A batch extraction method of small reservoirs based on high-precision remote sensing images

Abstract

The present invention provides a method for batch extracting small reservoirs based on high-precision remote sensing imagery. The method comprises collecting multiple years of high-resolution land classification data, and further comprises the following steps: fusing the high-resolution land classification data; utilizing multi-source data to remove polygons representing areas without reservoirs from water body polygons; capturing and annotating images from high-definition remote sensing imagery using water body polygons; and constructing and training a machine learning image classification model to classify the images. The method proposed in the present invention batch extracts small reservoirs based on high-precision remote sensing imagery. The method integrates multiple years of high-resolution land use data, utilizes environmental data such as water surface frequency and land use, vegetation, and population distribution to exclude water body data such as lakes and rivers, and utilizes a machine learning image recognition algorithm to extract small reservoirs from a large number of water surfaces.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing target recognition, in particular to a method for batch extraction of small reservoirs based on high-precision remote sensing images. Background Art

[0002] Small reservoirs have a significant impact on surface runoff. First, many modeling practitioners have suggested that the water storage in small reservoirs is one of the factors affecting model accuracy. Second, small reservoirs provide a large amount of water supply and irrigation capacity. Third, small reservoirs are more prone to dam failure, which can cause disasters such as flash floods and mudslides, resulting in losses. There are a large number of water conservancy projects represented by reservoirs around the world. The list and distribution of large reservoirs are relatively clear, but the location and number of small reservoirs are less clear. On the other hand, as reservoirs increase in service life, many reservoirs will be gradually decommissioned, and the number of reservoirs actually in use is dynamic. Therefore, rapid and timely acquisition of information on small reservoirs in use is of great value for flash flood disaster prediction and small and medium-sized river flood forecasting.

[0003] Currently, the most recognized reservoir datasets include HydroLAKES, OSM, GDAT, The Global DamTracker, GOODD, GRanD, GeoDAR, AQUASTAT, and the World Register of Dams (WRD). However, these data products mainly record large reservoirs and some medium-sized reservoirs, and lack the organization and recording of small reservoirs.

[0004] Patent application number CN110826394A discloses a reservoir identification method and device based on a convolutional neural network algorithm. The method includes the following steps: Step 1: Pre-establishing a reservoir classification neural network model combined with a convolutional neural network and training the reservoir classification neural network model using remote sensing images of reservoirs and non-reservoirs; Step 2: For a newly input original remote sensing image, using water-sensitive bands to extract the water area in the image, then using morphological operations to locate the block water area and crop this area from the original image; Step 3: Uniformly reducing the cropped image to a uniform square size; Step 4: Uniformly inputting the images into the reservoir classification neural network model for discrimination and identification of the reservoir area. A disadvantage of this method is that it only uses single-period remote sensing images for reservoir extraction. Single-period remote sensing images may cause reservoirs to dry up or have too little water surface to be extracted. This is particularly common during long-term droughts, where a large number of small reservoirs have no water surface or too little water surface. As a result, many reservoirs, especially small ones, are missed.

[0005] Patent application number CN109754025A discloses a method for identifying parameters of small reservoirs without data, combining hydrological simulation and continuous remote sensing imagery. The method includes the following steps: Step 1: Determine the reservoir area and upstream watershed of the small reservoir without data; Step 2: Determine the reservoir discharge parameters and discharge curve; Step 3: Verify the reservoir discharge parameters and discharge curve; Step 4: Identify the reservoir operation mode. A disadvantage of this method is that it first requires determining the location of the small reservoir, then identifying the reservoir area and extracting the reservoir parameters based on this. This method cannot identify and extract small reservoir objects over a large area when the location and distribution of the small reservoirs are unknown. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention proposes a batch extraction method of small reservoirs based on high-precision remote sensing images, which integrates many years of high-resolution land use data, uses water surface occurrence frequency and environmental data such as land use, vegetation, and population distribution to exclude water body data such as lakes and rivers, and uses machine learning image recognition algorithms to extract small reservoirs from a large number of water bodies.

