A method and system for calculating fine-scale land use types
Through the classification recognition model constructed by a layered perceptron and multi-grained decision-maker, combined with the pixel and spatial relationship attributes, the problem of insufficient precision in land use type calculation in the existing technology is solved, and efficient land use type of fine plots is achieved Identification and calculation improves productivity and reduces costs.
Patent Information
- Application Number
- CN202410668996.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-28
- Publication Date
- 2025-01-17
- Estimated Expiration
- 2044-05-28
AI Technical Summary
The prior art is difficult to achieve accurate judgment and calculation of fine plots in the calculation of land use types, resulting in insufficient information accuracy and inability to meet the needs of refined applications, especially in the absence of a large amount of training data.
The classification recognition model constructed by a layered perceptron and multi-grained decision maker is used to obtain original high-score images, extract fine plot-level maps and land use types, and adjust them according to the confidence level of land use types, combining cell attributes and Spatial relational properties are used for sample supplementation and training model optimization.
It significantly improves the accuracy of land use types of plots, reduces production costs, improves production efficiency, and achieves more accurate land use type identification with few samples.
Smart Images
Figure CN118608949B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of remote sensing computing technology, and in particular to a method and system for calculating fine plot land use types. Background Art
[0002] With the continuous development of artificial intelligence technology, related technologies are solving more and more complex problems. However, in many real-world scenarios, there is often a lack of a large amount of training data, which makes traditional machine learning technology unable to cope with these tasks well. In response to the above problems, few-shot learning or few-shot learning (FSL) is a method specifically designed for such problems.
[0003] Early land use and cover change (LUCC) information production mainly relied on the traditional human-computer interactive interpretation operation mode. On the one hand, its manpower and time costs were huge, and the cycle of information change and update was slow; on the other hand, its map morphology still represented a comprehensive geographical area, which was not refined to the level of real surface entities, resulting in fuzzy and confused benchmarks and inconsistent standards when multi-source data was merged with the map. Therefore, the accuracy of its information still cannot meet the actual needs of carrying out refined applications in many fields.
[0004] From the 1980s to the beginning of this century, machines were driven to simulate the manual visual interpretation process in order to achieve the fine extraction of LUCC information. Because the spatial resolution of remote sensing images at that time was low, the LUCC information obtained by the corresponding pixel-level classification method could not determine the exact boundary and could not express the morphological and structural characteristics of the surface LU. At the same time, the corresponding LU / LC classification system formulated also allowed LU and LC to coexist in one system in a fuzzy or confusing manner. It could only identify the major categories of land objects and carry out simple mapping at the medium and macro scales.
[0005] After 2000, with the continuous improvement of image spatial resolution, the image segmentation technology introduced by computer vision developed an object-oriented classification method, which achieved better mapping results and more complete analysis methods than pixel-level methods. The surface LU patches obtained by object-oriented segmentation can be integrated with GIS and other multi-source heterogeneous data at the feature level to carry out collaborative analysis of the surface composition and change mechanism of LUCC. However, the LU patches obtained by unsupervised segmentation are not accurate in form and category, and often cannot accurately correspond to real geographical entities. Therefore, the boundaries of LU and LC are still difficult to define and are mixed with each other. It is still difficult to implement LCC index calculation and pattern analysis based on internally homogeneous control units. Summary of the invention
[0006] The object of the present invention is to provide a method and system for calculating the land use type of a fine plot of land, which can enhance the discrimination and calculation of the land use type of a plot of land.
[0007] To achieve the above object, the present invention provides the following solutions:
[0008] A method for calculating fine-scale land use types, including:
[0009] Obtain the original high-resolution image of the target area;
[0010] Extracting the original high-resolution image to obtain fine plot-level patches and land use types;
[0011] The fine plot-level patches are input into a classification and recognition model to determine the land use type corresponding to each of the fine plot-level patches, and the land use type is adjusted according to the land use type confidence level; the classification and recognition model is constructed using a hierarchical perceptron and a multi-granularity decision maker.
[0012] Optionally, the training process of the classification recognition model is:
[0013] Obtain training data and corresponding labels, and build a training model;
[0014] The training data is input into the training model, and training is performed with the goal of minimizing the loss between the result output by the training model and the label of the training data, and the trained model is determined as the classification recognition model.
[0015] Optionally, before obtaining the training data and the corresponding labels, the following steps are also included:
[0016] The initial data is input into the training model, the attribute characteristics of the plot are calculated, and the reference points of the supplementary samples are determined according to the attribute characteristics, the samples are supplemented, and the supplemented sample data are determined as training data; the attribute characteristics include pixel attributes and spatial relationship attributes.
[0017] Optionally, the calculation process of the attribute characteristics of the land plot is:
[0018] The pixel attributes are calculated through the raster image corresponding to the plot patch. When calculating the spatial relationship attributes of the plot patch, the window is used as the basic calculation unit, each plot patch is divided into the corresponding window, and the land cover attributes in the buffer range of each window are calculated, and finally integrated into a single patch.
