Alphaearth foundations and machine learning based future year land use and land cover raster data downscaling method and system

Abstract

The application provides a land use and cover grid data downscaling method and system based on AlphaEarth Foundations and machine learning in future years, which identifies coarse scale pixels in which land classes change in future years by comparing coarse scale data in future years and a base year; maps the coarse scale pixels to fine scale data in the base year to determine a target area; counts the number of fine scale pixels of each land class in the target area to obtain the net increase of fine scale pixels of each land class in future years; in the target area, generates a conversion probability corresponding to each land class for each fine scale pixel by using a land class conversion probability prediction model; constructs an optimization model with the sum of the conversion probabilities corresponding to the land classes to which all fine scale pixels in the target area are assigned as an objective function, and solves the optimization model to obtain an optimal land class assignment result of the fine scale pixels in the target area; and fuses the optimal land class assignment result with the fine scale data in the base year to generate fine scale land use and cover grid data in future years after downscaling.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for downscaling future year land use and cover raster data based on AlphaEarth Foundations and machine learning. Background Technology

[0002] Raster data downscaling is a method for obtaining high spatial resolution raster data. It involves transforming low spatial resolution (hereinafter referred to as "coarse scale") raster data into high spatial resolution (hereinafter referred to as "fine scale") raster data through specific methods.

[0003] In the field of land use and cover change research, the demand for fine-scale raster data on the spatial distribution of land use is growing. Fine-scale raster data is not only more spatially refined, facilitating more accurate regional analysis, but also provides theoretical support for future land use management decisions. Typically, publicly available land use and cover raster data products for future years are not only few in number, but most are also coarse-scale raster data (spatial resolution is mostly 1 km or coarser).

[0004] Current downscaling methods for land use and cover raster data for future years are all based on processing existing coarse-scale raster data for future years. Corresponding technical means include machine learning methods, spatial suitability-based allocation models, and spatial rule-based allocation models. Specific research examples include the use of the ECOBRIDGE algorithm and spatiotemporal monitoring of land use and cover information based on prior knowledge.

[0005] However, among the two methods mentioned above, the ECOBRIDGE algorithm has high requirements for the quality of the input data. For example, obtaining vector data for surface features is difficult. If no vector data is used in the downscaling process, the output fine-scale land use and cover raster will appear visually jagged. At the same time, there is no uniform precedent in the academic community for setting land use conversion rules. Land use and cover changes for specific study areas are largely based on the personal experience of experts, which makes it difficult to guarantee the quality of input data and indirectly affects the quality of output results. The spatiotemporal monitoring method of land use and cover information based on prior knowledge lacks physical meaning in the spatial enhancement (i.e., setting the spatial distribution of land use categories in fine-scale pixels) stage, relying only on the two-dimensional texture features of raster data. In addition, because land use category conversion often has nonlinear and abrupt characteristics, the results of simple linear fitting are prone to deviation from reality. Summary of the Invention

[0006] This invention provides a method and system for downscaling future year land use and cover raster data based on AlphaEarth Foundations and machine learning, which addresses the shortcomings of existing technologies that heavily rely on manually defined rules and difficult-to-obtain auxiliary data, and achieves data-driven and automated optimization for downscaling future year land use and cover raster data.

[0007] This invention provides a method for downscaling raster data of land use and cover for future years, comprising the following steps:

[0008] The data includes fine-scale land use and cover raster data for the baseline year, pixel-level embedding vector data, and coarse-scale land use and cover raster data for future years. The pixel-level embedding vector data is sourced from the Google Satellite Embedding dataset released by AlphaEarthFoundations.

[0009] The fine-scale land use and cover raster data of the base year are aggregated to generate coarse-scale land use and cover raster data of the base year.

[0010] By comparing the coarse-scale land use and cover raster data of the future year with the coarse-scale land use and cover raster data of the baseline year, coarse-scale pixels with changes in land use type in the future year are identified; the coarse-scale pixels with changes in land use type are mapped onto the fine-scale land use and cover raster data of the baseline year, the range of fine-scale pixels covered by the coarse-scale pixels is determined, and the range of fine-scale pixels is defined as the target area.

[0011] By statistically analyzing the changes in the number of fine-scale pixels for each land type within the target area, the net increase in fine-scale pixels for each land type in future years can be obtained.

[0012] Within the target area, a land use conversion probability prediction model is used to generate conversion probabilities corresponding to each land use type for each fine-scale cell;

[0013] An optimization model is constructed with the objective function of maximizing the sum of the conversion probabilities corresponding to the land use types assigned to all fine-scale pixels within the target area;

[0014] Solving the optimization model yields the optimal land use classification results for fine-scale pixels within the target area;

[0015] The optimal land use classification result is fused with the fine-scale land use and cover raster data of the baseline year to generate downscaled fine-scale land use and cover raster data for future years.

[0016] According to the method for downscaling future land use and cover raster data provided by the present invention, the construction of the optimization model further includes setting the following constraints for the optimization model:

[0017] The total change in the number of fine-scale pixels for each land type is consistent with the net increase in the fine-scale pixels.

[0018] Each fine-scale cell can only be assigned one land class.