[0007] The purpose of the present invention is to provide a method for batch extraction of small reservoirs based on high-precision remote sensing images, which includes collecting high-resolution land classification data for many years and further includes the following steps:

[0008] Step 1: Fusing the high-resolution land classification data;

[0009] Step 2: Using multi-source data, remove polygons without reservoir distribution areas from water body polygons;

[0010] Step 3: Use water body polygons to capture and annotate images from high-definition remote sensing images;

[0011] Step 4: Build and train a machine learning image classification model to classify images.

[0012] Preferably, the high-resolution land classification data is a raster layer, the raster attribute values are numbers, and different numbers represent different land use types.

[0013] In any of the above solutions, preferably, step 1 includes the following sub-steps:

[0014] Step 11: Extract the water body type, form a new yearly raster layer, and assign the attribute value of the raster to 1;

[0015] Step 12: Add the rasters of all years to get a new raster layer:

[0016]

[0017] in,n is the year number, v i For the i The grid value of the year, V is the grid value after the sum operation, V In 0- n between;

[0018] Step 13: According to V Values are generated to generate two new raster layers. The RasterToPolygon tool in ArcGIS software is used to convert the two raster layers into two vector layers, polygon1 and polygon2.

[0019] Step 14: Extract two new vector layers based on the areas of the polygons in the two vector layers.

[0020] In any of the above solutions, preferably, step 13 includes: V ≥1, it is the first raster layer; when V When ≥2, it is the second raster layer.

[0021] In any of the above solutions, preferably, step 14 includes extracting polygons with an area greater than a first area threshold from polygon1 to generate layer1; and extracting polygons with an area greater than a second area threshold from polygon2 to generate layer2.

[0022] In any of the above solutions, preferably, step 14 further comprises using a spatial selection tool of ArcGIS software to select polygons from polygon2 that intersect polygon1 and remove them from polygon2.

[0023] In any of the above solutions, preferably, step 2 includes the following sub-steps:

[0024] Step 21: Based on polygon2, collect watershed layer attributes. If the area of each watershed is smaller than the first area threshold, extract three watershed layers according to the watershed layer attribute requirements, perform dissolve processing on these three layers, and remove discrete watersheds.

[0025] Step 22: Use the three watershed layers to overlay polygon2, select the intersecting polygons, and remove them from polygon2.

[0026] In any of the above solutions, preferably, the watershed layer attributes include average slope, population size and snow cover attributes.

[0027] In any of the above solutions, preferably, the watershed layer attribute requirements are:

[0028] 1) The average slope is less than 2 degrees;

[0029] 2) The population is 0;

[0030] 3) Snow cover is greater than 80%.

[0031] In any of the above solutions, preferably, the criterion for determining the discrete watershed is that the number of adjacent watershed units is less than 5.

[0032] In any of the above solutions, preferably, step 3 includes the following sub-steps:

[0033] Step 31: Overlay the online HD imagery under the water body polygon layer and export HD images of the polygons one by one;

[0034] Step 32: Use a random algorithm to select no less than 1% of the pictures, and use a manual interpretation method to manually classify the selected pictures into two groups: reservoirs and non-reservoirs, to form a picture sample library.

[0035] In any of the above solutions, preferably, step 4 includes the following sub-steps:

[0036] Step 41: using the picture samples in the picture sample library, obtaining the region of interest of each picture sample through image color analysis;

[0037] Step 42: Standardize and crop the region of interest to obtain a standardized sample for training and learning;

[0038] Step 43: Using the sample, the features in the image are learned and trained based on the CNN convolutional neural network to generate a machine learning image classification model;

[0039] Step 44: Import the image into the machine learning image classification model and output the classification result.

