A binary classification precision evaluation method and system based on boundary dissimilarity area and a storage medium

Abstract

The application belongs to the technical field of resource and environment analysis, and particularly relates to a binary classification precision evaluation method based on boundary different area, which comprises the following steps: S1, obtaining binary classification results of a target area to obtain extracted classification boundary vector data; S2, obtaining reference classification boundary vector data of the same target area; S3, determining a boundary strip evaluation area based on the reference boundary and the extracted boundary; S4, counting the area of classification inconsistent pixels in the boundary strip evaluation area; S5, calculating a boundary different area index BDAI; and S6, determining the extraction error type according to the proportional relationship between the false extraction area and the missed extraction area. The application quantifies the influence of boundary geometric errors on classification consequences to distinguish the reliability difference of different binary classification methods in boundary processing, and realizes quantitative evaluation of boundary precision.

Description

Technical Field

[0001] This invention belongs to the field of resource and environmental analysis technology, specifically relating to a binary classification accuracy evaluation method, system, and storage medium based on the area of ​​distinct boundaries. Background Technology

[0002] Traditional remote sensing classification accuracy evaluation methods primarily rely on pixel-level statistical indicators, such as overall accuracy (OA), Kappa coefficient, and F1 score. These indicators measure the overall performance of the classification task by comparing the matching degree between the classification results and reference samples. These methods typically use random sampling or manual point selection to obtain validation samples, focusing only on whether pixels are correctly classified as target or non-target features. However, ground features, as typical geographic objects, not only possess category attributes but also carry clearly defined shape and boundary information. The lack of spatial location information in existing evaluation systems makes it difficult to comprehensively reflect the accuracy of classification boundary extraction, which is precisely the key to measuring the ability of different algorithms to characterize ground feature details.

[0003] Existing technologies suffer from three core flaws: First, traditional methods focus solely on pixel classification accuracy, neglecting the spatial location information of land cover boundaries, making it impossible to distinguish the reliability differences of algorithms in boundary processing. Second, the impact of classification boundary errors (such as length scaling and spatial offset) on classification results is not systematically quantified; for example, the differences in mixed pixels at the land-water interface are difficult to capture using traditional indicators. Third, existing evaluation systems lack universality, failing to compare classification methods across resolutions and land cover types, and unable to pinpoint the types of algorithm errors (such as over-segmentation, over-smoothing, or boundary offset), leading to ambiguity in optimization direction. These problems make it difficult for traditional methods to meet the refined requirements for classification boundary accuracy in practical applications.

[0004] This invention uses the BDAI index to quantitatively assess the area of ​​dissimilar boundaries. The technical challenge lies in accurately delineating the boundary strip regions and statistically analyzing the areas of pixels with inconsistent classifications. For example, the strip regions must naturally form between the extracted boundary and the reference boundary, rather than relying on a preset buffer distance. This poses a challenge to adaptability to complex terrains (such as islands and nested polygons). Furthermore, the statistical analysis of misextracted and missed areas requires pixel-by-pixel comparison and normalization based on image resolution and the area of ​​the study region to eliminate the influence of scale differences. In addition, error type determination requires dynamic analysis of the misextracted / missed ratio and must consider the characteristics of different land cover classification tasks (such as water bodies, vegetation, and snow), further increasing the complexity of the technical implementation. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides a binary classification accuracy evaluation method based on dissimilar boundary areas. This method addresses how to quantify the impact of boundary geometric errors on classification outcomes, thereby distinguishing the reliability differences in boundary handling among different binary classification methods.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] One of the objectives of this invention is to provide a binary classification accuracy evaluation method based on the area of ​​dissimilar boundaries, which can achieve a quantitative assessment of boundary accuracy.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A binary classification accuracy evaluation method based on the area of ​​distinct boundaries includes the following steps:

[0010] S1. Obtain the binary classification result of the target region and extract the classification boundary vector data;

[0011] S2. Obtain reference classification boundary vector data for the same target region;

[0012] S3. Determine the boundary strip evaluation area based on the reference boundary and the extracted boundary;

[0013] S4. Statistically measure the area of ​​inconsistent pixels within the boundary band evaluation area to identify mis-extracted and under-extracted areas.

[0014] S5. Calculate the boundary dissimilarity area index (BDAI).

[0015] S6. Determine the type of extraction error based on the ratio of the incorrectly extracted area to the missed extracted area.

