Inland pond classification method and device at basin scale and electronic equipment

Abstract

The present application relates to a kind of inland pond classification method, device and electronic equipment of basin scale, the method includes: obtaining the multi-source geospatial data of target basin, including remote sensing image, digital elevation model, land cover data, JRC global land surface water product, city boundary data, photovoltaic power station distribution data;Radiometric calibration is carried out to remote sensing image, topographic correction, noise removal and spatial clipping;Inland pond water body baseline map is generated based on multi-source geospatial data;For the differentiated characteristics of aquaculture pond, city pond, photovoltaic pond, tailing pond, farmland pond, differential combination application is carried out, and the recognition and classification result of multi-class inland pond is output.The present application significantly improves the detection rate of small-scale pond through multi-source geospatial data fusion, improves the recognition accuracy of inland pond, the classification accuracy of whole process is significantly improved compared with single algorithm, time consumption is significantly reduced, and high-precision recognition and classification of multi-class pond in basin scale are realized.

Description

Technical Field

[0001] This invention relates to the field of remote sensing identification and classification technology for small water bodies, and in particular to a method, apparatus and electronic equipment for classifying inland ponds at the watershed scale. Background Technology

[0002] Inland ponds, as the main body of global pond systems (accounting for 64.5%, with 38.1% in China), play a vital role in maintaining biodiversity, supporting aquaculture, and regulating water resources. However, human-driven pond expansion has led to environmental risks such as eutrophication, algal blooms, and heavy metal pollution, necessitating refined classification and management. While remote sensing technology provides an efficient means of water body identification, significant technical bottlenecks remain in the classification of inland ponds.

[0003] Insufficient recognition accuracy: Traditional water index methods and pixel-level supervised classification are difficult to effectively distinguish between small-area, shape-variable inland ponds and natural water bodies, and are easily affected by clouds, shadows and turbid water; the similarity of pond spectral features leads to missed detections and misclassifications, and there is a lack of robust recognition methods for small-scale water bodies.

[0004] Lack of a classification system: Existing studies mostly focus on the independent identification of single types of ponds, lacking a unified classification framework that covers multiple categories; global water body datasets only identify the existence of water bodies and do not subdivide them by functional type, which cannot support differentiated ecological governance.

[0005] Limitations of algorithm adaptability: A single machine learning algorithm is difficult to take into account the differentiated characteristics of multiple types of ponds; the lack of a systematic classification framework that integrates multi-source data and multi-algorithm collaboration restricts the accuracy and efficiency of watershed-scale applications. Summary of the Invention

[0006] Therefore, it is necessary to provide a watershed-scale inland pond classification method, device, and electronic equipment to address the problems of insufficient identification technology, lack of multi-category classification system, and limited algorithm adaptability.

[0007] The present invention provides a method for classifying inland ponds at the watershed scale, the method comprising:

[0008] Acquire multi-source geospatial data for the target watershed, including remote sensing imagery, digital elevation models, land cover data, JRC global surface water products, urban boundary data, photovoltaic power station distribution data, mining area polygon data, lake datasets, and statistical yearbook data;

[0009] Radiometric calibration, topographic correction, noise removal, and spatial cropping of remote sensing images;

[0010] Baseline maps of inland pond water bodies are generated based on multi-source geospatial data;

[0011] Based on the differentiated characteristics of aquaculture ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds, differentiated combined applications are applied to output identification and classification results for multiple categories of inland ponds. The differentiated combined applications include: using object-oriented analysis combined with geometric feature decision trees and neighborhood space discrimination to identify aquaculture ponds; using land cover overlay analysis and buffer ratio verification to identify urban ponds; using random forest algorithm to train power station samples to identify photovoltaic ponds; constructing a convolutional neural network model combined with spectral texture features to identify tailings ponds; and applying digital elevation model and neighborhood analysis of cultivated land data to identify farmland ponds.

[0012] In one embodiment, the radiometric calibration, topographic correction, noise removal, and spatial cropping of the remote sensing imagery includes:

[0013] Calculate the median backscattering from Sentinel-1 SAR multi-year sequence data to eliminate salt-and-pepper noise;

[0014] Radiometric calibration and topographic correction were performed on the Sentinel-2 imagery, and the images were cropped according to the study area boundaries.

[0015] In one embodiment, generating an inland pond water baseline map based on multi-source geospatial data includes:

[0016] The dual-polarized water index was calculated based on JRC global surface water product and Sentinel-1 SAR data, and the threshold was determined by the water body frequency method to generate the initial water body mask.

[0017] By integrating Sentinel-2 data to optimize mask boundaries, a baseline map of inland ponds at the watershed scale is formed.

[0018] In one embodiment, the method of identifying aquaculture ponds using object-oriented analysis combined with geometric feature decision trees and neighborhood space discrimination includes:

[0019] Global city boundary data was used to delineate non-urban areas, and baseline maps were overlaid to extract the coarse range of aquaculture ponds.

[0020] Based on Sentinel-2 imagery and field sampling data, a geometric feature decision tree was constructed to screen aquaculture ponds.

[0021] In one embodiment, the construction of a geometric feature decision tree for screening aquaculture ponds includes:

[0022] The neighborhood expansion method is used to generate buffer zones for potential ponds, and isolated non-aquaculture ponds are identified based on the intersection of the buffer zones.

[0023] Fish ponds and shrimp / crab ponds are distinguished by the normalized difference index and the normalized difference vegetation index.

[0024] In one embodiment, the identification of urban ponds using land cover overlay analysis and buffer ratio verification includes:

[0025] Global city boundary data and SinoLC-1 city class layers are overlaid on the baseline map to filter candidate sets of city ponds;

[0026] Independence was verified using a neighborhood-based expansion method and further confirmed by the proportion of impermeable surfaces in the buffer zone.

