An automatic monitoring and early warning technology for water and soil loss based on transfer learning

Abstract

The application provides a soil erosion automatic updating monitoring and early warning technology based on transfer learning, combines the machine learning method of transfer learning which solves problems in different fields by using existing knowledge, realizes an automatic sequence process of data automatic warehousing, sample automatic extraction and optimization, soil erosion automatic monitoring and early warning, greatly improves the automation degree and practicability of soil erosion monitoring, meets the application requirement of full automation to a certain extent, is very practical and valuable, and can save a large amount of manpower and material resources for future soil erosion mapping work. Moreover, the operation steps of the method only need a small amount of manual participation, can be made into related automatic software, and provides an important scientific means for real-time dynamic monitoring and early warning of soil erosion problems.

Description

Technical Field

[0001] This invention relates to the field of automatic monitoring technology for soil erosion problems, and in particular to an automated monitoring and early warning technology for soil erosion based on transfer learning. Background Technology

[0002] Current monitoring technologies often require object-oriented segmentation and classification models such as HRNet to classify land use in remote sensing images, and then overlay indicators such as slope and vegetation cover to comprehensively determine the intensity of soil erosion. However, before land use classification, each pixel in the remote sensing image needs to be labeled with land use data to form sample data. This process requires extensive visual interpretation and field sampling, often consuming a significant amount of time. Furthermore, each time soil erosion data is updated, interpretation and sampling must be repeated. This manual collection method results in low efficiency for soil erosion monitoring, falling far short of automation and hindering the achievement of dynamic monitoring goals. Moreover, due to limitations imposed by natural conditions, field sampling often only covers a small area, presenting numerous constraints.

[0003] In fact, past map data contains a wealth of prior knowledge, but existing algorithms do not consider how to transfer and learn this knowledge to conduct soil erosion monitoring based on prior knowledge, which largely wastes these resources. Repeated sampling to update the calculation of soil erosion intensity not only wastes a lot of manpower and resources, but also makes it difficult to achieve the purpose of dynamic monitoring and early warning. Summary of the Invention

[0004] This invention proposes an automated monitoring and early warning technology for soil erosion based on transfer learning, which greatly improves the automation level and practicality of soil erosion monitoring, meets the application requirements of "full automation" to a certain extent, and has great practical application value. It can save a lot of manpower and material resources for future soil erosion mapping work.

[0005] The present invention adopts the following technical solution.

[0006] An automated monitoring and early warning technology for soil erosion based on transfer learning includes the following steps;

[0007] Step S1: Establish a multi-source remote sensing database;

[0008] Step S2: Use a multi-scale segmentation algorithm to divide the study area into multiple small patches S;

[0009] Step S3: Using automated machine scanning and extraction methods, extract the scope and intensity of historical soil erosion data in the study area to form historical data sample M;

[0010] Step S4: Using the overlay analysis method, overlay the small patch S with the historical sample M, and extract objects whose overlapping area is greater than 50% of the area of ​​patch S as potential soil erosion samples S. m The others are considered as non-soil erosion samples S n ;

[0011] Step S5: Assign the intensity attribute value of soil erosion to the potential sample S m As the output data of the classification model, it forms samples S with labeled values. m，t .

[0012] Step S6: Extract data information based on the multi-source remote sensing database, and assign information such as the spectral density and slope of the remote sensing image to sample S. m，t S n , as input data for the classification model;

[0013] Step S7: Using a preset ratio, adjust S... m，t S n The data is randomly divided into training data and test data, and fed into a deep learning network model (ResNet18) for training. The model accuracy is calculated. If the accuracy does not meet the requirements, the process returns to step S4, the overlap area threshold is increased, and the samples are re-screened to obtain a soil erosion sample purification model that meets the accuracy requirements.

[0014] Step S8: Calculate S for all potential soil erosion samples using the soil erosion sample purification model. m，t The confidence level;

[0015] Step S9: Extract high-confidence potential soil erosion samples and label them as pure samples S m，t，p The remaining low-confidence samples were marked as dirty samples. Samples with potential soil erosion were subjected to morphological erosion manipulation, shrinking the sample images to form shrunken samples S. m，t，d ;

[0016] Step S10: Return all samples to the sample pool, repeat steps S7 and S8, and calculate S for each dirty sample. m，t，d Scale up the confidence level, repeat step S9, and shrink the low-confidence samples again;

[0017] Step S11: After repeating this process multiple times, when the change in the number of pure samples approaches 0, the iteration ends, all pure samples are extracted, and a pure sample library of soil erosion is obtained.