[0019] According to the method for downscaling land use and cover raster data for future years provided by the present invention, the optimization model also sets the following constraints:

[0020] If a fine-scale pixel is found to have the same land use type in the fine-scale data of the base year and the pseudo-downscale fine-scale data, then the land use type allocation result of that pixel remains unchanged during the optimization process.

[0021] The pseudo-downscaling fine-scale data refers to coarse-scale data of future years that has been resampled to the same resolution.

[0022] According to the land use and cover raster data downscaling method for future years provided by the present invention, the step of statistically analyzing the changes in the number of fine-scale pixels for each land type within the target area to obtain the net increase in the number of fine-scale pixels for each land type in future years specifically includes:

[0023] The number of fine-scale pixels of each type in the target area in the fine-scale data of the base year and the number of fine-scale pixels of each type in the pseudo downscaled fine-scale data corresponding to the coarse-scale data of the future year are counted respectively. The difference between the number of fine-scale pixels of each type is calculated to obtain the net increase of fine-scale pixels of each type in the future year.

[0024] The pseudo-downscaled fine-scale data refers to coarse-scale data of future years that has been resampled to the same resolution.

[0025] According to the method for downscaling land use and cover raster data for future years provided by the present invention, the pixel-level embedded vector data is a high-dimensional vector obtained by feature extraction of satellite image blocks through a deep neural network, and the deep neural network is pre-trained on satellite images using a self-supervised learning method.

[0026] According to the land use and cover raster data downscaling method for future years provided by the present invention, the pixel-level embedded vector data is derived from the Google Satellite Embedding dataset released by AlphaEarth Foundations.

[0027] According to the method for downscaling land use and cover raster data for future years provided by the present invention, the land use conversion probability prediction model is obtained by machine learning training with sample cell-level embedded vector data as input and actual land use information of fine-scale land use and cover raster data of sample base year as training target; wherein, the land use conversion probability prediction model is a random forest model.

[0028] The method for downscaling land use and cover raster data for future years provided by the present invention is characterized in that the training process of the land use conversion probability prediction model further includes:

[0029] The hyperparameters of the land use conversion probability prediction model were tuned using the Bayesian optimization method.

[0030] According to the method for downscaling land use and cover raster data for future years provided by the present invention, the method for aggregating the fine-scale land use and cover raster data of the baseline year is the mode statistics method.

[0031] This invention also provides a system for downscaling raster data of land use and cover for future years, comprising the following modules:

[0032] The data acquisition module is used to acquire fine-scale land use and cover raster data, pixel-level embedding vector data and coarse-scale land use and cover raster data for the baseline year, wherein the pixel-level embedding vector data comes from the Google Satellite Embedding dataset released by AlphaEarth Foundations.

[0033] The data acquisition module is also used to aggregate the fine-scale land use and cover raster data of the reference year to generate coarse-scale land use and cover raster data of the reference year.

[0034] The change area determination module is used to compare the coarse-scale land use and cover raster data of the future year with the coarse-scale land use and cover raster data of the baseline year, identify coarse-scale pixels where land use types have changed in the future year; map the coarse-scale pixels where land use types have changed onto the fine-scale land use and cover raster data of the baseline year, determine the range of fine-scale pixels covered by the coarse-scale pixels, and define the range of fine-scale pixels as the target area.

[0035] The net increment calculation module is used to statistically analyze the changes in the number of fine-scale pixels of each land type within the target area, and to obtain the net increment of fine-scale pixels of each land type in future years.

[0036] The probability prediction module is used to generate conversion probabilities corresponding to each land use type for each fine-scale cell within the target area using a land use conversion probability prediction model.

[0037] The spatial optimization module is used to construct an optimization model with the objective function of maximizing the sum of the conversion probabilities corresponding to the land types assigned to all fine-scale pixels within the target area.

[0038] The spatial optimization module is also used to solve the optimization model to obtain the optimal land use classification result for fine-scale pixels within the target area;

[0039] The data fusion module is used to fuse the optimal land use allocation result with the fine-scale land use and cover raster data of the baseline year to generate downscaled fine-scale land use and cover raster data for future years.

[0040] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the land use and cover raster data downscaling method for future years as described above.

[0041] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for downscaling land use and cover raster data for future years as described above.

[0042] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the method for downscaling land use and cover raster data for future years as described above.

[0043] This invention provides a method and system for downscaling future land use and cover raster data based on AlphaEarth Foundations and machine learning. Utilizing embedded vectors as an objective data feature and a machine learning model, it automatically learns spatial distribution patterns from historical data. The results are stable and reproducible, avoiding biases introduced by human intervention and ensuring comparability of simulation results across different scenarios and regions. An optimization model is introduced to find the overall optimal spatial distribution scheme while meeting the constraint of total macro-level land demand, making the results more reasonable both locally at the pixel level and globally at the regional level. This transforms land use downscaling from relying on expert experience into a standardized, automated, and reproducible process. Furthermore, this invention eliminates the need to collect high-precision plot vector data and rely on experts with deep domain knowledge to formulate complex rules, lowering the technical threshold in this field and significantly enhancing the scientific and practical value of the simulation results. Attached Figure Description

[0044] 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.

[0045] Figure 1 This is one of the flowcharts illustrating the method for downscaling land use and cover raster data for future years provided by this invention.