[0027] In one embodiment, the step of using a random forest algorithm to train power station samples to identify photovoltaic ponds includes:

[0028] Collect sample points of photovoltaic power stations in the target area;

[0029] Set the parameters for the random forest model, train and test the model;

[0030] The distribution of photovoltaic ponds is determined by overlaying the model output with the baseline map.

[0031] In one embodiment, the construction of a convolutional neural network model incorporating spectral texture features to identify tailings pools includes:

[0032] The target area of ​​the tailings pool was delineated based on DEM, SinoLC-1 bare map layer and global mining area dataset;

[0033] A CNN model is constructed, inputting spectral and texture features, and trained to identify tailings pools through convolutional layers, max pooling layers, and fully connected layers.

[0034] In one embodiment, the identification of farmland ponds using the application of digital elevation model and neighborhood analysis of cultivated land data includes:

[0035] Overlay the DEM, SinoLC-1 cultivated land layer, and non-urban boundary data onto the baseline map;

[0036] Neighborhood analysis was used to verify whether the ponds were independently distributed in the farmland.

[0037] In one embodiment, the method further includes:

[0038] Output spatial distribution maps of inland ponds of multiple categories, and construct confusion matrices for fish ponds, shrimp and crab ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds;

[0039] The accuracy of the framework was assessed using the Nash coefficient (NS), correlation coefficient (R²), root mean square error (RMSE), and mean absolute error (MAE).

[0040] The present invention also provides a watershed-scale inland pond classification device, the device comprising:

[0041] The multi-source data acquisition module is used to acquire multi-source geospatial data of the target watershed, including remote sensing imagery, digital elevation models, land cover data, JRC global surface water products, urban boundary data, photovoltaic power station distribution data, mining area polygon data, lake datasets, and statistical yearbook data.

[0042] The data preprocessing module is used to perform radiometric calibration, topographic correction, noise removal, and spatial cropping on remote sensing images.

[0043] The water baseline map generation module is used to generate inland pond water baseline maps based on multi-source geospatial data.

[0044] The differentiated classification module is used to perform differentiated combined applications based on the differentiated characteristics of aquaculture ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds, and output the identification and classification results of multiple categories of inland ponds. The differentiated combined applications include: using object-oriented analysis combined with geometric feature decision trees and neighborhood space discrimination to identify aquaculture ponds; using land cover overlay analysis and buffer ratio verification to identify urban ponds; using random forest algorithm to train power station samples to identify photovoltaic ponds; constructing a convolutional neural network model combined with spectral texture features to identify tailings ponds; and applying digital elevation model and neighborhood analysis of cultivated land data to identify farmland ponds.

[0045] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the inland pond classification method at the watershed scale as described above.

[0046] The aforementioned watershed-scale inland pond classification methods, devices, and electronic equipment significantly improve the detection rate of small-scale ponds through multi-source geospatial data fusion, eliminating interference from clouds and water turbidity. Accurate classification is achieved through a combination of differentiated algorithms. For aquaculture ponds, a geometric feature decision tree combined with neighborhood spatial discrimination is used, achieving high accuracy. For urban ponds, land cover overlay and impermeable surface ratio verification are applied to avoid misclassification. For photovoltaic ponds, a lightweight random forest model is used to capture the spectral characteristics of the water surface and photovoltaic panels coexisting, achieving high accuracy and short processing time. For tailings ponds, a CNN is constructed to fuse spectral and texture features, improving the recognition rate compared to traditional methods. For farmland ponds, unsupervised and efficient identification is achieved based on digital elevation models and neighborhood analysis of cultivated land data, improving efficiency compared to supervised methods. Ultimately, the overall classification accuracy is significantly improved compared to single algorithms, while the processing time is significantly reduced, realizing high-precision identification and classification of multiple types of ponds at the watershed scale. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0048] Figure 1 A flowchart illustrating a watershed-scale inland pond classification method as an example;

[0049] Figure 2 A schematic diagram illustrating the principle of an inland pond classification method at the watershed scale, as shown in one embodiment.

[0050] Figure 3 Flowchart of an inland pond classification method at the watershed scale, as shown in another embodiment;

[0051] Figure 4 A schematic diagram of a classification decision tree extracted from aquaculture ponds;

[0052] Figure 5 Flowchart of a watershed-scale inland pond classification method as another embodiment;

[0053] Figure 6 Flowchart of an inland pond classification method at the watershed scale, as another embodiment;

[0054] Figure 7 A schematic diagram of a random forest extracted from a photovoltaic pond;

[0055] Figure 8 Flowchart of an inland pond classification method at the watershed scale, as another embodiment;

[0056] Figure 9 A schematic diagram of a convolutional neural network extracted from tailings ponds;

[0057] Figure 10 Flowchart of an inland pond classification method at the watershed scale, as another embodiment;

[0058] Figure 11 A schematic diagram of an inland pond classification device at the watershed scale, as an example;

[0059] Figure 12 This is an internal structural diagram of a computer device according to one embodiment. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Inland ponds, as the main body of global pond systems (accounting for 64.5%, with 38.1% in China), play a vital role in maintaining biodiversity, supporting aquaculture, and regulating water resources. However, human-driven pond expansion (such as food production and energy development) has led to environmental risks such as eutrophication, algal blooms, and heavy metal pollution, necessitating refined classification and management. While remote sensing technology provides an efficient means of water body identification, significant technical bottlenecks remain in the classification of inland ponds.

[0062] Insufficient recognition accuracy: Traditional water index methods (such as NDWI) and pixel-level supervised classification (MLC, SVM) are difficult to effectively distinguish between small-area, shape-variable inland ponds and natural water bodies, and are easily affected by clouds, shadows and turbid water. The similarity of spectral features of ponds (such as fish ponds and shrimp and crab ponds, photovoltaic ponds and ordinary ponds) leads to missed detections and misclassifications, and there is a lack of robust recognition methods for small-scale water bodies.