[0018] Step S12: Using pure soil erosion sample S m，t，p Compared with non-soil erosion sample S n Train the deep learning network model ResNet18 to obtain a soil erosion identification model that meets the accuracy requirements;

[0019] Step S13: Using the soil and water loss identification model in step S12, calculate the probability that all segmented small patches in the study area become soil and water loss patches, in order to determine the intensity of soil and water loss and obtain the updated soil and water loss monitoring results.

[0020] The automated monitoring and early warning technology for soil erosion also includes the following steps;

[0021] Step S14: Set the warning range and warning threshold for soil erosion. When soil erosion occurs in the warning area or the soil erosion area exceeds the threshold, trigger the soil erosion alarm system.

[0022] Step S15: Similarly, update the latest remote sensing data, repeat steps S2-S12, and update the soil and water loss monitoring results in real time. If the soil and water loss intensity exceeds the threshold, or the soil and water loss area affects important functional areas, trigger the soil and water loss alarm system to realize automatic monitoring and early warning of soil and water loss.

[0023] The multi-source remote sensing database in step S1 includes the latest high-resolution remote sensing image data, vegetation data, and topographic data;

[0024] In step S4, under normal circumstances, objects whose overlapping area is greater than 50% of the area of ​​patch S are extracted as potential soil erosion samples S. m In this step, the threshold for overlapping areas that are larger than the area of ​​patch S is set and adjusted according to the actual situation.

[0025] In step S5, soil erosion labels are divided into six categories.

[0026] S m，t (t = slight, mild, moderate, strong, very strong, severe)

[0027] When the classification systems for soil erosion intensity differ across study areas, the labels should be adjusted according to the actual situation.

[0028] In step S6, high-resolution image spectral, NDVI, DEM, and slope data information are extracted based on a multi-source remote sensing database; the high-resolution image spectral includes red, green, blue, and near-infrared multispectral.

[0029] In step S2, multi-scale segmentation technology is used to segment the map patches. Based on multi-source raster data, the study area is converted into multiple map patch objects for subsequent processing. The multi-source raster data includes high-resolution remote sensing imagery, vegetation index, topography, and slope. The core segmentation calculation formula is as follows:

[0030]

[0031]

[0032]

[0033] Where C Li Let C be the brightness value of the i-th pixel within the object in the L-th band, where n is the number of pixels in the object. L The average brightness of a single image object in the L band; S represents the average brightness of all objects in the image in the L-band, m is the total number of objects in the image, and S 2 The variance is represented by the value of the raster being merged into a patch. The variance value of the patch is continuously calculated during the process of merging the raster into a patch. When the variance value changes too much during the merging process, the raster merging is stopped and a patch object is generated.

[0034] In step S3, different techniques are used to extract samples based on the format of the soil and water loss data. For example, if the data is in jpg or tif format, image RGB analysis technology is used for sample extraction. If the data is in shp format, it is extracted directly using ArcGIS software tools.

[0035] In step S6, the average value method is usually used to input various types of information into the map patch. When adjustments are needed based on the actual situation, the variance value, maximum value, or minimum value is used to input the map patch.

[0036] Step S7 uses an 8:2 ratio to mix S... m，t s n The data is randomly divided into training data and test data. If the accuracy does not meet the requirements when calculating the model accuracy, return to step S4, increase the overlap area threshold by 10%, and re-screen the samples.

[0037] In step S8, samples with a confidence level higher than 95% are considered to have high confidence, but this can be adjusted according to the actual situation.

[0038] In step S9, the morphological erosion operation operator is as follows:

[0039]

[0040] In the formula, A represents the sample with low confidence, B is the corrosion factor, and the position in A is denoted as x. Shifting B to x yields B0. x If B x If A is contained within another element, then record point x. The set of x that satisfies all conditions is called the result of A being corroded by B. The structure of the corrosion factor B is as follows:

[0041] 0 1 0 1 1 1 0 1 0 .

[0042] In step S15, the warning zone and warning value are set according to the specific situation of the study area. The soil erosion warning zone is set in the area of ​​important drinking water source and important tributary rivers. The soil erosion warning threshold is determined according to the relevant indicator values ​​of the soil and water conservation plan.