[0040] This paper proposes a batch extraction method for small reservoirs based on high-precision remote sensing images. Based on multi-period resolution land use data and high-precision remote sensing images, it comprehensively utilizes the advantages of manual identification and machine learning to achieve rapid identification of small reservoirs over a large area, and avoids the problem of missing some small reservoirs due to drought or water use resulting in the water surface being too small.

[0041] The RasterToPolygon tool is a tool that converts raster data into vector data. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1The present invention is a flowchart of a preferred embodiment of a method for batch extraction of small reservoirs based on high-precision remote sensing images.

[0043] Figure 2 The flowchart is another preferred embodiment of the method for batch extraction of small reservoirs based on high-precision remote sensing images according to the present invention.

[0044] Figure 3 The figure is a schematic diagram of an embodiment of a summed raster file of a method for batch extraction of small reservoirs based on high-precision remote sensing images according to the present invention.

[0045] Figure 4 The figure is a schematic diagram of an embodiment of the water body polygon extraction method of small reservoirs based on high-precision remote sensing images according to the present invention.

[0046] Figure 5 This is a schematic diagram of a non-reservoir class according to an embodiment of the manual annotation classification results of the small reservoir batch extraction method based on high-precision remote sensing images according to the present invention.

[0047] Figure 6 This is a schematic diagram of reservoir classes according to an embodiment of the manual annotation classification results of the small reservoir batch extraction method based on high-precision remote sensing images according to the present invention.

[0048] Figure 7 A schematic diagram of reservoir classification according to an embodiment of the classification results of the method for batch extraction of small reservoirs based on high-precision remote sensing images of the present invention.

[0049] Figure 8 A schematic diagram of river classification according to an embodiment of the classification results of the method for batch extraction of small reservoirs based on high-precision remote sensing images of the present invention.

[0050] Figure 9 A schematic diagram of pond or lake classification according to an embodiment of the classification results of the method for batch extraction of small reservoirs based on high-precision remote sensing images of the present invention. DETAILED DESCRIPTION

[0051] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Example 1

[0052] like Figure 1 As shown, a batch extraction method for small reservoirs based on high-precision remote sensing images executes step 100 to collect high-resolution land classification data for many years. The high-resolution land classification data is a raster layer, and the raster attribute value is a number. Different numbers represent different land use types.

[0053] Executing step 110 to fuse the high-resolution land classification data includes the following sub-steps:

[0054] Execute step 111 to extract the water body type, form a new yearly raster layer, and assign the attribute value of the raster to 1.

[0055] Execute step 112 to sum the rasters of all years to obtain a new raster layer:

[0056]

[0057] in, n is the year number, v i For the i The grid value of the year, V is the grid value after the sum operation, V In 0- n between.

[0058] Execute step 113, according to V Value, generate two new raster layers, use the RasterToPolygon tool in ArcGIS software to convert the two raster layers into two vector layers, polygon1 and polygon2. V ≥1, it is the first raster layer; when V When ≥2, it is the second raster layer.

[0059] Execute step 114 to extract two new vector layers based on the areas of the polygons in the two vector layers. Extract polygons with an area greater than a first area threshold from polygon1 to generate layer1; extract polygons with an area greater than a second area threshold from polygon2 to generate layer2; use the spatial selection tool in ArcGIS software to select polygons from polygon2 that intersect with polygon1 and remove them from polygon2.

[0060] Executing step 120 to remove polygons without reservoir distribution areas from water body polygons using multi-source data includes the following sub-steps:

[0061] Execute step 121. Based on polygon2, collect watershed layer attributes. For each watershed whose area is less than the first area threshold, extract three watershed layers according to the watershed layer attribute requirements. Dissolve the three layers and remove the discrete watersheds. The watershed layer attributes include average slope, population, and snow cover. The watershed layer attribute requirements are:

[0062] 1) The average slope is less than 2 degrees;

[0063] 2) The population is 0;

[0064] 3) Snow cover is greater than 80%.

[0065] The criterion for determining a discrete watershed is that there are less than 5 adjacent watershed units.