[0063] Lack of a classification system: Existing studies mostly focus on the independent identification of single types of ponds (such as aquaculture ponds and tailings ponds), lacking a unified classification framework that covers multiple categories (aquaculture / urban / photovoltaic / tailings / farmland ponds); global water body datasets (such as JRC-GSW) only identify the existence of water bodies, without subdividing them by functional type, and cannot support differentiated ecological governance.

[0064] Limitations of algorithm adaptability: Single machine learning algorithms (such as random forests and CNNs) struggle to take into account the differentiated characteristics of various types of ponds (such as geometric regularity, spatial location, and texture complexity); the lack of a systematic classification framework that integrates multi-source data (optical / radar remote sensing, topography, and land use) with multi-algorithm collaboration (spatial analysis + decision tree + deep learning) restricts the accuracy and efficiency of watershed-scale applications.

[0065] Therefore, developing a classification method for inland ponds that integrates multi-source remote sensing data, couples spatial analysis and machine learning algorithms, and adapts to watershed scales is of great significance for improving the accuracy of small-scale water system mapping and supporting ecological protection decision-making.

[0066] The following is combined Figures 1-12 The present invention describes a method, apparatus, and electronic device for classifying inland ponds at the watershed scale.

[0067] like Figure 1 and Figure 2 As shown, in one embodiment, a method for classifying inland ponds at the watershed scale includes the following steps:

[0068] Step S110, Multi-source data collaborative processing: Acquire multi-source geospatial data of the target watershed, including remote sensing images, digital elevation models, land cover data, JRC global surface water products, urban boundary data, photovoltaic power station distribution data, mining area polygon data, lake datasets and statistical yearbook data.

[0069] The specific requirements for remote sensing imagery (including Sentinel-1 SAR data and Sentinel-2 optical imagery), digital elevation models (DEMs), land cover data (SinoLC-1), JRC global surface water products, urban boundary data, photovoltaic power plant distribution data, mining area polygon data, lake datasets, and statistical yearbook data are as follows: Sentinel-1 Synthetic Aperture Radar (SAR) data spanning multiple years for the target study area; Sentinel-2 satellite remote sensing imagery data for the corresponding years; SRTMGL1 global 1-arcsecond DEM data with a spatial resolution of 30 m; and data for China. The data includes m-resolution national-scale land cover maps, i.e., land cover data (SinolC-1); the China photovoltaic power station polygon geospatial dataset, which contains spatial distribution information of China's photovoltaic power stations from 2010 to 2022; the global-scale mining area polygons (version 2), which includes mining area polygon data from 2000 to 2019; vector boundary data of the target study area and global city boundary vector data; the corresponding Chinese lake dataset within the study area; and statistical data from the corresponding years of the "China Fisheries Statistical Yearbook" and the "China Statistical Yearbook".

[0070] Step S120: Radiometric calibration, topographic correction, noise removal, and spatial cropping are performed on the remote sensing image.

[0071] Noise removal from remote sensing images employs a time-series median composite method, specifically involving: calculating the median backscatter of Sentinel-1 SAR cross-year sequence data to eliminate salt-and-pepper noise generated by factors such as sensor interference, atmospheric particles, and the imaging environment during remote sensing imaging. Since the water volume of inland ponds remains relatively continuous over long periods, the backscatter values ​​remain stable within a specific range throughout the year. Compared to seasonal water bodies, median composite further enhances low backscatter characteristics. Utilizing the stability of backscatter values ​​throughout the year, median composite eliminates transient outliers (such as bird flocks and boat interference), thereby enhancing the low-scatter characteristics of the water body. Subsequently, radiometric calibration and topographic correction are performed on Sentinel-2 images, followed by cropping according to the study area boundaries. Radiometric calibration corrects sensor errors, and topographic correction eliminates shadow distortion, ensuring that subsequent classification is not affected by imaging conditions and reducing errors in small pond boundary extraction.

[0072] By fusing the spectral details of optical imagery (Sentinel-2) with the cloud and rain resistance of radar data (Sentinel-1), and combining this with DEM topographic constraints, the detection capability of small-scale ponds is significantly improved, overcoming the shortcomings of traditional single optical data which are susceptible to environmental interference. Through iterative optimization using JRC global surface water product data and localized remote sensing data, seasonal and temporary water body interference is eliminated, ensuring the spatial benchmark accuracy of inland ponds. Radiometric calibration and topographic correction reduce Sentinel-2 atmospheric distortion, ensuring that the input data quality meets the requirements of watershed-scale analysis.

[0073] Step S130, Dynamic construction of classification framework: Generate baseline map of inland pond water body based on multi-source geospatial data.

[0074] Optionally, a dual-polarized water index is calculated based on JRC global surface water product and Sentinel-1 SAR data. An initial water body mask is generated by determining the threshold using the water body frequency method. By fusing Sentinel-1 dual-polarized data (VV+VH), the distinction between water bodies and vegetation / bare soil is enhanced (improving water body identification rate by 20% compared to the traditional NDWI index). The water body frequency method is used to exclude seasonal water bodies (such as post-rainwater accumulation) through time-series statistics, ensuring that the baseline map retains only stable inland ponds. Subsequently, Sentinel-2 data is fused to optimize the mask boundaries, forming a watershed-scale inland pond baseline map. Sentinel-2 data is further optimized by adding a 10m resolution visible light band to refine pond boundaries (eliminating misclassifications of rivers / lakes), providing a clean spatial mask for classification.

[0075] Step S140 involves applying differentiated combinations of technologies to aquaculture ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds, based on their unique characteristics. This results in the identification and classification of multiple categories of inland ponds. The differentiated combinations of technologies include: using object-oriented analysis combined with geometric feature decision trees and neighborhood space discrimination to identify aquaculture ponds; using land cover overlay analysis and buffer ratio verification to identify urban ponds; using random forest algorithm to train power station samples to identify photovoltaic ponds; constructing a convolutional neural network model combining spectral texture features to identify tailings ponds; and applying digital elevation models and neighborhood analysis of cultivated land data to identify farmland ponds.