[0046] Figure 2 This is the second flowchart illustrating the method for downscaling land use and cover raster data for future years provided by this invention.

[0047] Figure 3 This is a schematic diagram of the structure of the future year land use and cover raster data downscaling system provided by the present invention.

[0048] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0050] The present invention will now be described in detail with reference to the accompanying drawings. The specific operation methods in the method embodiments can also be applied to the device embodiments or system embodiments. In the description of the present invention, unless otherwise stated, "at least one" includes one or more. "Multiple" refers to two or more. For example, at least one of A, B, and C includes: A existing alone, B existing alone, A and B existing simultaneously, A and C existing simultaneously, B and C existing simultaneously, and A, B, and C existing simultaneously. In the present invention, " / " means "or". For example, A / B can mean A or B. "And / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone.

[0051] The present invention will now be described in detail with reference to specific embodiments.

[0052] In some specific embodiments of the present invention, such as Figure 1As shown, this scheme provides a method for downscaling land use and cover raster data for future years, including:

[0053] Step 100: Obtain fine-scale land use and cover raster data, pixel-level embedding vector data and coarse-scale land use and cover raster data for the baseline year, wherein the pixel-level embedding vector data is from the Google Satellite Embedding dataset released by AlphaEarthFoundations.

[0054] Step 200: Aggregate the fine-scale land use and cover raster data of the reference year to generate coarse-scale land use and cover raster data of the reference year;

[0055] Step 300: Compare the coarse-scale land use and cover raster data of the future year with the coarse-scale land use and cover raster data of the baseline year to identify coarse-scale pixels where land use types have changed in the future year; map the coarse-scale pixels where land use types have changed onto the fine-scale land use and cover raster data of the baseline year to determine the range of fine-scale pixels covered by the coarse-scale pixels, and define the range of fine-scale pixels as the target area;

[0056] Step 400: Statistically analyze the changes in the number of fine-scale pixels for each land type within the target area to obtain the net increase in the number of fine-scale pixels for each land type in future years;

[0057] Step 500: Within the target area, use the land use conversion probability prediction model to generate conversion probabilities corresponding to each land use type for each fine-scale pixel;

[0058] Step 600: Construct an optimization model with the objective function of maximizing the sum of the conversion probabilities corresponding to the land types assigned to all fine-scale pixels within the target area;

[0059] Step 700: Solve the optimization model to obtain the optimal land use classification result for fine-scale pixels within the target area;

[0060] Step 800: The optimal land use allocation result is fused with the fine-scale land use and cover raster data of the baseline year to generate downscaled fine-scale land use and cover raster data for future years.

[0061] It should be noted that existing downscaling schemes for future years rely on costly auxiliary data (such as parcel vectors) and subjective expert rules. Since these rely on experts to formulate land use conversion rules, different experts may arrive at different results, making them difficult to reproduce and resulting in low-quality output. Furthermore, the common practice of directly determining the land use type of each pixel using simple rules or the maximum probability method easily leads to spatial jaggedness and unreasonable distribution. Therefore, the spatial distribution of the downscaling results lacks sufficient physical meaning, and the output is prone to deviating from reality.

[0062] Therefore, this invention acquires fine-scale data of the baseline year, pixel-level embedding vector data, and coarse-scale data of future years. Based on the comparison results of coarse-scale land use and cover raster data of future years and the baseline year, a target area is determined. An optimization model is constructed and solved with the objective function of maximizing the sum of the conversion probabilities corresponding to the land types assigned to all fine-scale pixels within the target area. The optimal land type allocation result is obtained and then fused with the fine-scale data of the baseline year to obtain downscaled fine-scale data of future years. Utilizing the objective data feature of embedding vectors and a machine learning model, spatial distribution patterns are automatically learned from historical data, resulting in stable and reproducible results. An optimization model is introduced to find the overall optimal spatial distribution scheme under the constraint of the total macro-level land demand (net increase). This makes the results more reasonable both locally (high probability) and globally (total balance). It reduces the difficulty and quality requirements of acquiring input data while enhancing the spatial distribution settings of land types in fine-scale pixels, obtaining more physically meaningful downscaled low-spatial-resolution (i.e., coarse-scale) land use and cover raster data of future years.

[0063] The above steps will be explained in detail below with reference to specific embodiments.

[0064] In some specific embodiments of the present invention, such as Figure 2 As shown, the machine learning component of the method for downscaling coarse-scale land use and cover raster data for future years proposed in this invention includes the following steps:

[0065] Step 1: Process the spatial reference system (including projection, coordinates, etc.) and spatial resolution of the acquired data to ensure complete consistency with the geographical features of the land use and cover raster data of the high spatial resolution study area in the baseline year, and form a machine learning training database. The database includes land use zones, independent variable zones, dependent variable zones, and machine learning model zones.

[0066] In some possible embodiments of the present invention, the pixel-level embedded vector data is a high-dimensional vector obtained by feature extraction of satellite image blocks through a deep neural network, which is pre-trained on satellite images using a self-supervised learning method.

[0067] Specifically, this embodiment provides an implementation method for embedding vector data at the pixel level.