[0019] Optionally, the method for acquiring the training data and the corresponding labels is: defining the category field of each image spot in the training data and the corresponding calculation label.
[0020] Optionally, the adjusting the land use type according to the land use type confidence level specifically includes:
[0021] Calculate the land use type confidence and compare the land use type confidence with the preset threshold. When the land use type confidence is greater than or equal to the preset threshold, replace the land use type output at the previous moment with the currently output land use type; when the land use type confidence is less than the preset threshold, modify it through manual verification.
[0022] The present invention also provides a system for calculating fine-scale land use types, comprising:
[0023] A data acquisition unit, used to obtain original high-resolution images of the target area;
[0024] A data extraction unit, used to extract the original high-resolution image to obtain fine plot-level patches and land use types;
[0025] The land use type classification calculation unit is used to input the fine plot-level image into the classification recognition model, determine the land use type corresponding to each of the fine plot-level images, and adjust the land use type according to the land use type confidence; the classification recognition model is constructed using a hierarchical perceptron and a multi-granularity decision maker.
[0026] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0027] The present invention discloses a method and system for calculating fine land use types of plots, the method comprising obtaining an original high-resolution image of a target area; extracting the original high-resolution image to obtain fine plot-level patches and land use types; inputting the fine plot-level patches into a classification recognition model, determining the land use types corresponding to each of the fine plot-level patches, and adjusting the land use types according to the land use type confidence; the classification recognition model is constructed using a hierarchical perceptron and a multi-granularity decision maker. The present invention can enhance the discrimination and calculation of land use types of plots. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0029] Figure 1 For the present invention
[0030] Figure 2 In this embodiment DETAILED DESCRIPTION
[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0032] The object of the present invention is to provide a method and system for calculating the land use type of a fine plot of land, which can enhance the discrimination and calculation of the land use type of a plot of land.
[0033] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0034] like Figure 1 As shown, the present invention provides a method for calculating the land use type of a fine plot, comprising:
[0035] Step 100: Acquire the original high-resolution image of the target area;
[0036] Step 200: extracting the original high-resolution image to obtain fine plot-level patches and land use types;
[0037] Step 300: Input the fine plot-level patches into a classification recognition model, determine the land use type corresponding to each of the fine plot-level patches, and adjust the land use type according to the land use type confidence; the classification recognition model is constructed using a hierarchical perceptron and a multi-granularity decision maker.
[0038] This embodiment aims at the production process of a fine plot land use type calculation product with a few samples enhanced for a typical production task. First, the fine plot-level patches and their land use types extracted from the high-resolution image are input into the model, and the patch pixel attributes and spatial relationship attributes are calculated to determine the reference points for supplementary samples. A small number of samples are supplemented according to the features selected by the model, and the model parameters are determined. Then, the model is trained, and the confidence of the plot land use type is calculated for the training results. If the confidence of the training result meets the predefined accuracy requirement, then the training result is reliable and the land use type of the plot is modified. If the accuracy is not met, the patch pixel attributes and spatial relationship attribute features are recalculated, and the above steps are repeated until the confidence of the training result meets the accuracy requirement.
[0039] As a specific implementation step, Figure 2 As shown, it mainly includes:
[0040] (1) Data initialization: Define the category field and corresponding calculation label of each patch in the shp file. Divide the patches in the shp file into two categories, set the TRAIN_FLAG field of the patches with clear categories to true and define their category fields, and set the TRAIN_FLAG field of the remaining patches to false.
[0041] (2) Calculate the patch attribute characteristics: including pixel attributes and spatial relationship attributes, and update the results back to the shp file. Calculate the pixel attributes through the raster image corresponding to the patch. When calculating the patch spatial relationship characteristics, use the window as the basic calculation unit, divide each patch into the corresponding window, calculate the land cover attributes in the buffer range of each window, and finally integrate them into a single patch.
[0042] (3) Export shp file: Export the attribute table in the shp file into CSV format for storage. According to the TRAIN_FALG field, export the data into toTrain.csv and toPredict.csv. The TRAIN_FLAG field of toTrain.csv is true, which is used to determine the model parameters. The TRAIN_FLAG field of toPredict.csv is false, which is a collection of land use types to be calculated.
[0043] (4) Training model: Determine model parameters through the exported CSV file;
[0044] (5) Calculate the confidence of the land use type of the map: Determine the confidence of the land use type of the map to be calculated through the trained model, and save the calculation results to a CSV file;
[0045] (6) Update shp file: import the calculated CSV file into the updated shp file;
[0046] (7) Determine accuracy: If the confidence level of the calculation result meets the requirement, the land use type of the map block is modified and the calculation is completed;
[0047] (8) Determine the land use types of some low-confidence spots: manually check some low-confidence spots and modify them to land use types that can be determined;
[0048] (9) repeat the steps;
[0049] The step (4) training model: defining multiple classifiers, dividing toTrain.csv into a training set, a validation set, and a test set, and determining the parameters of the multiple classifiers through the above sets.