[0076] Based on literature and previous research, this paper summarizes the main classification categories and descriptions of inland ponds at the watershed scale. Pond categories include: aquaculture ponds (fish ponds, shrimp and crab ponds), urban ponds, photovoltaic ponds, tailings ponds, farmland ponds, and other ponds. Specifically, aquaculture ponds: fish ponds are used for fish farming, are regularly shaped, usually concentrated, and arranged in a regular pattern with uniform water content. Shrimp and crab ponds are mainly used for shrimp and crab farming, are regularly shaped, usually concentrated, and arranged in a regular pattern, with plant cultivation within the ponds; urban ponds: located within urban boundaries, usually independently distributed, serving a landscaping function; photovoltaic ponds: photovoltaic modules are located on the pond surface; tailings ponds: ponds in mining areas, storing mining waste, at high altitudes, with high heavy metal content and unique spectral characteristics; farmland ponds: located around farmland, scattered; other ponds: mainly natural, abandoned, or unused ponds with minimal human interference, but excluding the categories specified above.

[0077] A basin-scale framework for identifying and classifying inland ponds is constructed, utilizing multi-source data and multiple machine learning algorithms for pond identification and classification. For example, spatial analysis methods from geography (overlay analysis, buffer analysis, neighborhood analysis) are combined with various machine learning algorithms such as decision trees, random forests (RF), and convolutional neural networks (CNN). Optimal identification and classification results are obtained based on the characteristics of each pond type, and accuracy analysis is performed. Optionally, for five target categories—aquaculture ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds—the following methods are dynamically selected and combined based on the morphology, location, and spectral texture characteristics of each type of pond: spatial neighborhood analysis and geometric feature decision trees, land cover overlay and buffer ratio verification, machine learning classification model training, and deep learning feature extraction. By integrating parallel classification frameworks for aquaculture ponds (fish ponds / shrimp / crab ponds), urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds, collaborative pond identification is achieved. To address the regular geometric characteristics of aquaculture ponds, a geometric feature decision tree (e.g., area, shape index) and a spatial neighborhood discrimination method are constructed. The geometric feature decision tree identifies regular shapes and excludes irregular boundaries of natural water bodies. The neighborhood spatial discrimination method (buffer zone expansion + intersection test) filters isolated water bodies. NDPI (plankton enrichment index) and NDVI (submarine vegetation index) are combined to distinguish between fishponds (NDPI > 0.2) and shrimp / crab ponds (NDVI > 0.3), improving classification accuracy. For urban ponds, land cover overlay analysis and buffer zone ratio verification are employed. Urban boundary data + SinoLC-1 overlay quickly locates urban water bodies, and accuracy is improved through buffer zone impermeable surface ratio verification. For photovoltaic ponds, a random forest model is used to train photovoltaic power station samples, learning the spectral characteristics of photovoltaic panels and water surface coexistence, reducing processing time while maintaining high accuracy. For tailings ponds, the convolutional neural network model is input with DEM elevation (tailings ponds are located in steep slope areas), Sentinel-2 spectra (heavy metal pollution causes a sharp drop in short-wave infrared light), and texture features (mud-like texture). The model automatically extracts pollution markers through convolutional layers (32 / 16 filter banks), improving the recognition rate from 30% to 64%. For farmland ponds, the SinoLC-1 cultivated land layer is overlaid with non-urban boundaries, combined with a 50m buffer analysis to ensure the ponds are completely surrounded by farmland, achieving rapid identification without samples, with an efficiency improvement of 70% compared to supervised classification.

[0078] This embodiment of the inland pond classification method at the watershed scale significantly improves the detection rate of small-scale ponds by fusing multi-source geospatial data and eliminating interference from clouds and water turbidity. It achieves accurate classification through a combination of differentiated algorithms. For aquaculture ponds, it employs a geometric feature decision tree combined with neighborhood spatial discrimination, supplemented by NDPI (plankton enrichment) and NDVI (submarine vegetation) to distinguish between fish ponds and shrimp / crab ponds, achieving high accuracy. For urban ponds, it uses land cover overlay and impermeable surface ratio verification to avoid misclassification. For photovoltaic ponds, it uses a lightweight random forest model to capture the spectral characteristics of the water surface and photovoltaic panels coexisting, achieving high accuracy and short processing time. For tailings ponds, it constructs a CNN to fuse spectral and texture features, improving the recognition rate compared to traditional methods. For farmland ponds, it achieves unsupervised and efficient identification based on digital elevation models and neighborhood analysis of cultivated land data, improving efficiency compared to supervised methods. Ultimately, the overall classification accuracy is significantly improved compared to single algorithms, while the processing time is significantly reduced, achieving high-precision identification and classification of multiple types of ponds at the watershed scale.

[0079] like Figure 3 As shown, in one embodiment, object-oriented analysis combined with geometric feature decision trees and neighborhood space discrimination is used to identify aquaculture ponds, including the following steps:

[0080] Step S310: Use global city boundary data to delineate non-urban areas and overlay baseline maps to extract the coarse range of aquaculture ponds.

[0081] By using global city boundary data to delineate the non-urban boundaries of the target area, and combining this with a watershed-scale inland pond baseline map, a rough estimate of the aquaculture pond range is extracted.

[0082] Step S320: Based on Sentinel-2 imagery and field sampling data, construct a geometric feature decision tree to screen aquaculture ponds.