[0068] In some possible embodiments of the present invention, the pixel-level embedding vector data is derived from the Google Satellite Embedding dataset released by AlphaEarthFoundations.

[0069] Specifically, in this embodiment, the following data structure is created:

[0070] Land Class Zone: Used to store the pixel values ​​corresponding to all possible land classes in the downscaling experiment.

[0071] The independent variable area stores the 64-dimensional embedding vectors at a global resolution of 10 meters provided by the Google Satellite Embedding dataset published by AlphaEarth Foundations. Geographic features have been removed, and the 64-dimensional vector of each pixel is used as the input independent variable. Finally, it is stored as a large number of one-dimensional arrays in the independent variable area.

[0072] The dependent variable region is composed of fine-scale land use and cover data from the baseline year. For each pixel, a conversion probability is set according to its land use attribute. The dimension of the conversion probability vector is determined based on the number of land use categories stored in the land use region. The value of the dimension corresponding to the pixel's land use category is set to 1.0, and the values ​​of all other dimensions are set to 0.0. A large number of one-dimensional arrays are thus constructed and stored in the dependent variable region, with the same number of arrays as the independent variable region and a one-to-one correspondence.

[0073] Machine learning model area: Used to store the machine learning model obtained during the training process of the above-mentioned independent variable area data and dependent variable area data.

[0074] Step 2: Divide the array data in the independent variable region into training group and validation group, and select the corresponding array data in the dependent variable region to be grouped into the same group.

[0075] In some possible embodiments of the present invention, the land use conversion probability prediction model is obtained by machine learning training with sample cell-level embedded vector data as input and actual land use information of fine-scale land use and cover raster data of sample base year as training target; wherein, the land use conversion probability prediction model is a random forest model.

[0076] Specifically, the training data is used to train a random forest regression model.

[0077] Step 3: After training, obtain a machine learning model with the same number of land categories as the land category storage area. Use the validation area data to perform validation tests to ensure the accuracy of the model.

[0078] In some possible embodiments of the present invention, the training process of the land use conversion probability prediction model further includes:

[0079] The hyperparameters of the land use conversion probability prediction model were tuned using the Bayesian optimization method.

[0080] In the above embodiments, if the test results are relatively poor, the Bayesian optimization method is used to optimize the machine learning model parameters, and the training and model testing are repeated.

[0081] Specifically, the existing fine-scale land use raster data is first compared with the fine-scale land use raster data obtained by hard downscaling the coarse-scale raster data of the new year, and the increase or decrease of pixels for each land use category is obtained (i.e., which land use categories need to be increased and which land use categories need to be decreased).

[0082] Subsequently, a Fishnet vector dataset with the same spatial resolution as the coarse-scale raster is established. The vector surface feature data of each plot in the vector dataset is used as a mask to extract the existing fine-scale land use raster data (all effective class values ​​must be determined in advance). The number of pixels of fine-scale land use raster data and each land use type within the data range of each plot is counted, and the dominant land use type (i.e., the dominant land use type) within the data range of each plot is determined based on this.

[0083] The following land use optimization process is performed based on land suitability. All land use type pixels that need to be reduced are considered as candidates, and all land use types that need to be added are iterated through. For each land use type that needs to be added, given the number of pixels that need to be added, all candidate pixels are iterated through. Then, based on the land suitability map of the land use type to be added, all candidate pixels are sorted from highest to lowest land suitability, and pixels are selected for modification from highest to lowest suitability. Before modification, it is determined whether the dominant land use type within the corresponding plot vector range will change due to this modification. If it does not change, the land use type of that pixel is modified normally; if the dominant land use type changes, it is not modified, and the pixel with the second highest land suitability is selected for further evaluation.

[0084] In the previous comparison of "existing fine-scale land use raster data with coarse-scale raster data from the new year, obtained by hard downscaling," if the land use value of a cell at the same location in the two raster datasets is the same, that cell is "locked," meaning it does not participate in subsequent land use changes, and the determination of whether a cell's land use can be changed does not take into account the "locked" cell. For each cell whose land use has been modified, it is also "locked."

[0085] After the above process is completed, a map based on land suitability and land use classification will be generated, which is the final downscaled result.

[0086] In some possible embodiments of the present invention, the step of statistically analyzing the changes in the number of fine-scale pixels of various landforms within the target area to obtain the net increase in the number of fine-scale pixels of various landforms in future years specifically includes:

[0087] The number of fine-scale pixels of each type in the target area in the fine-scale data of the base year and the number of fine-scale pixels of each type in the pseudo downscaled fine-scale data corresponding to the coarse-scale data of the future year are counted respectively. The difference between the number of fine-scale pixels of each type is calculated to obtain the net increase of fine-scale pixels of each type in the future year.

[0088] The pseudo-downscaled fine-scale data refers to coarse-scale data of future years that has been resampled to the same resolution.

[0089] Specifically, the coarse-scale land use and cover raster data for future years are compared pixel-by-pixel with the coarse-scale land use and cover raster data for the baseline year to identify pixels where land use types have changed, which are then used as a mask. Simultaneously, the coarse-scale land use and cover raster data for future years are hard-downscaled (i.e., the spatial resolution is directly changed without any additional operations) to obtain pseudo-downscaled fine-scale land use and cover raster data for future years. Pixels from the pseudo-downscaled fine-scale land use and cover raster data for future years and the fine-scale land use and cover raster data for the baseline year are extracted using the mask. The area change and the corresponding number of pixel changes for each land use type within the mask area are statistically analyzed (hereinafter referred to as "net increment of land use type").

