[0048] Example 2: Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 As shown, a large-scale wetland data automatic extraction method based on cloud computing and an example of large-scale wetland data automatic extraction based on cloud computing in the Yellow River Basin.

[0049] Step 1: Data Collection

[0050] The example area is the Yellow River Basin, where a large number of wetlands are distributed and form a network structure. GEE is used to retrieve continuous Sentinel 2 images, SRTM elevation data, Globa LALOS terrain data, and tidal height data for the Yellow River Basin throughout the year in 2020.

[0051] Step 2: Separation of inland coastal wetlands and coastal wetlands

[0052] First, the Yellow River Delta coastline was buffered by 20 km. All Sentinel 2 images covering the coastal boundary of the Yellow River Delta region in 2020 were used to establish the statistical relationship between the spectral index and tidal height. The NDWI mean of the full time series of the Yellow River basin was calculated using the threshold method ( Figure 2 and Figure 3 ), the maximum and minimum water surfaces of the images available in the year in the area were filtered out respectively, and the tidal flat wetlands in the offshore and coastal wetlands were divided by difference calculation.

[0053] Slope distribution was analyzed based on elevation data, and areas with slopes greater than 5° were extracted and masked. The DEM was segmented using thresholds and the Normalized Difference Vegetation Index (NDVI) was used to remove mountain forest vegetation coverage and mountain shadows. The forest coverage areas with NDVI>0.3 in areas above 300m were then screened using altitude classification. Finally, bare land, agricultural land, buildings, roads and other impermeable surfaces were covered by masks. The potential wetland distribution area in the Yellow River Basin was obtained.

[0054] Step 3: Extract water body wetland segmentation data

[0055] After masking out all classified wetland extents and non-wetland extents containing disturbance data, a phenological classifier was constructed using the random forest algorithm to draw mixed wetland maps of vegetation and water bodies. Correlation analysis of marsh wetland sample points with NDVI revealed that herbaceous marshes, shrub marshes, and other wetland cover types with obvious phenological rhythms showed significant differences in time series. The time series of Sentinel 2 images, EVI, NDVI, NDWI, and RVI were connected to the raster collection, and the three types of training samples were redistributed, including lake wetlands, marsh wetlands, and non-wetlands.

[0056] In order to further refine the classification, the annual maximum water surface coverage is first extracted according to the flooding frequency, and then the SN IC object segmentation algorithm is used in combination with the flooding frequency difference to refine the extraction of wetlands covered by water surface. In this part, the main types of wetlands covered by water bodies are detected: lakes, rivers, reservoirs / breeding ponds / salt fields, canals / channels and estuarine waters. In addition, although agricultural ponds, breeding ponds and rice fields have similar shape characteristics. However, agricultural ponds are usually covered by water bodies throughout the year; rice fields in the Yellow River Basin are only covered by water bodies at a specific time of the year (single-season rice), and breeding ponds are cleared and drained in the rainy season or dry season, and the dry period is often 7-20 days. Estuarine waters are greatly affected by the frequency of flooding, and there is a significant water level drop between the rainy season and the dry season. Therefore, based on data such as flooding frequency, artificial waters can be further divided into agricultural ponds and rice fields, and natural waters (lakes and rivers) can be divided into rivers, lakes and estuarine waters. Taking river wetlands as an example, the flooding frequency is as follows: Figure 4 and Figure 5 As can be seen, river wetland extraction Figure 6 shown.

[0057] Step 4. Finally, five types of wetland data are extracted, including swamp wetlands, lake wetlands, river wetlands, offshore and coastal wetlands, and artificial wetlands in the Yellow River basin.

[0058] from Figure 7 It can be seen from the results of wetland data extraction in the Yellow River Basin that better results can be obtained through cloud computing for large-scale wetland data. The extraction of wetland data can meet the application needs of research related to wetland spatial distribution and wetland management and protection.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A large-scale wetland data automatic extraction method based on cloud computing, characterized in that: Through the cloud computing platform, the regular hexagon segmentation mechanism is introduced, and the simple non-iterative clustering algorithm and random forest algorithm are used as the basic model to autonomously learn the relationship between pixels and the change pattern of the region itself, comprehensively consider the wetland flooding dynamics, phenological characteristics, spectral texture and geographic spatial distribution characteristics, and automatically extract large-scale wetland data. The following steps are included: Step 1: Separation of inland offshore and coastal wetlands. Step 2: Extract water body wetland segmentation data.