[0066] Execute step 122 to overlay polygon2 with the three watershed layers, select intersecting polygons, and remove them from polygon2.

[0067] Executing step 130 to capture and annotate images from the high-definition remote sensing image using water body polygons includes the following sub-steps:

[0068] Execute step 131 to overlay the online high-definition image under the water body polygon layer and export high-definition images of the polygons one by one;

[0069] Execute step 132, use a random algorithm to select no less than 1% of the pictures, and use a manual interpretation method to manually classify the selected pictures into two groups: those that are reservoirs and those that are not reservoirs, thereby forming a picture sample library.

[0070] Execute step 140 to build and train a machine learning image classification model to classify images, including the following sub-steps:

[0071] Executing step 141, using the picture samples in the picture sample library, obtaining a region of interest of each picture sample through image color analysis;

[0072] Executing step 142, standardizing and cropping the region of interest to obtain a standardized sample for training and learning;

[0073] Execute step 143 to use the sample to learn and train features in the image based on a CNN convolutional neural network to generate a machine learning image classification model;

[0074] Execute step 144 to import the image into the machine learning image classification model and output the classification result. Example 2

[0075] This paper proposes a method for extracting small reservoirs over a large area based on years of high-resolution land use data and high-precision remote sensing imagery. First, the method extracts and fuses water polygons from the years of high-resolution land use data to identify possible small reservoir polygons. Then, a remote sensing image of each polygon is captured from the high-resolution remote sensing imagery, extending a certain range. Finally, an image recognition neural network is trained to identify small reservoirs.

[0076] (1) Integrate high-resolution land use data from multiple years and use environmental data such as water surface frequency and land use, vegetation, and population distribution to exclude water body data such as lakes and rivers;

[0077] (2) Using machine learning image recognition algorithms, small reservoirs are extracted from a large number of water bodies.

[0078] The calculation process of the present invention is as follows Figure 2 shown.

[0079] Step 1: Fusion of multi-year high-resolution land use classification data.

[0080] We collected high-resolution land classification data over many years. This data is a raster layer with numeric attribute values, representing different land use types. We first extracted the water body types to form a new yearly raster layer. We then assigned the attribute values of the raster to 1. We then added up the rasters for all years to create a new raster layer:

[0081]

[0082] Where, n is the year number, v i For the i The grid value of the year, here, v i The value of is unified to 1, which is the grid value after the addition operation. V In 0- n between.

[0083] according to V Value, generate two new raster layers, if V ≥1, for the first raster layer, if V ≥2, which is the second raster layer. Then use the RasterToPolygon tool in ArcGIS software to convert the two raster layers into two vector layers, polygon1 and polygon2.

[0084] According to the area of the polygons in the two vector layers, two new vector layers are extracted: polygons with an area greater than 1km² are extracted from polygon1 to generate layer1. Here, if the area is greater than 1km 2, generally medium-to-large reservoirs, do not require extraction and can be obtained from existing datasets. This area threshold can be adjusted in practice. Polygons with an area greater than 1 ha² are extracted from polygon2 to generate layer2. Water bodies with an area less than 1 ha² are generally not large enough to accommodate small reservoirs. This area threshold can be adjusted in practice. Using ArcGIS's spatial selection tools, select polygons from polygon2 that intersect polygon1 and remove them from polygon2. This operation removes medium-to-large reservoirs from polygon2. This operation uses the ArcGIS Select By Location command, with the Intersect the Source Layer Feature option selected in the Overlay option.

[0085] In the second step, multi-source data are used to remove polygons without reservoir distribution areas from water body polygons.

[0086] Based on polygon2 obtained in step 1, collect watershed layers (including average slope, population, and snow cover). Each watershed should be less than 100 km². Extract watershed layers for areas with an average slope less than 2 degrees, a population of 0, and a snow cover greater than 80%. Dissolve these three layers separately and remove discrete watersheds. If there are fewer than five adjacent watershed units, remove them.

