Land change prediction system based on improved PINN network

Abstract

The invention discloses a land change prediction system based on an improved PINN network, and the system comprises a collection and processing module which is used for collecting land change data and carrying out the preprocessing of the data, and obtaining the processed data; the model improvement module is used for improving the loss function on the basis of a traditional PINN network to obtain a PINN network model for optimizing the improved loss function; the model training module is used for training a PINN network model for improving loss function optimization through dynamic adjustment of an exponential decay learning rate strategy by using an Adam optimization algorithm; the prediction module is used for predicting the land to be predicted by using the trained model; according to the invention, high-precision land change prediction conforming to physical rules is realized, a system with comprehensive functions and convenient operation is provided, and scientific management of land resources is assisted.

Description

Technical Field

[0001] The present invention relates to a land change prediction system, and particularly to a land change prediction system based on an improved PINN network, belonging to the technical fields of artificial intelligence, geographic information science, and environmental science and technology. Background Art

[0002] Land use and cover change is an important content in the study of global environmental change, which has a profound impact on ecological environment, urban planning, resource management, etc. With the acceleration of urbanization process, population growth, and climate change and other factors, land change presents a complex and dynamic situation.

[0003] Traditional land change prediction methods have many limitations. Statistical models rely on simple correlation analysis and are difficult to handle complex non-linear relationships; although the cellular automata model can simulate spatial dynamics, the state transition rules are highly subjective; economic models consider non-economic factors insufficiently; agent-based models have high computational costs and demanding data requirements.

[0004] Physics-informed neural network (PINN), as an emerging deep learning method, integrates physical laws into neural networks and shows advantages in solving complex scientific problems. However, when the traditional PINN network is applied to land change prediction, there are problems such as poor adaptability to the complex physical processes of land change, low processing efficiency for large-scale geospatial data, and room for improvement in prediction accuracy. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to provide a land change prediction system based on an improved PINN network that can improve the prediction accuracy.

[0006] Technical Solution: A land change prediction system based on an improved PINN network according to the present invention includes:

[0007] A data collection and processing module, configured to collect land change data and perform preprocessing to obtain processed data; the land change data includes land use status data, topographic and geomorphic data, social and economic data, climate data, and soil data;

[0008] A model improvement module, configured to improve the loss function based on the traditional PINN network to obtain a PINN network model for optimizing the improved loss function. The improved loss function includes a data fitting loss and a physical constraint loss. The data fitting loss uses the mean square error loss function to measure the difference between the predicted value of land use data and the land use status data, and the physical constraint loss uses the logic of land total conservation and land type conversion for measurement;

[0009] A model training module for training a PINN network model that optimizes an improved loss function by using the Adam optimization algorithm and dynamically adjusting it through an exponential decay learning rate strategy.

[0010] A prediction module for predicting the land to be predicted using the trained model.

[0011] Further, in the collection and processing module, the preprocessing includes data cleaning, filling missing values, and format standardization.

[0012] Further, in data cleaning, the 3-sigma rule is used to remove outliers, linear regression algorithm is used to fill missing values, and the Pandas data analysis library and UTM projection conversion algorithm are used for format standardization.

[0013] Further, in the collection and processing module, the current land use data is obtained by remote sensing image interpretation, the topographic and geomorphic data is obtained from digital elevation model (DEM) data, and the social and economic data is obtained from statistical websites.

[0014] Further, the collection and processing module further includes extracting slope and aspect features from the topographic and geomorphic data. Specifically:

[0015] The slope calculation formula is:

[0016]

[0017] where Slope is the slope, z is the elevation value, and x and y are the planar coordinates, and are the partial derivatives of z with respect to x and y respectively, calculated by the numerical difference method;

[0018] The aspect calculation formula is:

[0019]

[0020] where Aspect is the aspect.