[0083] Using Sentinel-2 satellite remote sensing imagery combined with statistical data from on-site aquaculture pond sampling, multiple geometric features were selected as sensitive parameters for aquaculture ponds, and threshold values ​​for these geometric parameters were determined. A decision tree for pond selection was constructed using these geometric features and parameters to screen aquaculture ponds. Pond selection was based on regular shape, area threshold (>0.1 hectares), and clustering degree (neighborhood pond density >5 / km²), thus addressing misclassification caused by irregular shapes in natural water bodies.

[0084] Optionally, a geometric feature decision tree is constructed to screen aquaculture ponds, including: generating buffer zones for potential ponds using the neighborhood expansion method, and identifying isolated non-aquaculture ponds based on the intersection of the buffer zones; and distinguishing fish ponds from shrimp and crab ponds using the Normalized Difference Pond Index (NDPI) and the Normalized Difference Vegetation Index (NDVI). The neighborhood expansion method, using the intersection of buffer zones to identify isolated non-aquaculture ponds, achieves an exclusion rate of 93%. NDPI enhances the extraction of phytoplankton-rich water bodies (fish ponds), and NDVI detects underwater vegetation (a core feature of shrimp and crab ponds), resulting in a fish pond / shrimp and crab pond differentiation accuracy of 87.67% (a 35% improvement over single-spectrum classification).

[0085] In the classification of aquaculture ponds, see Figure 4 Object-oriented classification decision trees have the potential to address the need for accurate identification of aquaculture ponds and can serve as an effective support for classical mapping techniques. Global urban boundary data was used to delineate the non-urban boundaries of the target area, and combined with inland pond baseline maps at the watershed scale, a rough range of aquaculture ponds was extracted. Sentinel-2 satellite remote sensing imagery combined with statistical data from on-site aquaculture pond sampling was used to select multiple geometric features as sensitive parameters for the aquaculture ponds and determine the thresholds for these geometric parameters. A decision tree for pond selection was constructed using these geometric features and parameters to screen aquaculture ponds. Since aquaculture ponds are typically concentrated in low-altitude and flat rural areas, a neighborhood-based expansion method and a spatial discrimination method were used. All potential aquaculture ponds were expanded with a buffer of a certain size and then superimposed based on whether they intersected with other buffers. If there was no intersection, it was identified as an isolated non-aquaculture pond. To further differentiate between fish ponds and shrimp / crab ponds, two indices were used: the Normalized Difference Pond Index (NDPI) and the Normalized Difference Vegetation Index (NDVI). NDPI reflects the status of surface plankton and flora, enhancing the extraction effect of aquaculture ponds with low and abundant plankton mobility. More importantly, shrimp and crab ponds typically have a large amount of underwater vegetation (distinguished by NDVI) to facilitate molting and avoid predators, which is significantly different from fish ponds.

[0086] like Figure 5 As shown, in one embodiment, urban ponds are identified using land cover overlay analysis and buffer ratio verification, including the following steps:

[0087] Step S510: Overlay global city boundary data with the SinoLC-1 city class layer onto the baseline map to filter the candidate set of city ponds. By overlaying city boundaries, non-urban water bodies are directly excluded, and the candidate set of city ponds is narrowed down.

[0088] Step S520: The independence is verified by using a neighborhood-based expansion method and confirmed a second time by the proportion of impermeable surface in the buffer zone.

[0089] The accuracy of the method was improved to 75.75% by using the impermeable surface ratio verification and calculating the proportion of the surrounding impermeable surface through buffer zone analysis (>60% is judged as urban landscape water body).

[0090] In the classification of urban ponds, urban ponds are considered important resources for environmental aesthetics, ecological maintenance, environmental quality protection, and public health. A three-step classification approach is employed to identify urban ponds. In the first step, urban boundaries and the SinoLC-1 map are overlaid on a watershed-scale inland pond baseline map to filter urban ponds. Secondly, a neighborhood-based expansion method is used to ensure isolation. The final step involves expanding these potential ponds into a buffer zone of a certain size and then calculating the proportion of different land cover categories within the intersection area. If a large proportion of impervious surfaces or grasslands exist in the overlapping area, it is identified as an urban pond.

[0091] like Figure 6 As shown, in one embodiment, a random forest algorithm is used to train power station samples to identify photovoltaic ponds, including the following steps:

[0092] Step S610: Collect sample points of photovoltaic power stations in the target area. 1000 photovoltaic power station sample points are used to cover different regions and scales of scenarios, addressing the sample bias issue.

[0093] Step S620: Set the parameters of the random forest model, train and test the model. A lightweight random forest model (number of trees = 20) is used, which improves training efficiency by 5 times compared to CNN while maintaining 87% accuracy, meeting the needs of rapid mapping at the watershed scale.

[0094] Step S630: Overlay the model output with the baseline map to determine the distribution of photovoltaic ponds.

[0095] We collected sample points of photovoltaic power stations in the target study area, set the parameters of the random forest model, and trained and tested it to obtain the final results. Combined with the baseline map of inland ponds at the watershed scale, we determined the accurate distribution of photovoltaic ponds.

[0096] See the photovoltaic pond classification section. Figure 7 The innovative fishery-photovoltaic model involves installing photovoltaic modules on the water surface to generate electricity using solar energy while conducting aquaculture activities. Random Forest (RF) is an ensemble classifier that uses a set of decision trees to predict classification or regression, offering advantages such as high accuracy, high efficiency, and good stability. RF is sensitive to sampling design; suitable training samples are crucial for the classification accuracy and stability of the RF model. The samples were labeled as photovoltaic power stations, and 1000 sample points were collected. The number of trees was set to 20, and other parameters remained at their default values. The constructed RF model was trained and tested to obtain the final results.

[0097] like Figure 8 As shown, in one embodiment, constructing a convolutional neural network model that incorporates spectral texture features to identify tailings ponds includes the following steps:

[0098] Step S810: Delineate the target area of ​​the tailings pool based on the DEM, SinoLC-1 bare map layer, and global mining area dataset.