[0050] The step (5) calculates the confidence of the land use type of the fine patch: the confidence is defined by integrating the classification results of multiple classifiers.
[0051] The step (6) updates the shp file: importing the calculated CSV file into the updated shp file;
[0052] Step (7) determines the accuracy: if the confidence of the calculation result meets the requirement, the land use type of the image patch is modified and the calculation is completed;
[0053] The step (8) determines the land use types of some low-confidence patches: manually verify, modify the category field attributes, and modify the TRAIN_FLAG field to true.
[0054] The technical method of this embodiment uses the machine learning mechanism of "reinforcement learning to drive production process optimization". First, it integrates the layered perceptron to improve the internal feedback mechanism and active sample supplement mechanism of various networks, and then cooperates with the multi-granularity decision maker to dynamically supplement the external incremental knowledge to the fine plot scene, so as to continuously enhance the discrimination and calculation of the land use type of the plot.
[0055] The overall production efficiency using the technology of this embodiment is about 20 times higher than that of products obtained by pure manual drawing, and the production cost is greatly reduced to about 1 / 5. With continuous reinforcement and memory, the production efficiency will continue to improve and the cost will continue to decrease.
[0056] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.
[0057] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A method for calculating fine-scale land use types, characterized in that: include: Obtain the original high-resolution image of the target area; Extracting the original high-resolution image to obtain fine plot-level patches and land use types; Inputting the fine plot-level image patches into a classification recognition model, determining the land use type corresponding to each of the fine plot-level image patches, and adjusting the land use type according to the land use type confidence level; The classification and recognition model is constructed using a layered perceptron and a multi-granularity decision maker; The fine plot-level patches are input into the classification recognition model to determine the land use type corresponding to each fine plot-level patch, and the land use type is adjusted according to the land use type confidence. The specific processing process is as follows: (1) Data initialization: define the category field and corresponding calculation label of each patch in the shp file; divide the patches in the shp file into two categories, set the TRAIN_FLAG field of the patches with clear categories to true and define their category fields, and set the TRAIN_FLAG field of the remaining patches to false; (2) Calculate the patch attribute characteristics: including pixel attributes and spatial relationship attributes, and update the results back to the shp file; calculate the pixel attributes through the raster image corresponding to the patch. When calculating the patch spatial relationship characteristics, use the window as the basic calculation unit, divide each patch into the corresponding window, calculate the land cover attributes in the buffer range of each window, and finally integrate them into a single patch; (3) Export shp file: Export the attribute table in the shp file into CSV format for storage; export the data into toTrain.csv and toPredict.csv according to the TRAIN_FALG field. The TRAIN_FLAG field of toTrain.csv is true, which is used to determine the model parameters. The TRAIN_FLAG field of toPredict.csv is false, which is a collection of land use types to be calculated. (4) Training model: Determine model parameters through the exported CSV file; (5) Calculate the confidence of the land use type of the map: Determine the confidence of the land use type of the map to be calculated through the trained model, and save the calculation results to a CSV file; (6) Update shp file: import the calculated CSV file into the updated shp file; (7) Determine accuracy: If the confidence level of the calculation result meets the requirement, the land use type of the map block is modified and the calculation is completed; (8) Determine the land use types of some low-confidence spots: Manually check some low-confidence spots and modify them to land use types that can be determined; (9) Repeat steps (1) to (8).
2. The method for calculating the fine land use type according to claim 1, characterized in that: The training process of the classification recognition model is: Obtain training data and corresponding labels, and build a training model; The training data is input into the training model, and training is performed with the goal of minimizing the loss between the result output by the training model and the label of the training data, and the trained model is determined as the classification recognition model.
3. The method for calculating the fine land use type of claim 2, characterized in that: Before obtaining the training data and corresponding labels, it also includes: The initial data is input into the training model, the attribute characteristics of the plot are calculated, and the reference points of the supplementary samples are determined according to the attribute characteristics, the samples are supplemented, and the supplemented sample data are determined as training data; the attribute characteristics include pixel attributes and spatial relationship attributes.
4. The method for calculating the fine land use type according to claim 2, characterized in that: The method for acquiring the training data and the corresponding labels is: defining the category field of each image spot in the training data and the corresponding calculation label.
5. A system for calculating fine-scale land use type, applied to the method according to any one of claims 1 to 4, characterized in that: include: A data acquisition unit, used to obtain original high-resolution images of the target area; A data extraction unit, used to extract the original high-resolution image to obtain fine plot-level patches and land use types; A land use type classification calculation unit, used for inputting the fine plot-level image patches into a classification recognition model, determining the land use type corresponding to each of the fine plot-level image patches, and adjusting the land use type according to the land use type confidence level; The classification and recognition model is constructed by using a layered perceptron and a multi-granularity decision maker.