[0043] This invention combines transfer learning, a machine learning method that uses existing knowledge to solve problems in different related fields, and proposes an automatic monitoring and early warning technology for soil erosion based on transfer learning. This technology greatly improves the automation and practicality of soil erosion monitoring, meeting the application requirements of "full automation" to a certain extent. It has significant practical application value and can save a lot of manpower and resources for future soil erosion mapping work. Furthermore, the operation steps of this invention require only minimal human intervention and can be developed into related automated software, providing an important scientific means for real-time dynamic monitoring and early warning of soil erosion problems. Attached Figure Description

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0045] Appendix Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0046] like Figure 1 As shown, an automated monitoring and early warning technology for soil erosion based on transfer learning includes the following steps;

[0047] Step S1: Establish a multi-source remote sensing database;

[0048] Step S2: Use a multi-scale segmentation algorithm to divide the study area into multiple small patches S;

[0049] Step S3: Using automated machine scanning and extraction methods, extract the scope and intensity of historical soil erosion data in the study area to form historical data sample M;

[0050] Step S4: Using the overlay analysis method, overlay the small patch S with the historical data sample M, and extract objects whose overlapping area is greater than the area threshold of patch S as potential soil erosion samples S. m The others are considered as non-soil erosion samples S n ;

[0051] Step S5: Assign the intensity attribute value of soil erosion to the potential sample S m As the output data of the classification model, it forms samples S with labeled values. m，t .

[0052] Step S6: Extract data information based on the multi-source remote sensing database, and assign information such as the spectral density and slope of the remote sensing image to sample S.m，t S n , as input data for the classification model;

[0053] Step S7: Using a preset ratio, adjust S... m，t S n The data are randomly divided into training data and test data, and fed into a deep learning network model (ResNet18) for training. The model accuracy is calculated. If the accuracy does not meet the requirements, the process returns to step S4, the overlap area threshold is increased, and the samples are re-screened to obtain a soil erosion sample purification model that meets the accuracy requirements.

[0054] Step S8: Calculate S for all potential soil erosion samples using the soil erosion sample purification model. m，t The confidence level;

[0055] Step S9: Extract high-confidence potential soil erosion samples and label them as pure samples S m，t，p The remaining low-confidence samples were marked as dirty samples. Samples with potential soil erosion were subjected to morphological erosion manipulation, shrinking the sample images to form shrunken samples S. m，t，d ;

[0056] Step S10: Return all samples to the sample pool, repeat steps S7 and S8, and calculate S for each dirty sample. m，t，d Scale up the confidence level, repeat step S9, and shrink the low-confidence samples again;

[0057] Step S11: After repeating this process multiple times, when the change in the number of pure samples approaches 0, the iteration ends, all pure samples are extracted, and a pure sample library of soil erosion is obtained.

[0058] Step S12: Using pure soil erosion sample S m，t，p Compared with non-soil erosion sample S n Train the deep learning network model ResNet18 to obtain a soil erosion identification model that meets the accuracy requirements;

[0059] Step S13: Using the soil and water loss identification model in step S12, calculate the probability that all segmented small patches in the study area become soil and water loss patches, in order to determine the intensity of soil and water loss and obtain the updated soil and water loss monitoring results.

[0060] The automated monitoring and early warning technology for soil erosion also includes the following steps;

[0061] Step S14: Set the warning range and warning threshold for soil erosion. When soil erosion occurs in the warning area or the soil erosion area exceeds the threshold, trigger the soil erosion alarm system.

[0062] Step S15: Similarly, update the latest remote sensing data, repeat steps S2-S12, and update the soil and water loss monitoring results in real time. If the soil and water loss intensity exceeds the threshold, or the soil and water loss area affects important functional areas, trigger the soil and water loss alarm system to realize automatic monitoring and early warning of soil and water loss.

[0063] The multi-source remote sensing database in step S1 includes the latest high-resolution remote sensing image data, vegetation data, and topographic data;

[0064] In step S4, under normal circumstances, objects whose overlapping area is greater than 50% of the area of ​​patch S are extracted as potential soil erosion samples S. m In this step, the threshold for overlapping areas that are larger than the area of ​​patch S is set and adjusted according to the actual situation.

[0065] In step S5, soil erosion labels are divided into six categories.