[0099] Step S820: Construct a CNN model, input spectral and texture features, and train it to identify tailings pools through convolutional layers (32 / 16 filters), max pooling layers, and fully connected layers.

[0100] By combining spectral and textural inputs, and utilizing the spectral anomalies caused by heavy metal pollution (sudden drop in shortwave infrared reflectance) and the unique texture of tailings sludge (CNN automatically extracts mottled features), the method resolves the confusion between tailings ponds and ordinary bare ponds. Employing a multi-layer convolutional structure, the method captures multi-scale features through a 32 / 16 filter bank, and the max-pooling layer enhances spatial pattern recognition in polluted areas, thereby increasing the detection rate of tailings ponds from 30% in traditional methods to 64%.

[0101] In tailings pool classification, a CNN model combining spectral and texture features is established based on target tailings pool areas obtained from DEM, SinoLC-1 (bare land), and global mining datasets to further identify tailings pools. See also Figure 9 The CNN architecture consists of two convolutional layers, one max-pooling layer, and three fully connected layers. Using a segmented image with a pixel size of 350×150, inputs of different sizes are fed into the two convolutional layers, with filter sizes of 32 and 16, and a kernel size of 3×3. The max-pooling layer extracts features by summing the maximum values ​​of the inputs to the convolutional filters. After feature extraction by the convolutional layers, the pooling layer downsamples the features, and then a subset of these features is mapped to the CNN output through a flattening layer and a fully connected layer. Finally, after the three fully connected layers are applied, the network output is obtained by mapping the features to a 1×1 size tailings sample.

[0102] like Figure 10 As shown, in one embodiment, the identification of farmland ponds using a digital elevation model and neighborhood analysis of cultivated land data includes the following steps:

[0103] Step S1010: Overlay the DEM, SinoLC-1 cultivated land layer, and non-urban boundary data onto the baseline map. By overlaying the SinoLC-1 cultivated land layer and the non-urban boundary data, it is ensured that the pond is entirely within the farmland area.

[0104] Step S1020: Apply neighborhood analysis to verify whether the ponds are independently distributed within the farmland. Using the neighborhood expansion method, the independence of the ponds is tested (distance from the nearest water body > 50m), excluding linear water bodies such as irrigation ditches, achieving rapid identification without sample labeling (70% more efficient than machine learning).

[0105] In the classification of farmland ponds, farmland ponds are generally widely distributed, scattered around low-lying, flat farmland, and mainly used for irrigation and drainage. Therefore, we utilize DEM, SinoLC-1 (cultivated land), and non-urban boundary data, while simultaneously overlaying and analyzing inland pond baseline maps at the watershed scale. We apply the aforementioned domain analysis to examine whether the ponds are independently located within farmland, thereby identifying farmland ponds.

[0106] All ponds outside the categories of aquaculture ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds are considered "other ponds." These may be unused or abandoned ponds, or ponds with multiple functions (whose characteristics are not obvious or easily confused), and therefore need to be listed separately.

[0107] After classifying inland ponds into multiple categories using differentiated combination applications, spatial distribution maps of these ponds are output, and confusion matrices are constructed for fish ponds, shrimp and crab ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds. The accuracy is quantified using the confusion matrices, with each of the six pond categories assessed independently (e.g., fish ponds have an overall accuracy of 91.50% and a Kappa coefficient of 79.27%), identifying the main misclassification types (e.g., tailings ponds mistakenly identified as bare land).

[0108] The accuracy of the framework was assessed using the Nash coefficient (NS), correlation coefficient (R²), root mean square error (RMSE), and mean absolute error (MAE). Cross-validation with hydrological indicators was employed, and the spatial distribution rationality was verified using the Nash coefficient and correlation coefficient. The RMSE constrained the area estimation error, ensuring that the results could directly support the quantitative management of watershed water resources.

[0109] Specifically, confusion matrices were established for fishponds, shrimp and crab ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds. The accuracy of the watershed-scale inland pond identification and classification framework was analyzed, with indicators including: Nash coefficient (NS) and correlation coefficient (R²). 2 ), root mean square error (RMSE) and mean absolute error (MAE).

[0110] Among them, the overall accuracy of fish ponds, shrimp and crab ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds were 91.50%, 87.67%, 75.75%, 87.00%, 64.00%, and 62.00%, respectively; the Kappa coefficients (%) were 79.27, 75.34, 48.27, 41.23, 27.42, and 10.02, respectively. The Nash coefficient (NS) and correlation coefficient (R) of the watershed-scale inland pond identification and classification framework were also discussed. 2The root mean square error (RMSE) and mean absolute error (MAE) are 0.96, 0.96, 9.91, and 8.83, respectively. The development, verification, and practical application of this invention demonstrate that the estimation of inland ponds at the watershed scale is accurate and up-to-date.

[0111] The aforementioned watershed-scale inland pond classification method employs multi-source data collaboration. By fusing Sentinel-1SAR (which penetrates clouds and is unaffected by illumination) and Sentinel-2 optical data (high spectral resolution), combined with DEM topographic constraints, it significantly improves the detection capability of small-scale ponds and overcomes the shortcomings of traditional single optical data being susceptible to environmental interference. A refined processing workflow is adopted, using time-series median synthesis to eliminate transient noise, and overlaying JRC water body products with localized Sentinel data to optimize the baseline map, ensuring that the water body mask accuracy meets the requirements of watershed-scale applications, thereby breaking through the bottleneck of accuracy identification.