[0021] Further, in the PINN network model that optimizes the improved loss function,

[0022] the mean squared error loss function L data has the formula:

[0023]

[0024] where y i is the i-th current land use data sample, is the predicted value of the i-th land use data, and the number of samples is N;

[0025] The total land quantity conservation constraint loss \(L\) area is formulated as follows:

[0026]

[0027] where \(A\) j is the actual area of the \(j\)-th land type, is the predicted area of the \(j\)-th land type, the total number of land types is \(M\), and the total area is \(S\);

[0028] Dividing the area by the preset reference area \(S_0\) gives the relative area and The loss function of the PINN network becomes:

[0029]

[0030] where,

[0031] The land type conversion logic constraint loss \(L\) trans is formulated as follows:

[0032]

[0033] where \(P\) is the probability matrix, and \(P\) jk represents the probability of converting from land type \(j\) to land type \(k\); \(n\) jk is the actually observed number of conversions from land type \(j\) to land type \(k\), is the predicted value of the conversion quantity;

[0034] Considering normalization, let \(N\) total be the total number of observed land conversions, then the formula for the loss function is:

[0035]

[0036] where,

[0037] Combining the data fitting loss and the physical constraint loss, the total loss function \(L\) is obtained:

[0038] \(L = L\) data +\(\lambda_1L\) area +\(\lambda_2L\) trans

[0039] where \(\lambda_1\) and \(\lambda_2\) are weight coefficients used to balance the importance of different loss terms.

[0040] Furthermore, the model training module is specifically implemented as follows:

[0041] Divide the data processed by the collection and processing module into a training set, a validation set, and a test set according to a preset ratio;

[0042] Input the training set into the PINN network model optimized by the improved loss function, use the Adam optimization algorithm and the exponential decay learning rate strategy, update the model parameters according to the formula, perform multiple rounds of training, and minimize the loss function;

[0043] During the training process, evaluate the model performance using the validation set at a preset frequency, and calculate the loss L val on the validation set, accuracy, recall, precision, and root mean square error, observe whether the model is overfitting or underfitting, and adjust the model hyperparameters according to the validation results;

[0044] After the training is completed, evaluate the model using the test set based on socioeconomic data, climate data, and soil data.

[0045] Furthermore, when the prediction module uses the trained model to predict the land to be predicted, it combines the slope and aspect for prediction.

[0046] Furthermore, it further includes: an interaction module for providing a user interaction interface, displaying prediction results and data visualization, and realizing data import, export, update, and query.

[0047] Furthermore, the interaction module realizes data import, export, update, and query in the user interaction interface; provides users with the configuration of training parameters and monitors the training progress; displays prediction results for users and generates reports, and displays data and results in maps or charts.

[0048] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: (1) Improve the PINN network, incorporate physical constraints, focus on key factors, allocate the network structure, effectively learn complex land change laws, and predict more accurately than traditional models, overcoming the problem that traditional models are difficult to capture complex relationships; (2) Adopt the Adam optimization algorithm and the exponential decay learning rate strategy to accelerate model convergence and shorten the training time. Compared with traditional training strategies, the model training can be completed faster, improving the training efficiency; (3) Provide an interaction interface covering functions such as data management, parameter configuration, and progress monitoring, and can also visually display the results, with convenient operation, in contrast to the lack of intuitive interaction in traditional technologies, improving the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the system module diagram of the embodiment of the present invention;

[0050] Figure 2 It is the improved PINN network structure framework of the embodiment of the present invention;

[0051] Figure 3Comparison chart of the predicted results of land use type distribution in the embodiments of the present invention;

[0052] Figure 4 Comparison chart of the predicted results of the change in land type area in the embodiments of the present invention. Detailed implementation manners

[0053] In order to enable those skilled in the art to better understand the solution of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this application.

[0054] As shown in the attached Figure 1 drawing, the land change prediction system based on the improved PINN network in this embodiment includes:

[0055] A data collection and processing module, configured to collect land change data and perform preprocessing to obtain processed data; the land change data includes land use status data, topographic and geomorphic data, social and economic data, climate data, and soil data;

[0056] A model improvement module, configured to improve the loss function based on the traditional PINN network to obtain a PINN network model for optimizing the improved loss function. The improved loss function includes a data fitting loss and a physical constraint loss. The data fitting loss uses the mean square error loss function to measure the difference between the predicted value of land use data and the land use status data, and the physical constraint loss is measured using the logic of land total conservation and land type conversion;

[0057] A model training module, configured to use the Adam optimization algorithm and dynamically adjust through an exponentially decaying learning rate strategy to train the PINN network model for optimizing the improved loss function;

[0058] A prediction module, configured to use the trained model to predict the land to be predicted;

[0059] An interaction module, configured to provide a user interaction interface, display the prediction results and data visualization, and implement data import, export, update, and query.