[0090] Using the machine learning model obtained from the previous random forest training process, a one-dimensional array consisting of 64 independent variables corresponding to each pixel within the mask area is imported to calculate the suitability (i.e., conversion probability) of each land use type corresponding to a specific pixel, and then normalize it. Following the above method, a two-dimensional array (hereinafter referred to as the "probability matrix") is obtained to store the conversion probability data. Each row in the array stores the conversion probability of each land use type for a pixel. Simultaneously, a dictionary is used to store the row number of the pixel in the array and the row and column coordinates of the pixel in the fine-scale land use and cover raster data of the study area.

[0091] In some possible embodiments of the present invention, the construction of the optimization model further includes setting the following constraints for the optimization model:

[0092] The total change in the number of fine-scale pixels for each land type is consistent with the net increase in the fine-scale pixels.

[0093] Each fine-scale cell can only be assigned one land class.

[0094] In some possible embodiments of the present invention, the optimization model also sets the following constraints:

[0095] If a fine-scale pixel is found to have the same land use type in the fine-scale data of the base year and the pseudo-downscale fine-scale data, then the land use type allocation result of that pixel remains unchanged during the optimization process.

[0096] Specifically, an optimization model is constructed, with the objective of "maximizing the sum of fitness of each pixel," and the following three constraints are set:

[0097] Based on the net increase in land use type, the increase or decrease of pixels for each land use type is set;

[0098] Each cell can be set to one and only one land type;

[0099] If the land use type of a specific pixel in the base year is consistent with the land use type of the corresponding pixel in the pseudo-downscale high-resolution land use and cover raster data of future years, then that pixel will not participate in land use change.

[0100] By importing the probability matrix into the optimization model and importing the output into the corresponding pixels through the dictionary, we can obtain optimized fine-scale land use spatial distribution raster data within the mask area. This data is then overlaid onto the corresponding pixels of the base year's fine-scale land use and cover raster to obtain the final downscaled high spatial resolution land use and cover raster data.

[0101] In some possible embodiments of the present invention, the method for aggregating the fine-scale land use and cover raster data of the reference year is the mode method.

[0102] Specifically, the mode method is a criterion for classifying things based on the frequency of their occurrence. In this embodiment, it refers to counting the occurrence frequency of each land type pixel in the target category set within a given spatial unit, and determining the category that occurs most frequently and reaches a relative majority (not necessarily an absolute majority) as the dominant land type of that unit.

[0103] For example, suppose a fishing net plot contains 100 pixels of 10 m × 10 m each, of which 10 pixels are of category 0 (water), 45 pixels are of category 1 (forest), 40 pixels are of category 4 (arable land), and 5 pixels are of category 6 (construction land). The mode statistics show that the forest land pixel appears most frequently, and the dominant land type of this plot is ultimately determined to be 1 (forest).

[0104] In general, for example, we use 64-dimensional Google Satellite Embedding raster data (10m globally) as independent variables and a 10m land cover map of the baseline year as labels. Each pixel is made into a one-hot vector (the dimension of the land class = 1, and the rest = 0). The embedding is first processed to remove projection differences and then resampled to a 10m grid perfectly aligned with the training area, forming sample pairs (assuming there are k land classes, the training data format is: 64-dimensional vector, 1×k-dimensional one-hot). The training and validation processes are divided into training and validation parts. A random forest classification model is used for training, with a separate regression tree for each class. The probability of "this pixel being classified as 1" is output as the land suitability of this land class. If the validation accuracy is low, parameters are adjusted using Bayesian optimization to automatically search the tree depth, minimum leaf samples, and maximum number of features until the validation set error converges. After training, k regression models are obtained. Then, the 64-dimensional data of the target study area is imported into the models to obtain k land suitability maps for the next step of downscaling experiments.

[0105] Suppose the study area mask contains K pixels, and the land cover set is C={1,2,...,K}. Execute the following procedure:

[0106] Step 1: Decision Variables

[0107] =1, if pixel i is assigned to land class c; otherwise =0.

[0108] Step 2, Probability Matrix (Land Type Suitability Matrix):

[0109] ,and .

[0110] Step 3, Objective Function (Maximizing Fitness):

[0111] .

[0112] Step 4, Constraints:

[0113] Constraint 1: Net increase in land use constraint

[0114] .

[0115] in Indicates whether pixel i in the base year belongs to land category c. This represents the net increase / decrease in land category c.

[0116] Constraint 2: Single land category constraint:

[0117] , .

[0118] Constraint 3: Fixed constraints that do not participate in the change of pixels:

[0119] If the base year and the pseudo-downscaled land use categories are consistent in future years:

[0120] This invention first constructs a database with a unified geographic framework: the land type area stores all possible land types (used to define the dimensionality of subsequent data); the independent variable area stores 64-dimensional satellite embedding vectors, i.e., 64 independent variable raster data; the dependent variable area stores the spatial suitability (transformation probability) of each pixel in the base year using one-hot encoding; and the machine learning model area stores a random forest regression model trained on a land type-by-land type basis, supplemented by an iterative parameter tuning method based on Bayesian optimization to ensure the reliability of the machine learning model.