[0016] Further optimization of the design results in the binary classification, which divides the land surface into two categories: target features and non-target features.

[0017] Further optimization of the design involves obtaining the reference classification boundary vector data through manual visual interpretation of high-resolution remote sensing images. Specifically, this includes selecting images whose imaging date and image extraction time differ by no more than 30 days and whose cloud cover is less than 5%.

[0018] In a further optimized design, the boundary strip evaluation area is defined by the extraction boundary and the reference boundary. Specifically, it includes the area where the extraction boundary extends outward to the reference boundary, the area where the reference boundary extends outward to the extraction boundary, and the area formed by their intersection and overlap. This area does not require a preset buffer distance but is naturally formed by the actual spatial relationship.

[0019] Further design optimization involves calculating the area of ​​inconsistent pixels by comparing the extracted results with the reference true value in raster data pixel by pixel, using a confusion matrix classification statistical method to calculate the misextracted area and the missed area respectively.

[0020] Further optimizing the design, the formula for calculating the boundary dissimilarity area index (BDAI) is as follows:

[0021]

[0022] Where Acommission is the area of ​​incorrect extraction, Aomission is the area of ​​missed extraction, R is the spatial resolution of the image, and S is the area of ​​the study region.

[0023] The design is further optimized so that the error type is determined by comparing the ratio of the incorrectly extracted area to the missed extracted area:

[0024] If the area of ​​incorrect extraction is significantly larger than the area of ​​missing extraction, it is judged as an oversegmentation bias.

[0025] If the area of ​​missed extraction is significantly larger than the area of ​​incorrect extraction, it is judged as an oversmoothing bias.

[0026] If the two are comparable, it is determined to be a boundary offset type deviation.

[0027] With further design optimization, the method is applicable to various land cover classification tasks, including but not limited to water body extraction, vegetation identification, and land use mapping. Only the reference boundary and extraction boundary need to be input to achieve a quantitative assessment of the classification results.

[0028] The second objective of this invention is to provide a binary classification accuracy evaluation system based on the area of ​​distinct boundaries, for running the aforementioned binary classification accuracy evaluation method based on the area of ​​distinct boundaries on a computer.

[0029] To achieve the above objectives, the technical solution of the present invention is as follows:

[0030] A binary classification accuracy evaluation system based on the area of ​​distinct boundaries includes a data acquisition module, a region division module, a statistical analysis module, and an error determination module.

[0031] The data acquisition module is configured to: acquire the binary classification result of the target region and obtain the extracted classification boundary vector data, and acquire the reference classification boundary vector data of the same target region;

[0032] The region division module is configured to determine the boundary strip evaluation region based on the reference boundary and the extracted boundary.

[0033] The statistical analysis module is configured to: statistically analyze the area of ​​inconsistent pixels within the boundary strip evaluation region and calculate the boundary dissimilarity area index;

[0034] The error determination module is configured to determine the extraction error type based on the ratio of the misextracted area to the missed area.

[0035] A third objective of this invention is to provide a computer-readable storage medium for implementing the aforementioned binary classification accuracy evaluation method based on the area of ​​dissimilar boundaries.

[0036] To achieve the above objectives, the technical solution of the present invention is as follows:

[0037] A computer-readable storage medium storing program instructions that, when executed by a processor, are used to implement the binary classification accuracy evaluation method and optimization design based on boundary dissimilar area as described in one of the objectives of this invention.

[0038] Compared with the prior art, the present invention has the following advantages:

[0039] 1. This invention defines a strip-shaped region between the extraction boundary and the reference boundary, and statistically analyzes the misextracted area and the missed area pixel by pixel. This makes the impact of boundary geometric error on the classification result explicit, breaking through the neglect of spatial location information by traditional pixel statistical indicators, and achieving accurate characterization of classification boundary precision.

[0040] 2. This invention quantifies the area density of inconsistent pixels by using the BDAI index, and combines the ratio of mis-extracted to under-extracted areas to distinguish error types such as over-segmentation, over-smoothing, or boundary offset, providing a clear direction for error diagnosis for algorithm optimization.

[0041] 3. This invention does not rely on specific land cover types or preset parameters. It only requires inputting reference boundaries and extraction boundaries to achieve quantitative evaluation of classification results. It is applicable to binary classification tasks of various land cover types such as water bodies, vegetation, and construction land.