[0087] Use the three layers obtained to overlay polygon2, select the intersecting polygons, and remove them from polygon2. This operation uses the Select By Location command in ArcGIS software, and select intersect the source layer feature in the overlay option.

[0088] Step 3: Use water body polygons to capture and label images from high-definition remote sensing images.

[0089] Using ArcGIS VBA, we overlaid online high-definition imagery on the water body polygon layer and exported high-definition images of each polygon individually. To include surrounding information, we expanded the bounding rectangle of each polygon by 1.5 times. Due to the large number of images, we launched multiple programs in parallel to export them, resulting in a large number of images.

[0090] Using a random algorithm, select no less than 1% of the images (if the number is less than 10,000, select 10,000), such as the 1st, 101st, 201st, etc. Then use a manual interpretation method to manually classify the selected images into two groups: those that are reservoirs and those that are not reservoirs.

[0091] Step 4: Build and train a machine learning image classification algorithm to classify all images.

[0092] Using the image samples obtained in step 3, we analyze the image color to identify the region of interest (i.e., the water body boundary) for each image sample. This region is then normalized and cropped to obtain a standardized training sample. Using this sample, a convolutional neural network (CNN) is used to learn and train the image features. Once trained, the model can effectively identify the type of water bodies in remote sensing impact images. Example 3

[0093] This embodiment discloses an example of obtaining a small reservoir according to the extraction method of the present application.

[0094] In the first step, global land use raster data with a 10-meter grid was collected for seven consecutive years from 2017 to 2023. The dataset is sliced at 6° longitude and 8° latitude, and includes 712 slices per year. Using ArcGIS software, the water body grids are first extracted, and then the grid values are set to 1. The grids for the seven years are then summed to obtain 712 summed raster files, as shown in the following figure. Figure 3 shown.

[0095] Extract all rasters with raster values greater than or equal to 1 and raster values greater than or equal to 2 from the raster file and convert them into vector layers. Extract polygons with an area greater than 1 km² and an area greater than 1 ha² from the vector layers to obtain layers 1 and 2 respectively. Then remove polygons from layer 2 that intersect with polygons in layer 1. Finally, 7.87 million polygons are obtained. The results are as follows: Figure 4 As shown, the range is 110°-111° east longitude and 40°-41° north latitude.

[0096] The second step was to collect global BasinATLAS_v10 watershed data, which contains 1.18 million watersheds. Each watershed's attributes included average slope, population, and snow cover. Three layers were extracted based on the conditions for an average slope greater than 2 degrees, zero population, and snow cover greater than 80%. Discrete watersheds (fewer than five adjacent watersheds) were removed from each layer. These three layers were then overlaid with the water polygon layers obtained in the previous step. Water polygons that intersected with the watershed polygons in the three layers were removed, resulting in 1.51 million polygons.

[0097] Step 3: Use water polygons to capture and annotate images from high-definition remote sensing images. The remote sensing images use ESRI's free World Imagery data, which includes three resolutions: 30cm, 60cm, and 15m. The company processed the data into image data at 14 scales, ranging from 1.9cm to 150m. A batch processing program was used to generate 1.51 million images, and 1% of them were selected for manual annotation and classification. Figure 5 The images shown are confirmed to be not reservoirs, such as Figure 6 Shown is an image of what is confirmed to be a reservoir.

[0098] Step 4: Use the image samples obtained in step 3 to build and train a machine learning image classification algorithm to classify all images.

[0099] Based on the image samples in step 3, the water bodies in the images are preliminarily identified through image color analysis to obtain the region of interest of the image samples. The region is then standardized and cropped to obtain standardized training learning samples.

[0100] The standardized sample is used as a learning sample, and the CNN convolutional neural network model is used for feature-based recognition. According to the needs of daily work and the characteristics of the sample images, the training samples are classified into Pool (pond or lake), River (river) and Reservoir (reservoir). The results are as follows: Figure 7 The reservoir is shown; Figure 8 The river is shown; Figure 9 Pond or lake shown.