[0121] A change mask is generated by comparing coarse-scale land use and cover raster data from future years with coarse-scale land use and cover raster data from the baseline year pixel by pixel. Then, the net increment of land use type for each land cover is obtained by statistically analyzing fine-scale land use and cover raster data from the baseline year and hard-downscaled fine-scale land use and cover raster data from future years within the mask area. The aforementioned machine learning model is then used to output and normalize the conversion probability of each land cover type for each pixel within the mask, subsequently forming a probability matrix and a corresponding "row number: coordinate" dictionary.

[0122] Finally, with the objective of "maximizing the sum of suitability of all pixels within the mask area," three constraints, including net increment balance, were set. Land use and cover raster data from the baseline year were imported into the mask area, and the optimal spatial distribution of land types within the mask area was obtained through optimization model solving. These results were then overlaid onto the fine-scale land use and cover raster of the baseline year, outputting downscaled, high-spatial-resolution land use and cover raster data for future years, thus completing the entire downscaling process.

[0123] First, the probabilities of various possibilities are given, and then the optimal solution is sought through optimization. The objective function of the optimization model is "maximizing the overall fitness", so that each piece of land is assigned to its most suitable use (determined by the machine learning model based on historical patterns). This provides a clear and interpretable geographical logic for the spatial pattern of the downscaling results.

[0124] Compared with existing technologies, in terms of data storage and acquisition, this invention completely eliminates the reliance on high-cost surface element vector data and subjective land use conversion rules based on expert experience, reducing the difficulty and cost of data acquisition. Regarding the setting of land use type spatial distribution at a fine scale, this invention optimizes the spatial distribution of land use type pixels in the model, giving physical meaning to the spatial distribution of land use types in each pixel. Simultaneously, because it does not use a unified land use conversion rule for batch processing of fine-scale pixels, it reduces the visual jaggedness of the output raster to some extent. In terms of process execution, this invention can be run with a single click via script, while avoiding the uncertainty of results caused by factors such as differences in expert experience, providing an easily reproducible technical path for land use scenario simulation and land management decision-making.

[0125] The following describes the land use and cover raster data downscaling system for future years provided by the present invention. The land use and cover raster data downscaling system for future years described below can be referred to in correspondence with the land use and cover raster data downscaling method for future years described above.

[0126] In some specific embodiments of the present invention, such as Figure 3 As shown, this solution provides a system for downscaling raster data of land use and cover for future years, including:

[0127] Data acquisition module 10 is used to acquire fine-scale land use and cover raster data, pixel-level embedding vector data and coarse-scale land use and cover raster data for the baseline year, wherein the pixel-level embedding vector data comes from the Google Satellite Embedding dataset released by AlphaEarth Foundations.

[0128] The data acquisition module 10 is also used to aggregate the fine-scale land use and cover raster data of the reference year to generate coarse-scale land use and cover raster data of the reference year.

[0129] The change area determination module 20 is used to compare the coarse-scale land use and cover raster data of the future year with the coarse-scale land use and cover raster data of the baseline year, identify coarse-scale pixels where land use has changed in the future year; map the coarse-scale pixels where land use has changed onto the fine-scale land use and cover raster data of the baseline year, determine the range of fine-scale pixels covered by the coarse-scale pixels, and define the range of fine-scale pixels as the target area.

[0130] The net increment calculation module 30 is used to statistically analyze the changes in the number of fine-scale pixels of each land type in the target area, and to obtain the net increment of fine-scale pixels of each land type in future years.

[0131] The probability prediction module 40 is used to generate a conversion probability corresponding to each land use type for each fine-scale cell within the target area using a land use conversion probability prediction model.

[0132] The spatial optimization module 50 is used to construct an optimization model with the objective function of maximizing the sum of the conversion probabilities corresponding to the land types assigned to all fine-scale pixels in the target area;

[0133] The spatial optimization module 50 is also used to solve the optimization model to obtain the optimal land use allocation result of fine-scale pixels within the target area;

[0134] The data fusion module 60 is used to fuse the optimal land use allocation result with the fine-scale land use and cover raster data of the baseline year to generate downscaled fine-scale land use and cover raster data of future years.

[0135] Optionally, the land use conversion probability prediction model is obtained by machine learning training using sample cell-level embedded vector data as input and actual land use information of fine-scale land use and cover raster data of the sample base year as training target; wherein, the land use conversion probability prediction model is a random forest model.