[0112] This study is the first to integrate a parallel classification framework for aquaculture ponds (fish ponds / shrimp and crab ponds), urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds, filling the gap in existing technologies that only focus on a single type. For the regular geometric characteristics of aquaculture ponds, a geometric feature decision tree (such as area and shape index) and a spatial neighborhood discrimination method are constructed, using NDPI / NDVI to distinguish between fish ponds and shrimp and crab ponds (with an accuracy of 87.67%). Based on the locational characteristics of urban ponds, the spectral confusion problem is solved by overlaying land cover data and verifying the proportion of impermeable surfaces in buffer zones. For tailings ponds with complex spectral textures, CNNs are used to automatically extract deep features (such as special reflection patterns caused by heavy metal pollution), improving the recognition rate of difficult-to-distinguish targets, thus establishing a comprehensive classification system.

[0113] A differentiated model deployment is adopted, selecting the optimal algorithm combination based on the characteristics of pond type. For example, Random Forest (RF) is adapted to the sample-driven characteristics of photovoltaic ponds, and CNN is adapted to the texture complexity of tailings ponds, avoiding the insufficient generalization ability of a single algorithm. Spatial analysis is used to enhance decision-making by introducing geospatial methods such as neighborhood analysis and buffer validation to strengthen the utilization of contextual information (such as the independent distribution determination of farmland ponds) and solve the spatial misjudgment problem of pixel-level classification. A full-process verification mechanism is adopted, using confusion matrix (accuracy of 62%~91% for each type) and hydrological model indicators (NS=0.96, R²=0.96) for dual verification to ensure the robustness of the framework, thereby achieving performance optimization through multi-algorithm collaboration.

[0114] This invention enables high-precision automated mapping of five major types of inland ponds (aquaculture / urban / photovoltaic / tailings / farmland) at the watershed scale, with an overall average accuracy of 81.18% (32% higher than a single algorithm) and a 50% reduction in classification time, providing standardized technical support for the management of small and micro water resources.

[0115] The watershed-scale inland pond classification device provided by the present invention will be described below. The watershed-scale inland pond classification device described below can be referred to in correspondence with the watershed-scale inland pond classification method described above.

[0116] like Figure 11 As shown, in one embodiment, an inland pond classification device at the watershed scale includes a multi-source data acquisition module 1110, a data preprocessing module 1120, a water body baseline map generation module 1130, and a differential classification module 1140.

[0117] The multi-source data acquisition module 1110 is used to acquire multi-source geospatial data of the target watershed, including remote sensing images, digital elevation models, land cover data, JRC global surface water products, urban boundary data, photovoltaic power station distribution data, mining area polygon data, lake datasets, and statistical yearbook data.

[0118] The data preprocessing module 1120 is used to perform radiometric calibration, topographic correction, noise removal, and spatial cropping on remote sensing images.

[0119] The water body baseline map generation module 1130 is used to generate inland pond water body baseline maps based on multi-source geospatial data.

[0120] The differentiated classification module 1140 is used to perform differentiated combined applications based on the differentiated characteristics of aquaculture ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds, and output the identification and classification results of multiple categories of inland ponds. The differentiated combined applications include: using object-oriented analysis combined with geometric feature decision trees and neighborhood space discrimination to identify aquaculture ponds; using land cover overlay analysis and buffer ratio verification to identify urban ponds; using random forest algorithm to train power station samples to identify photovoltaic ponds; constructing a convolutional neural network model combined with spectral texture features to identify tailings ponds; and applying digital elevation model and neighborhood analysis of cultivated land data to identify farmland ponds.

[0121] Figure 12 This example illustrates a schematic diagram of the physical structure of an electronic device, which can be a smart terminal. Its internal structure diagram can be as follows: Figure 12 As shown. The electronic device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a watershed-scale inland pond classification method, which includes:

[0122] Acquire multi-source geospatial data for the target watershed, including remote sensing imagery, digital elevation models, land cover data, JRC global surface water products, urban boundary data, photovoltaic power station distribution data, mining area polygon data, lake datasets, and statistical yearbook data;

[0123] Radiometric calibration, topographic correction, noise removal, and spatial cropping of remote sensing images;

[0124] Baseline maps of inland pond water bodies are generated based on multi-source geospatial data;

[0125] Based on the differentiated characteristics of aquaculture ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds, differentiated combined applications are applied to output identification and classification results for multiple categories of inland ponds. The differentiated combined applications include: using object-oriented analysis combined with geometric feature decision trees and neighborhood space discrimination to identify aquaculture ponds; using land cover overlay analysis and buffer ratio verification to identify urban ponds; using random forest algorithm to train power station samples to identify photovoltaic ponds; constructing a convolutional neural network model combined with spectral texture features to identify tailings ponds; and applying digital elevation model and neighborhood analysis of cultivated land data to identify farmland ponds.

[0126] Those skilled in the art will understand that the structures shown in the inland pond classification at the watershed scale are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the electronic devices to which the present invention is applied. Specific electronic devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0127] On the other hand, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements a watershed-scale inland pond classification method, the method comprising:

[0128] Acquire multi-source geospatial data for the target watershed, including remote sensing imagery, digital elevation models, land cover data, JRC global surface water products, urban boundary data, photovoltaic power station distribution data, mining area polygon data, lake datasets, and statistical yearbook data;

[0129] Radiometric calibration, topographic correction, noise removal, and spatial cropping of remote sensing images;

[0130] Baseline maps of inland pond water bodies are generated based on multi-source geospatial data;

[0131] Based on the differentiated characteristics of aquaculture ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds, differentiated combined applications are applied to output identification and classification results for multiple categories of inland ponds. The differentiated combined applications include: using object-oriented analysis combined with geometric feature decision trees and neighborhood space discrimination to identify aquaculture ponds; using land cover overlay analysis and buffer ratio verification to identify urban ponds; using random forest algorithm to train power station samples to identify photovoltaic ponds; constructing a convolutional neural network model combined with spectral texture features to identify tailings ponds; and applying digital elevation model and neighborhood analysis of cultivated land data to identify farmland ponds.