[0042] 4. This invention eliminates the influence of images of different resolutions on the evaluation results by performing resolution normalization processing in the BDAI formula, and supports the comparison of classification methods between multi-scale and multi-resolution images.

[0043] 5. By analyzing the magnitude of BDAI values ​​and the ratio of false positives to false negatives, this invention can accurately pinpoint the error type of the classification method, providing direct technical basis for algorithm improvement and enhancing the reliability and spatial consistency of classification results. Attached Figure Description

[0044] Figure 1A flowchart of a binary classification accuracy evaluation method based on the area of ​​dissimilar boundaries is provided for an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of BDAI in a binary classification accuracy evaluation method based on dissimilar boundary area provided in an embodiment of the present invention. NDWI is used to extract the water body boundary and the reference water body boundary interpreted from high-resolution remote sensing images on Google Earth.

[0046] Figure 3 Box plots of the BDAI results of various indices in a binary classification accuracy evaluation method based on the area of ​​dissimilar boundaries are provided in this embodiment of the invention.

[0047] Figures 4-5 This invention provides a comparison diagram of the extraction boundary of typical index results and the boundary of a reference water body in a binary classification accuracy evaluation method based on the area of ​​dissimilar boundaries.

[0048] Figure 6 This invention provides a functional module framework diagram of a binary classification accuracy evaluation system based on the area of ​​distinct boundaries. Detailed Implementation

[0049] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0050] To better explain the technical solution, the following technical terms are explained:

[0051] 1. Binary Classification

[0052] Definition: A classification task that divides ground features or targets into two categories, such as water / non-water, vegetation / non-vegetation, buildings / non-buildings, etc. Technical Application: The core application scenario of this invention is to evaluate classification accuracy by quantifying boundary errors.

[0053] 2. Boundary Discrepancy Area Index (BDAI)

[0054] Definition: A precision evaluation index based on the area of ​​regions with distinct boundaries, the formula is:

[0055] BDAI = (Acommission + Aomission) / (R² × S);

[0056] Where Acommission is the area of ​​incorrect extraction, Aomission is the area of ​​missed extraction, R is the image spatial resolution, and S is the area of ​​the study region. Technical function: To quantify the absolute value of boundary errors and eliminate the influence of resolution through normalization, enabling cross-scale comparisons.

[0057] 3. Strip Evaluation Zone

[0058] Definition: The region jointly defined by the extracted boundary (the boundary generated by the algorithm) and the reference boundary (the true boundary), i.e., the spatial range between the two. Technical function: To focus on the "fuzzy zone" of boundary error, rather than the global classification result, and to avoid the boundary error being masked by the overall classification accuracy.

[0059] 4. Commission Area

[0060] Definition: The total area of ​​pixels within a strip-shaped evaluation region where the algorithm incorrectly labels non-target features as target features. Technical significance: Reflects the type of error caused by over-segmentation in the algorithm, such as misclassifying non-water bodies (e.g., buildings) as water bodies.

[0061] 5. Omission Area

[0062] Definition: The total area of ​​target land features missed by the algorithm within a strip-shaped evaluation region. Technical function: Reflects the type of error caused by excessive smoothing in the algorithm, such as missing wetlands or small lakes.

[0063] 6. High-Resolution Remote Sensing Image

[0064] Definition: Remote sensing imagery with a spatial resolution better than 1 meter, such as WorldView, QuickBird, and Jilin-1. Technical Function: Serves as a reference boundary data source for both visual interpretation and digital delineation, ensuring high accuracy of the reference boundary.

[0065] 7. Confusion Matrix

[0066] Definition: A two-dimensional table used for classification accuracy assessment, comparing the classification results of pixels with the true values, including true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN). Technical application: Used to calculate the area of ​​false positives and the area of ​​false negatives, for example:

[0067] False positive area = FP (number of false positive pixels)

[0068] Area of ​​missed detection = FN (number of false negative pixels)

[0069] 8. Spatial Resolution

[0070] Definition: The actual ground size (unit: meters) corresponding to a single pixel in a remote sensing image. Technical function: Used in the BDAI formula to normalize error values, ensuring fairness in algorithm evaluation for images of different resolutions.