[0066] S m，t (t = slight, mild, moderate, strong, very strong, severe)

[0067] When the classification systems for soil erosion intensity differ across study areas, the labels should be adjusted according to the actual situation.

[0068] In step S6, high-resolution image spectral, NDVI, DEM, and slope data information are extracted based on a multi-source remote sensing database; the high-resolution image spectral includes red, green, blue, and near-infrared multispectral.

[0069] In step S2, multi-scale segmentation technology is used to segment the map patches. Based on multi-source raster data, the study area is converted into multiple map patch objects for subsequent processing. The multi-source raster data includes high-resolution remote sensing imagery, vegetation index, topography, and slope. The core segmentation calculation formula is as follows:

[0070]

[0071]

[0072]

[0073] Where C Li Let C be the brightness value of the i-th pixel within the object in the L-th band, where n is the number of pixels in the object. L The average brightness of a single image object in the L band; S represents the average brightness of all objects in the image in the L-band, m is the total number of objects in the image, and S 2 The variance is represented by the value of the raster being merged into a patch. The variance value of the patch is continuously calculated during the process of merging the raster into a patch. When the variance value changes too much during the merging process, the raster merging is stopped and a patch object is generated.

[0074] In step S3, different techniques are used to extract samples based on the format of the soil and water loss data. For example, if the data is in jpg or tif format, image RGB analysis technology is used for sample extraction. If the data is in shp format, it is extracted directly using ArcGIS software tools.

[0075] In step S6, the average value method is usually used to input various types of information into the map patch. When adjustments are needed based on the actual situation, the variance value, maximum value, or minimum value is used to input the map patch.

[0076] Step S7 uses an 8:2 ratio to mix S... m，t S n The data is randomly divided into training data and test data. If the accuracy does not meet the requirements when calculating the model accuracy, return to step S4, increase the overlap area threshold by 10%, and re-screen the samples.

[0077] In step S8, samples with a confidence level higher than 95% are considered to have high confidence, but this can be adjusted according to the actual situation.

[0078] In step S9, the morphological erosion operation operator is as follows:

[0079]

[0080] In the formula, A represents the sample with low confidence, B is the corrosion factor, and the position in A is denoted as x. Shifting B to x yields B0. x If B x If A is contained within another element, then record point x. The set of x that satisfies all conditions is called the result of A being corroded by B. The structure of the corrosion factor B is as follows:

[0081] 0 1 0 1 1 1 0 1 0 In step S15, the warning zone and warning value are set according to the specific conditions of the study area. The soil erosion warning zone is set in the area of ​​important drinking water sources and important tributaries. The soil erosion warning threshold is determined according to the relevant indicator values ​​of the soil and water conservation plan.

[0082] In this example, step S15 uses Fuzhou as an example. The area of ​​soil erosion in Fuzhou needs to be less than 7% of the country's land area, which is approximately 799 km². 2 .

Claims

1. An automated monitoring and early warning method for soil erosion based on transfer learning, characterized in that: Includes the following steps; Step S1: Establish a multi-source remote sensing database; Step S2: Use a multi-scale segmentation algorithm to divide the study area into multiple small patches S; Step S3: Using automated machine scanning and extraction methods, extract the scope and intensity of historical soil erosion data in the study area to form a historical data sample. ; Step S4: Using the overlay analysis method, compare the small patch S with historical results samples. Overlay, extract overlapping areas larger than the patch Objects exceeding the area threshold are considered potential samples for soil erosion. The others were used as non-soil erosion samples. ; Step S5: Assign the intensity attribute value of soil erosion to the potential samples. As output data of the classification model, samples with labeled values ​​are formed. ; Step S6: Extract data information based on the multi-source remote sensing database and assign various types of information to the samples. , , as input data for the classification model; Step S7: Using a preset ratio, , The data are randomly divided into training data and test data, and fed into the deep learning network model ResNet18 for training. The model accuracy is calculated. If the accuracy does not meet the requirements, the process returns to step S4, increases the overlap area threshold, and re-screens the samples to obtain a soil erosion sample purification model that meets the accuracy requirements. Step S8: Calculate all potential soil erosion samples using the soil erosion sample purification model. The confidence level; Step S9: Extract high-confidence potential soil erosion samples and label them as pure samples. The remaining low-confidence samples were marked as dirty samples. Morphological erosion was then applied to shrink the sample images, resulting in shrunken samples. ; Step S10: Return all samples to the sample pool, repeat steps S7 and S8, and calculate the value of each dirty sample. Scale up the confidence level, repeat step S9, and shrink the low-confidence samples again; Step S11: After repeating this process multiple times, when the change in the number of pure samples approaches 0, the iteration ends, all pure samples are extracted, and a pure sample library of soil erosion is obtained. Step S12: Using pure samples of soil erosion Compared with non-soil erosion samples Train the deep learning network model ResNet18 to obtain a soil erosion identification model that meets the accuracy requirements; Step S13: Using the soil and water loss identification model in step S12, calculate the probability that all segmented small patches in the study area become soil and water loss patches, so as to determine the intensity of soil and water loss and obtain the updated soil and water loss monitoring results. The multi-source remote sensing database in step S1 includes the latest high-resolution remote sensing image data, vegetation data, and topographic data.