[0136] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440. The processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a downscaling method for land use and cover raster data from future years. This method includes acquiring fine-scale land use and cover raster data from a baseline year, pixel-level embedded vector data, and coarse-scale land use and cover raster data from future years. The pixel-level embedded vector data originates from Google Satellite data released by AlphaEarth Foundations. Embedding dataset; aggregating the fine-scale land use and cover raster data of the baseline year to generate coarse-scale land use and cover raster data of the baseline year; comparing the coarse-scale land use and cover raster data of future years with the coarse-scale land use and cover raster data of the baseline year to identify coarse-scale pixels where land use types change in future years; mapping the coarse-scale pixels where land use types change onto the fine-scale land use and cover raster data of the baseline year to determine the range of fine-scale pixels covered by the coarse-scale pixels, and defining the range of fine-scale pixels as the target area; statistically analyzing the changes in the number of fine-scale pixels for each land use type within the target area to obtain the net increase in fine-scale pixels for each land use type in future years; within the target area, utilizing... A land use conversion probability prediction model is used to generate conversion probabilities corresponding to each land use type for each fine-scale pixel. This model is trained via machine learning, using sample pixel-level embedded vector data as input and the actual land use information from the fine-scale land use and cover raster data of the baseline year as the training objective. An optimization model is constructed with the objective function of maximizing the sum of conversion probabilities corresponding to the land use types assigned to all fine-scale pixels within the target area. The optimization model is solved to obtain the optimal land use type allocation result for the fine-scale pixels within the target area. This optimal allocation result is then fused with the fine-scale land use and cover raster data of the baseline year to generate downscaled fine-scale land use and cover raster data for future years.

[0137] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0138] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the land use and cover raster data downscaling method for future years provided by the methods described above. This method includes: acquiring fine-scale land use and cover raster data of the baseline year, pixel-level embedded vector data, and coarse-scale land use and cover raster data for future years. The pixel-level embedded vector data is sourced from Google Satellite data released by AlphaEarthFoundations. Embedding dataset; aggregating the fine-scale land use and cover raster data of the baseline year to generate coarse-scale land use and cover raster data of the baseline year; comparing the coarse-scale land use and cover raster data of future years with the coarse-scale land use and cover raster data of the baseline year to identify coarse-scale pixels where land use types change in future years; mapping the coarse-scale pixels where land use types change onto the fine-scale land use and cover raster data of the baseline year to determine the range of fine-scale pixels covered by the coarse-scale pixels, and defining the range of fine-scale pixels as the target area; statistically analyzing the changes in the number of fine-scale pixels for each land use type within the target area to obtain the net increase in fine-scale pixels for each land use type in future years; within the target area, utilizing... A land use conversion probability prediction model is used to generate conversion probabilities corresponding to each land use type for each fine-scale pixel. This model is trained via machine learning, using sample pixel-level embedded vector data as input and the actual land use information from the fine-scale land use and cover raster data of the baseline year as the training objective. An optimization model is constructed with the objective function of maximizing the sum of conversion probabilities corresponding to the land use types assigned to all fine-scale pixels within the target area. The optimization model is solved to obtain the optimal land use type allocation result for the fine-scale pixels within the target area. This optimal allocation result is then fused with the fine-scale land use and cover raster data of the baseline year to generate downscaled fine-scale land use and cover raster data for future years.

[0139] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for downscaling land use and cover raster data for future years provided by the methods described above. This method includes: acquiring fine-scale land use and cover raster data for a baseline year, pixel-level embedded vector data, and coarse-scale land use and cover raster data for future years, wherein the pixel-level embedded vector data originates from Google Satellite data released by AlphaEarth Foundations. Embedding dataset; aggregating the fine-scale land use and cover raster data of the baseline year to generate coarse-scale land use and cover raster data of the baseline year; comparing the coarse-scale land use and cover raster data of future years with the coarse-scale land use and cover raster data of the baseline year to identify coarse-scale pixels where land use types change in future years; mapping the coarse-scale pixels where land use types change onto the fine-scale land use and cover raster data of the baseline year to determine the range of fine-scale pixels covered by the coarse-scale pixels, and defining the range of fine-scale pixels as the target area; statistically analyzing the changes in the number of fine-scale pixels for each land use type within the target area to obtain the net increase in fine-scale pixels for each land use type in future years; within the target area, utilizing... A land use conversion probability prediction model is used to generate conversion probabilities corresponding to each land use type for each fine-scale pixel. This model is trained via machine learning, using sample pixel-level embedded vector data as input and the actual land use information from the fine-scale land use and cover raster data of the baseline year as the training objective. An optimization model is constructed with the objective function of maximizing the sum of conversion probabilities corresponding to the land use types assigned to all fine-scale pixels within the target area. The optimization model is solved to obtain the optimal land use type allocation result for the fine-scale pixels within the target area. This optimal allocation result is then fused with the fine-scale land use and cover raster data of the baseline year to generate downscaled fine-scale land use and cover raster data for future years.