[0071] 9. Error Type Classification

[0072] Definition: Based on the ratio of incorrectly extracted area to missed extracted area, determine the type of algorithm error:

[0073] The area of ​​incorrectly extracted features is significantly larger than the area of ​​missed extraction features → Oversegmentation bias (e.g., the algorithm draws too many non-target features);

[0074] The area missed is significantly larger than the area incorrectly captured → Over-smoothing bias (e.g., the algorithm misses the target ground features);

[0075] Both are equivalent → Boundary offset type deviation (e.g., the algorithm shifts overall but the shape remains). Technical role: Provides direction for algorithm improvement, such as adjusting the segmentation threshold or smoothing parameters.

[0076] 10. Study Area Size (S)

[0077] Definition: The total geographical area of ​​the region being evaluated (unit: square kilometers). Technical function: Used in the BDAI formula to normalize the error value and avoid evaluation bias caused by differences in region size.

[0078] This application provides a binary classification accuracy evaluation method based on boundary dissimilarity area, which can be widely applied to the binary classification accuracy evaluation of various land features in remote sensing images. For ease of understanding, the following uses water body extraction as an example to describe the specific implementation of the invention in detail. Specifically, it includes the following steps:

[0079] A1. Acquire multispectral remote sensing images of the target area.

[0080] A2. Calculate the water index of the target area based on remote sensing images that have undergone preprocessing such as atmospheric correction;

[0081] The preferred values ​​are MNDWI, NDWI, AWEI0, AWEI1, and WI2015. The calculation formula is as follows:

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] In the formula: Blue, Green, Red, SWIR, SWIR1, SWIR2, and NIR are the remote sensing reflectivities of the blue band, green band, red band, shortwave infrared band, shortwave infrared 1 band, shortwave infrared 2 band, and near-infrared band, respectively, corresponding to B2, B3, B4, B8, B11, B12, and B13 in Sentinel-2.

[0088] A3. Based on the water index results after processing such as smoothing filtering, image enhancement or edge suppression, a threshold segmentation method is introduced to extract water bodies.

[0089] The preferred method is the OTU (Otsu method) adaptive threshold segmentation method, which automatically finds the segmentation threshold that maximizes the inter-class variance based on the statistical characteristics of the image histogram.

[0090] A4. High-resolution remote sensing imagery provided by the Google Earth platform was selected as the base map for interpretation. This platform's imagery has a spatial resolution of 0.3 to 0.5 meters, enabling clear identification of surface water body boundaries. Clear images with an imaging date no more than 30 days different from the imaging time of the remote sensing image used to extract the water body to be evaluated, and a cloud cover of less than 5%, were selected. Professional interpreters conducted manual visual interpretation and digital delineation along the visible land-water boundary line in the imagery, using the continuous outer edge of the water body as the boundary. No subjective speculation was made for shaded or obscured areas to ensure the accuracy of the reference water body boundary.

[0091] A5. Import the extracted water body boundary obtained in step A3 and the reference water body boundary obtained in step A4 into the ArcGIS platform. Within the strip area between the reference water body boundary and the extracted water body boundary, compare the water body extraction result raster with the reference water body boundary raster pixel by pixel. Count the number of pixels whose extraction result is water body but whose reference true value is non-water body as the misextracted area, and count the number of pixels whose extraction result is non-water body but whose reference true value is water body as the missed area. Add the two to get the total number of pixels with inconsistent classification, and substitute them into the formula to calculate the BDAI value.

[0092] A6. The extraction results of five water body index methods (NDWI, MNDWI, AWEI0, AWEI1, and WI2015) were compared pixel-by-pixel with the reference water body boundary. The total number of inconsistent pixels within the boundary band region for each sample was counted, and the BDAI value of each sample was calculated using the formula. The results were then compiled into a table, with each sample corresponding to a set of data, ensuring a one-to-one correspondence between the BDAI value of each index method and the reference boundary. The calculated BDAI values ​​of each index were statistically summarized, and the mean, median, standard deviation, minimum, maximum, and interquartile range of each method were calculated to evaluate the overall performance and stability of each method in terms of boundary classification reliability. The smaller the BDAI mean, the less the classification consequences caused by the geometric error of the extracted water body boundary, and the higher the classification reliability; the smaller the standard deviation, the more stable the performance of the method among different samples.