[0101] In order to better understand the present invention, the above is described in detail in conjunction with the specific embodiments of the present invention, but it is not intended to limit the present invention. Any simple modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention. Each embodiment in this specification focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

Claims

1. A method for batch extraction of small reservoirs based on high-precision remote sensing images, including the collection of high-resolution land classification data over many years, and further comprising the following steps: Step 1: Fusion of the high-resolution land classification data, including the following sub-steps: Step 11: Extract the water body type, form a new yearly raster layer, and assign the attribute value of the raster to 1; Step 12: Add the rasters of all years to get a new raster layer: ， in, n is the year number, v i For the i The grid value of the year, V is the grid value after the sum operation, V In 0- n between; Step 13: According to V Values are generated to generate two new raster layers. The RasterToPolygon tool in ArcGIS software is used to convert the two raster layers into two vector layers, polygon1 and polygon2. Step 14: Extract two new vector layers based on the areas of the polygons in the two vector layers; Step 2: Using multi-source data, remove the polygons without reservoir distribution areas from the water body polygons, including the following sub-steps: Step 21: Based on polygon2, collect watershed layer attributes. If the area of each watershed is smaller than the first area threshold, extract three watershed layers according to the watershed layer attribute requirements, perform dissolve processing on these three layers, and remove discrete watersheds. Step 22: Overlay the three watershed layers with polygon2, select the intersecting polygons, and remove them from polygon2; Step 3: Use water body polygons to capture and annotate images from high-definition remote sensing images; Step 4: Build and train a machine learning image classification model to classify images.

2. The method for batch extraction of small reservoirs based on high-precision remote sensing images according to claim 1, characterized in that: The step 13 includes when V ≥1, it is the first raster layer; when V When ≥2, it is the second raster layer.

3. The method for batch extraction of small reservoirs based on high-precision remote sensing images according to claim 2, characterized in that: The step 14 includes extracting polygons with an area greater than a first area threshold from polygon1 to generate layer1; and extracting polygons with an area greater than a second area threshold from polygon2 to generate layer2.

4. The method for batch extraction of small reservoirs based on high-precision remote sensing images according to claim 3 is characterized in that: The step 14 further includes using a spatial selection tool of ArcGIS software to select polygons from polygon2 that intersect polygon1 and remove them from polygon2.

5. The method for batch extraction of small reservoirs based on high-precision remote sensing images according to claim 4 is characterized in that: The watershed layer attributes include average slope, population, and snow cover attributes.

6. The method for batch extraction of small reservoirs based on high-precision remote sensing images according to claim 5, characterized in that: The watershed layer attribute requirements are: 1) The average slope is less than 2 degrees; 2) The population is 0; 3) Snow cover is greater than 80%.

7. The method for batch extraction of small reservoirs based on high-precision remote sensing images according to claim 5, characterized in that: The criterion for determining a discrete watershed is that there are less than 5 adjacent watershed units.

8. The method for batch extraction of small reservoirs based on high-precision remote sensing images according to claim 7, characterized in that: Step 3 includes the following sub-steps: Step 31: Overlay the online HD imagery under the water body polygon layer and export HD images of the polygons one by one; Step 32: Use a random algorithm to select no less than 1% of the pictures, and use a manual interpretation method to manually classify the selected pictures into two groups: reservoirs and non-reservoirs, to form a picture sample library.

9. The method for batch extraction of small reservoirs based on high-precision remote sensing images according to claim 8, characterized in that: The step 4 includes the following sub-steps: Step 41: using the picture samples in the picture sample library, obtaining the region of interest of each picture sample through image color analysis; Step 42: Standardize and crop the region of interest to obtain a standardized sample for training and learning; Step 43: Using the sample, the features in the image are learned and trained based on the CNN convolutional neural network to generate a machine learning image classification model; Step 44: Import the image into the machine learning image classification model and output the classification result.

Images