[0140] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0141] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for downscaling raster data of land use and cover for future years, characterized in that, include: The data includes fine-scale land use and cover raster data for the baseline year, pixel-level embedding vector data, and coarse-scale land use and cover raster data for future years. The pixel-level embedding vector data is sourced from the Google Satellite Embedding dataset released by AlphaEarth Foundations. The fine-scale land use and cover raster data of the base year are aggregated to generate coarse-scale land use and cover raster data of the base year. By comparing the coarse-scale land use and cover raster data of the future years with the coarse-scale land use and cover raster data of the baseline year, coarse-scale pixels in which land use types change in the future years are identified. The coarse-scale pixels showing changes in land use are mapped onto the fine-scale land use and cover raster data of the baseline year to determine the range of fine-scale pixels covered by the coarse-scale pixels, and the range of fine-scale pixels is defined as the target area. The changes in the number of fine-scale pixels for each land type within the target area are statistically analyzed to obtain the net increase in fine-scale pixels for each land type in future years, specifically including: The number of fine-scale pixels of each type in the target area in the fine-scale data of the base year and the number of fine-scale pixels of each type in the pseudo-downscale fine-scale data corresponding to the coarse-scale data of the future year are counted respectively. The difference between the number of fine-scale pixels of each type is calculated to obtain the net increase of fine-scale pixels of each type in the future year. The pseudo-downscale fine-scale data refers to the coarse-scale data of the future year that has been resampled to the same resolution. Within the target area, a land use conversion probability prediction model is used to generate conversion probabilities corresponding to each land use type for each fine-scale cell; An optimization model is constructed with the objective function of maximizing the sum of the conversion probabilities corresponding to the land use types assigned to all fine-scale pixels within the target area; the optimal land use type assignment result for fine-scale pixels within the target area is obtained by solving the optimization model. The optimal land use classification result is fused with the fine-scale land use and cover raster data of the baseline year to generate downscaled fine-scale land use and cover raster data for future years.

2. The method for downscaling land use and cover raster data for future years according to claim 1, characterized in that, The construction of the optimization model also includes setting the following constraints for the optimization model: The total change in the number of fine-scale pixels for each land type is consistent with the net increase in the fine-scale pixels. Each fine-scale cell can only be assigned one land class.

3. The method for downscaling land use and cover raster data for future years according to claim 2, characterized in that, The optimization model also sets the following constraints: If a fine-scale pixel is found to have the same land use type in the fine-scale data of the base year and the pseudo-downscale fine-scale data, then the land use type allocation result of that pixel remains unchanged during the optimization process. The pseudo-downscaled fine-scale data refers to coarse-scale data of future years that has been resampled to the same resolution.

4. The method for downscaling land use and cover raster data for future years according to claim 1, characterized in that, The pixel-level embedded vector data is a high-dimensional vector obtained by feature extraction of satellite image blocks through a deep neural network, which is pre-trained on satellite images using a self-supervised learning method.

5. The method for downscaling land use and cover raster data for future years according to claim 1, characterized in that, The land use conversion probability prediction model is obtained by training through machine learning, using sample pixel-level embedded vector data as input and actual land use and cover raster data of the sample base year as training targets; wherein, the land use conversion probability prediction model is a random forest model.

6. The method for downscaling land use and cover raster data for future years according to claim 5, characterized in that, The training process of the land use conversion probability prediction model also includes: The hyperparameters of the land use conversion probability prediction model were tuned using the Bayesian optimization method.

7. The method for downscaling land use and cover raster data for future years according to claim 1, characterized in that, The method for aggregating the fine-scale land use and cover raster data of the base year is the mode method.

8. A system for downscaling raster data of land use and cover for future years, characterized in that, include: The data acquisition module is used to acquire fine-scale land use and cover raster data, pixel-level embedding vector data and coarse-scale land use and cover raster data for the baseline year, wherein the pixel-level embedding vector data comes from the Google Satellite Embedding dataset released by AlphaEarth Foundations. The data acquisition module is also used to aggregate the fine-scale land use and cover raster data of the reference year to generate coarse-scale land use and cover raster data of the reference year. The change area determination module is used to compare the coarse-scale land use and cover raster data of the future year with the coarse-scale land use and cover raster data of the baseline year to identify coarse-scale pixels where land use types have changed in the future year. The coarse-scale pixels showing changes in land use are mapped onto the fine-scale land use and cover raster data of the baseline year to determine the range of fine-scale pixels covered by the coarse-scale pixels, and the range of fine-scale pixels is defined as the target area. The net increment calculation module is used to statistically analyze the changes in the number of fine-scale pixels for each land type within the target area, and to obtain the net increment of fine-scale pixels for each land type in future years. Specifically, it is used for: The number of fine-scale pixels of each type in the target area in the fine-scale data of the base year and the number of fine-scale pixels of each type in the pseudo-downscale fine-scale data corresponding to the coarse-scale data of the future year are counted respectively. The difference between the number of fine-scale pixels of each type is calculated to obtain the net increase of fine-scale pixels of each type in the future year. The pseudo-downscale fine-scale data refers to the coarse-scale data of the future year that has been resampled to the same resolution. The probability prediction module is used to generate the conversion probability corresponding to each land use type for each fine-scale cell within the target area using a land use conversion probability prediction model. The land use conversion probability prediction model is obtained by machine learning training with sample cell-level embedded vector data as input and actual land use information of fine-scale land use and cover raster data of the sample reference year as the training target. The spatial optimization module is used to construct an optimization model with the objective function of maximizing the sum of the conversion probabilities corresponding to the land types assigned to all fine-scale pixels within the target area. The spatial optimization module is also used to solve the optimization model to obtain the optimal land use classification result for fine-scale pixels within the target area; The data fusion module is used to fuse the optimal land use allocation result with the fine-scale land use and cover raster data of the baseline year to generate downscaled fine-scale land use and cover raster data for future years.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method for downscaling land use and cover raster data for future years as described in any one of claims 1 to 7.

Images