[0093] In this embodiment, to systematically test the universality and robustness of the boundary dissimilarity area index, this study selected 50 representative reservoirs in China as the study area. The selection strictly adhered to the principle of diversity across five dimensions: climate, geomorphology, environmental background, season, and reservoir type. The study area covers major climate zones including humid, semi-humid, semi-arid, and arid, and is distributed across various geomorphic units such as plains, hills, basins, plateaus, and mountains. The shoreline environment encompasses typical land surface types such as urban areas, cultivated land, forest land, bare land, and snow and ice. Image acquisition spans spring, summer, autumn, and winter to reflect phenological changes. The reservoir samples include both large and medium-sized reservoirs, and encompass different morphological types such as riverine and lacustrine reservoirs. Through this multi-dimensional system design, the validation set can simulate complex scenarios in real-world applications to the greatest extent possible, providing a solid geographical empirical foundation for method evaluation.

[0094] To further analyze the extraction effect of the boundary areas, three 100m×100m (1 hectare) square quadrats were manually identified and determined on key sections of the shoreline boundary of each reservoir (such as inlets, convex banks, concave banks, and shaded areas). The selection criteria for the quadrats were: (1) containing a clear water-land boundary; (2) covering a major interference background (such as building shadows, mountain shadows, and mixed vegetation areas); and (3) being stably identifiable on 10-meter resolution images. A total of 150 boundary quadrats constituted the basic spatial unit for the accuracy evaluation of this study.

[0095] Figure 2 This is a schematic diagram of BDAI, using NDWI to extract the water body boundary and the reference water body boundary interpreted from high-resolution remote sensing imagery on Google Earth.

[0096] Figure 3Box plots of the BDAI results for each index are shown. Based on the mean, the BDAI ranking of the five methods is: NDWI (1046.84) < AWEI1 (1084.15) < WI2015 (1128.72) < MNDWI (1320.57) < AWEI0 (1641.28). NDWI performed best, with the fewest average discrete pixels per square kilometer at 1046.84; AWEI1 and WI2015 were next, with mean values ​​of 1084.15 and 1128.72 respectively; MNDWI performed poorly, with a mean of 1320.57; and AWEI0 performed worst, with a mean of 1641.28, 1.57 times that of NDWI.

[0097] Based on the median, the ranking is as follows: NDWI (604.83) < AWEI1 (646.29) < WI2015 (686.78) < MNDWI (957.10) < AWEI0 (1274.87). The medians for all methods are significantly smaller than the means, indicating that the BDAI distribution is right-skewed, meaning that a few large error samples inflate the overall mean. The NDWI shows the largest difference between its mean and median (1046.84 vs 604.83), indicating that it is most significantly affected by extremely high values.

[0098] Standard deviation reflects the stability of the method. AWEI0 has the smallest standard deviation (1556.96), but its mean is the highest, indicating that although the method has relatively small fluctuations, its overall bias is large. The standard deviation of MNDWI is 1636.70. The standard deviations of AWEI1, NDWI, and WI2015 are relatively large, at 1752.62, 1711.91, and 1731.23, respectively. Combined with their low means, this indicates that these three methods perform well on most samples, but may exhibit large areal biases in some complex scenarios.

[0099] The interquartile range (IQR) reflects the error distribution range of the middle 50% of the samples. AWEI1 has the smallest IQR (570.06), followed by WI2015 (605.84), NDWI (655.88), and MNDWI (653.55), indicating that the middle 50% of the samples are relatively concentrated in these four methods. AWEI0 has the largest IQR (1232.77), which is 2.16 times that of AWEI1, further confirming its instability.

[0100] In summary, the accuracy of the differentiation area among the five methods is ranked as follows: NDWI ≈ AWEI1 ≈ WI2015 > MNDWI > AWEI0. NDWI, AWEI1, and WI2015 perform excellently in terms of mean and median; MNDWI performs moderately; and AWEI0 has the largest overall bias and performs the worst.

[0101] Figures 4-5 A comparison chart showing the effect of extracting the boundary from the typical index results and the boundary of the reference water body.

[0102] This application provides a binary classification accuracy evaluation system based on the area of ​​distinct boundaries, including a data acquisition module 21, a region division module 22, a statistical analysis module 23, and an error determination module 24.

[0103] The data acquisition module 21 is configured to: acquire the binary classification result of the target region and obtain the extracted classification boundary vector data, and acquire the reference classification boundary vector data of the same target region;

[0104] The region division module 22 is configured to determine the boundary strip evaluation region based on the reference boundary and the extracted boundary;

[0105] The statistical analysis module 23 is configured to: statistically analyze the area of ​​inconsistent pixels within the boundary strip evaluation region and calculate the boundary dissimilarity area index;

[0106] The error determination module 24 is configured to determine the extraction error type based on the ratio of the misextracted area to the missed area.