2. According to the cloud computing-based large-scale wetland data automatic extraction method of claim 1, it is characterized in that: Step 1: Separation of inland offshore and coastal wetlands, including the following steps: Step 1.1: Dynamic tidal monitoring. All Sentinel 2 images of the study area and the study period were used in the Google Earth Engine (GEE) cloud platform to establish the statistical relationship between the spectral index and the tidal height. The normalized difference water index (NDWI) of each part was calculated from the Sentinel 2 images of the entire tidal cycle to obtain the NDWI value of the entire time series. The maximum NDWI value was used to obtain the maximum sea surface range, and the minimum NDWI value was used to obtain the time point image of the minimum sea surface range. Seawater detection was performed. The maximum and minimum water surfaces of the images available in the area were filtered out through high tide images and low tide images, respectively. The tidal flat wetlands in the offshore and coastal wetlands were divided by difference calculation to separate the sea and land boundaries. Step 1.2: Multi-feature masking of non-wetland areas. For non-wetland areas, firstly, the slope mask is used to select areas with a slope greater than 5°. Then, the digital elevation model (DEM) and the normalized difference vegetation index (NDVI) are segmented by thresholds to perform mountain shadow masking and remove mountain forest vegetation coverage areas and mountain shadows. Finally, the normalized difference building index (NDBI), NDVI, and NDWI are used in the filter to perform bare land masking, and the impervious surface mask is calculated by difference.

3. The method for automatically extracting large-scale wetland data based on cloud computing according to claim 1 is characterized in that: Step 2: Extracting water body wetland segmentation data, including the following steps: Step 2.1, phenological classification. After masking out all classified wetland ranges and non-wetland ranges containing interference data, a phenological classifier was constructed using the random forest (RF) algorithm to draw a mixed wetland map of vegetation and water bodies. In GEE, the time series of Sentinel 2 images, enhanced vegetation index (EVI), NDVI, NDWI and ratio vegetation index (RVI) were connected to the raster collection, and three types of training samples were redistributed, including lake wetlands, swamp wetlands and non-wetlands. The phenological image was obtained using the random forest algorithm, and finally the phenological frequency was obtained by calculating the difference between the NDVI, NDWI, EVI and RVI indices at different phenological periods. Step 2.2, phenological feature segmentation, shape features and NDWI are combined, and the training samples are trained through the simple non-iterative clustering (SNIC) algorithm. K initial centroids are generated in the image plane using a regular grid, and then K corresponding elements are created. Each element includes spatial position, International Illumination Commission Color System (CIELab) color, pixel label, and the distance from the pixel centroid to the candidate pixel. The algorithm quantifies the distance metric through a 5-dimensional feature vector to achieve local clustering of image pixels. Through the calculation of the above process in the GEE cloud platform, the data of five major types of wetlands are segmented and extracted: lake wetlands, river wetlands, swamp wetlands, offshore and coastal wetlands, and artificial wetlands.

4. The method for automatically extracting large-scale wetland data based on cloud computing according to claim 3 is characterized in that: Step 2.1: The conditions for identifying phenological characteristics in phenological classification are: (1) EVI improves the spectral detection sensitivity of chlorophyll concentration and biomass, and uses aerosol drag coefficients (C1, C2), canopy background adjustment (L), and gain factor (G) to significantly compensate for the effects of atmosphere, water, shadow, and soil background. (2) RVI uses near-infrared and red bands to directly infer vegetation growth data from vegetation leaf reflectance. (3) The phenological dynamics of wetland vegetation can be captured through the NDVI, NDWI, EVI, and RVI indices, and the phenological characteristics of vegetation and water bodies can be identified.

5. The method for automatically extracting large-scale wetland data based on cloud computing according to claim 3 is characterized in that: Step 2.2 of feature segmentation includes the calculation of flooding frequency (IF). The flooding frequency (IF) is the frequency at which a single pixel is identified as a water body within a year. Different water bodies in each wetland type have similar spectral characteristics. The spatial resolution of the green and near-infrared bands used to calculate NDWI using Sentinel 2 images is 10m. The detailed wetland types are distinguished: aquaculture ponds and canals. The IF calculation formula is: IF=f_water / f(1) Where f_water is the number of times it is identified as a water body, and f is the number of valid observations of the pixel.

Images