2. The automated monitoring and early warning method for soil erosion based on transfer learning according to claim 1, characterized in that: The automated monitoring and early warning technology for soil erosion also Includes the following steps; Step S14: Set the warning range and warning threshold for soil erosion. When soil erosion occurs in the warning area or the soil erosion area exceeds the threshold, trigger the soil erosion alarm system. Step S15: Similarly, update the latest remote sensing data, repeat steps S2-S12, and update the soil and water loss monitoring results in real time. If the soil and water loss intensity exceeds the threshold, or the soil and water loss area affects important functional areas, trigger the soil and water loss alarm system to realize automatic monitoring and early warning of soil and water loss.

3. The automated monitoring and early warning method for soil erosion based on transfer learning according to claim 1, characterized in that: In step S4, under normal circumstances, the overlapping part is extracted to be larger than the patch. 50% of the area was used as a potential sample for soil erosion. In this step, the overlapping area is larger than the map patch. The threshold for overlapping areas should be set and adjusted according to the actual situation. In step S5, soil erosion labels are divided into six categories. , When the classification systems for soil erosion intensity differ across study areas, the labels should be adjusted according to the actual situation. In step S6, high-resolution image spectral, NDVI, DEM, and slope data information are extracted based on a multi-source remote sensing database; the high-resolution image spectral includes red, green, blue, and near-infrared multispectral.

4. The automated monitoring and early warning method for soil erosion based on transfer learning according to claim 1, characterized in that: In step S2, multi-scale segmentation technology is used to segment the patches. Based on multi-source raster data, the study area is converted into multiple patch objects for subsequent processing. The multi-source raster data includes high-resolution remote sensing images, vegetation index, topography, and slope.

5. The automated monitoring and early warning method for soil erosion based on transfer learning according to claim 1, characterized in that: In step S3, different techniques are used to extract samples based on the format of the soil and water loss data. For example, if the data is in jpg or tif format, image RGB analysis technology is used for sample extraction. If the data is in shp format, it is extracted directly using ArcGIS software tools.

6. The automated monitoring and early warning method for soil erosion based on transfer learning according to claim 1, characterized in that: In step S6, various types of information are input into the map patch using variance values, maximum values, or minimum values. Step S7 uses an 8:2 ratio to... , The data is randomly divided into training data and test data. If the accuracy does not meet the requirements when calculating the model accuracy, return to step S4, increase the overlap area threshold by 10%, and re-select the samples.

7. The automated monitoring and early warning method for soil erosion based on transfer learning according to claim 1, characterized in that: In step S8, samples with a confidence level higher than 95% are considered to have high confidence, but this can be adjusted according to the actual situation.

8. The automated monitoring and early warning method for soil erosion based on transfer learning according to claim 1, characterized in that: In step S9, the morphological erosion operation operator is as follows: In the formula, For samples with low confidence, As a corrosive agent, The position in is denoted as ,Will Translate to Later obtained ,like Included in Then record point x, and the set of x that satisfies all conditions is called the set of x. quilt As a result of corrosion, the structure of corrosion factor B is as follows: 。 9. The automated monitoring and early warning method for soil erosion based on transfer learning according to claim 2, characterized in that: In step S15, the warning zone and warning value are set according to the specific situation of the study area. The soil erosion warning zone is set in the area of ​​important drinking water source and important tributary rivers. The soil erosion warning threshold is determined according to the relevant indicator values ​​of the soil and water conservation plan.

Images