[0107] This embodiment provides a computer-readable storage medium storing computer-executable instructions. When executed by a processor, these instructions are used to implement a binary classification accuracy evaluation method based on the area of ​​dissimilar boundaries.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications should be covered within the scope of the claims of the present invention.

Claims

1. A method for evaluating the accuracy of binary classification based on the area of ​​distinct boundaries, characterized in that, Includes the following steps: S1. Obtain the binary classification result of the target region and extract the classification boundary vector data; S2. Obtain reference classification boundary vector data for the same target region; S3. Determine the boundary strip evaluation area based on the reference boundary and the extracted boundary; S4. Statistically measure the area of ​​inconsistent pixels within the boundary band evaluation area to identify mis-extracted and under-extracted areas. S5. Calculate the boundary dissimilarity area index (BDAI). S6. Determine the type of extraction error based on the ratio of the incorrectly extracted area to the missed extracted area.

2. The binary classification accuracy evaluation method based on distinct boundary areas as described in claim 1, characterized in that: The binary classification results divide the land surface into two categories: target features and non-target features.

3. The binary classification accuracy evaluation method based on boundary dissimilar area as described in claim 2, characterized in that: The reference classification boundary vector data is obtained through manual visual interpretation of high-resolution remote sensing images, specifically including images whose imaging date and image extraction time differ by no more than 30 days and whose cloud cover is less than 5%.

4. The binary classification accuracy evaluation method based on distinct boundary areas as described in claim 3, characterized in that: The boundary strip evaluation area is jointly defined by the extraction boundary and the reference boundary. Specifically, it includes the area where the extraction boundary extends outward to the reference boundary, the area where the reference boundary extends outward to the extraction boundary, and the area formed by the intersection and overlap of the two. This area does not require a preset buffer distance but is naturally formed by the actual spatial relationship.

5. The binary classification accuracy evaluation method based on distinct boundary areas as described in claim 4, characterized in that: The area statistics of the inconsistent pixels are achieved by comparing the extracted results with the reference ground truth raster data pixel by pixel, using a mixed... The confusion matrix classification and statistical method is used to calculate the area of ​​incorrect extraction and the area of ​​missed extraction respectively.

6. The binary classification accuracy evaluation method based on boundary dissimilar area as described in claim 5, characterized in that: The formula for calculating the boundary dissimilarity area index (BDAI) is as follows: Where Acommission is the area of ​​incorrect extraction, Aomission is the area of ​​missed extraction, R is the spatial resolution of the image, and S is the area of ​​the study region.

7. The binary classification accuracy evaluation method based on distinct boundary areas as described in claim 6, characterized in that, The error type is determined by comparing the ratio of incorrectly extracted area to the missed extracted area: If the area of ​​incorrect extraction is significantly larger than the area of ​​missing extraction, it is judged as an oversegmentation bias. If the area of ​​missed extraction is significantly larger than the area of ​​incorrect extraction, it is judged as an oversmoothing bias. If the two are comparable, it is determined to be a boundary offset type deviation.

8. The binary classification accuracy evaluation method based on boundary dissimilar area as described in claim 7, characterized in that: The method is applicable to various land cover classification tasks, including but not limited to water body extraction, vegetation identification, and land use mapping. Only the reference boundary and extraction boundary need to be input to achieve a quantitative assessment of the classification results.

9. A binary classification accuracy evaluation system based on the area of ​​distinct boundaries, characterized in that, It includes a data acquisition module, a region division module, a statistical analysis module, and an error determination module. The data acquisition module is configured to: acquire the binary classification result of the target region and obtain the extracted classification boundary vector data, and acquire the reference classification boundary vector data of the same target region; The region division module is configured to determine the boundary strip evaluation region based on the reference boundary and the extracted boundary. The statistical analysis module is configured to: statistically analyze the area of ​​inconsistent pixels within the boundary strip evaluation region and calculate the boundary dissimilarity area index; The error determination module is configured to determine the extraction error type based on the ratio of the misextracted area to the missed area.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that, when executed by a processor, implement the method as described in any one of claims 1 to 8.

Images