[0132] In another aspect, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium, and when the processor executes the computer instructions, implements a watershed-scale inland pond classification method, the method comprising:

[0133] Acquire multi-source geospatial data for the target watershed, including remote sensing imagery, digital elevation models, land cover data, JRC global surface water products, urban boundary data, photovoltaic power station distribution data, mining area polygon data, lake datasets, and statistical yearbook data;

[0134] Radiometric calibration, topographic correction, noise removal, and spatial cropping of remote sensing images;

[0135] Baseline maps of inland pond water bodies are generated based on multi-source geospatial data;

[0136] Based on the differentiated characteristics of aquaculture ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds, differentiated combined applications are applied to output identification and classification results for multiple categories of inland ponds. The differentiated combined applications include: using object-oriented analysis combined with geometric feature decision trees and neighborhood space discrimination to identify aquaculture ponds; using land cover overlay analysis and buffer ratio verification to identify urban ponds; using random forest algorithm to train power station samples to identify photovoltaic ponds; constructing a convolutional neural network model combined with spectral texture features to identify tailings ponds; and applying digital elevation model and neighborhood analysis of cultivated land data to identify farmland ponds.

[0137] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.

[0138] By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0139] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0140] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for classifying inland ponds at the watershed scale, characterized in that, The method includes: Acquire multi-source geospatial data for the target watershed, including remote sensing imagery, digital elevation models, land cover data, JRC global surface water products, global city boundary data, photovoltaic power plant distribution data, mining area polygon data, lake datasets, and statistical yearbook data; Radiometric calibration, topographic correction, noise removal, and spatial cropping of remote sensing images; The process of generating inland pond water baseline maps based on multi-source geospatial data includes: calculating the dual-polarized water index based on JRC global surface water product and Sentinel-1 SAR data, determining the threshold using the water frequency method to generate an initial water mask, and optimizing the mask boundary by fusing Sentinel-2 data to form a watershed-scale inland pond baseline map. Based on the baseline map of inland ponds, differentiated combined applications are applied to identify and classify inland ponds of various types, including aquaculture ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds, taking into account their different characteristics. These differentiated combined applications include: using object-oriented analysis combined with geometric feature decision trees and neighborhood space discrimination to identify aquaculture ponds; using land cover overlay analysis and buffer ratio verification to identify urban ponds; using random forest algorithm to train photovoltaic power station samples to identify photovoltaic ponds; constructing a convolutional neural network model combining spectral texture features to identify tailings ponds; and applying digital elevation models and neighborhood analysis of cultivated land data to identify farmland ponds.

2. The method for classifying inland ponds at the watershed scale according to claim 1, characterized in that, The radiometric calibration, topographic correction, noise removal, and spatial cropping of remote sensing images include: Calculate the median backscattering from Sentinel-1 SAR multi-year sequence data to eliminate salt-and-pepper noise; Radiometric calibration and topographic correction were performed on the Sentinel-2 imagery, and the images were cropped according to the study area boundaries.

3. The method for classifying inland ponds at the watershed scale according to claim 1, characterized in that, The method of identifying aquaculture ponds using object-oriented analysis combined with geometric feature decision trees and neighborhood space discrimination includes: Global city boundary data was used to delineate non-urban areas, and baseline maps were overlaid to extract the coarse extent of aquaculture ponds. Based on Sentinel-2 imagery and field sampling data, a geometric feature decision tree was constructed to screen aquaculture ponds.

4. The method for classifying inland ponds at the watershed scale according to claim 3, characterized in that, The construction of a geometric feature decision tree for screening aquaculture ponds includes: The neighborhood expansion method is used to generate buffer zones for potential ponds, and isolated non-aquaculture ponds are identified based on the intersection of the buffer zones. Fish ponds and shrimp / crab ponds are distinguished by the normalized difference index and the normalized difference vegetation index.

5. The method for classifying inland ponds at the watershed scale according to claim 1, characterized in that, The method of identifying urban ponds using land cover overlay analysis and buffer ratio verification includes: Global city boundary data and SinoLC-1 city class layers are overlaid on the baseline map to filter candidate sets of city ponds; Independence was verified using a neighborhood-based expansion method and further confirmed by the proportion of impermeable surfaces in the buffer zone.

6. The method for classifying inland ponds at the watershed scale according to claim 1, characterized in that, The method of using a random forest algorithm to train photovoltaic power station samples to identify photovoltaic ponds includes: Collect sample points of photovoltaic power stations in the target area; Set the parameters for the random forest model, train and test the model; The distribution of photovoltaic ponds is determined by overlaying the model output with the baseline map.

7. The method for classifying inland ponds at the watershed scale according to claim 1, characterized in that, The construction of a convolutional neural network model combining spectral texture features for identifying tailings pools includes: The target area of ​​the tailings pond was delineated based on the digital elevation model, the SinoLC-1 bare map layer, and the global mining area dataset. A CNN model is constructed, inputting spectral and texture features, and trained to identify tailings pools through convolutional layers, max pooling layers, and fully connected layers.

8. The method for classifying inland ponds at the watershed scale according to claim 1, characterized in that, The application of digital elevation model and neighborhood analysis of cultivated land data to identify farmland ponds includes: Overlay digital elevation model, SinoLC-1 cultivated map layer and non-urban boundary data onto baseline map; Neighborhood analysis was used to verify whether the ponds were independently distributed in the farmland.

9. The method for classifying inland ponds at the watershed scale according to any one of claims 1 to 8, characterized in that, The method further includes: Output spatial distribution maps of inland ponds of multiple categories, and construct confusion matrices for fish ponds, shrimp and crab ponds, urban ponds, photovoltaic ponds, tailings ponds, and farmland ponds; The accuracy of the framework was evaluated using the Nash coefficient (NS), correlation coefficient (R²), root mean square error (RMSE), and mean absolute error (MAE).

Images