[0060] Specifically, in the data collection and processing module, the method for obtaining and processing the land change-related data is specifically as follows:

[0061] (1) The land use status data is obtained by interpreting high-resolution remote sensing images; the topographic and geomorphic data is extracted from digital elevation model (DEM) data; the social and economic data collects statistical yearbooks published by statistical departments, covering information such as population quantity, GDP, and industrial structure.

[0062] (2) The data preprocessing part includes data cleaning and format standardization;

[0063] (3) Extract features such as slope and aspect from the terrain data. The slope calculation formula is:

[0064]

[0065] where z is the elevation value, and x and y are the planar coordinates, and are the partial derivatives of z with respect to x and y respectively, which can be calculated by the numerical difference method.

[0066] The aspect calculation formula is:

[0067]

[0068] (4) For categorical data, one-hot encoding is used for feature encoding regarding land use types; for numerical features, normalization or standardization processing is adopted. The normalization formula is:

[0069]

[0070] where x is the original feature value, x min and x max are the minimum and maximum values of this feature respectively.

[0071] The standardization formula is:

[0072]

[0073] where μ is the mean of the feature and σ is the standard deviation of the feature.

[0074] In the model improvement module, as Figure 2 shown, the PINN network model optimized according to the improved loss function is specifically:

[0075] (1) The data fitting term is the mean square error loss function, and the formula for the mean square error loss function L data is:

[0076]

[0077] where y i is the i-th land use status data sample, is the predicted value of the i-th land use data predicted by the model, and the number of samples is N.

[0078] (2) The formula for the land total amount conservation constraint loss L area is:

[0079]

[0080] where A jis the actual area of the j-th land type, is the area of the j-th land type predicted by the model. The total number of land types is M, and the total area is S.

[0081] In actual calculations, normalization and other processes need to be performed according to the specific situation of the data. Divide the area by a preset reference area S0 to obtain the relative area and Then the loss function of the PINN network becomes:

[0082]

[0083] Among them,

[0084] (3) The logical constraint loss L of land type conversion trans The formula is:

[0085]

[0086] Among them, there is a certain probability matrix P for land type conversion. P jk represents the probability of converting from land type j to land type k. n jk is the number of conversions from land type j to land type k actually observed, is the number of conversions predicted by the model.

[0087] If normalization is considered, let N total be the total number of observed land conversions, then the formula of the loss function is:

[0088]

[0089] Among them,

[0090] (4) Combine the data fitting loss and the physical constraint loss to obtain the total loss function L:

[0091] L = L data + 1L area + λ2L trans

[0092] Among them, λ1 and λ2 are weight coefficients used to balance the importance of different loss terms and are adjusted by methods such as cross-validation according to the actual situation.

[0093] In the model training module, the Adam optimization algorithm is specifically:

[0094] (1) Prepare the training dataset D related to land changes train, input the data into the PINN network model that improves the loss function optimization, use the Adam optimization algorithm and the exponential decay learning rate strategy, update the model parameters according to the formula, perform multiple rounds of training, and minimize the loss function.

[0095] (2) Prepare the validation dataset D val , regularly evaluate the model performance using the validation set during the training process, and calculate the loss L on the validation set val and other evaluation metrics such as accuracy, recall, etc., observe whether the model is overfitting or underfitting, and adjust the model hyperparameters according to the validation results.

[0096] (3) After the training is completed, use the test dataset D test to conduct the final evaluation of the model, calculate the evaluation metrics on the test set, evaluate the ability of the model to accurately identify land changes, and if the metrics do not meet the requirements, the model can be further adjusted or retrained.

[0097] Let the model parameters be θ, including the weight W and the bias b, initialize the first moment estimate m0 = 0, the second moment estimate v o = 0, the time step t = 0, the learning rate α, the exponential decay rates β1, β2 of the moment estimates, usually β1 = 0.9, β2 = 0.999, and the small constant ε = 10 to prevent division by zero -8 .

[0098] At each step t of the training process, for the given training data x and the corresponding label y, calculate the gradient of the loss function L(θ) with respect to the parameter θ

[0099] The first moment estimate m t = β1m t-1 + (1 - β1)g t , the second moment estimate v t = β2v t-1 + (1 - β2)g t 2 . Perform bias correction, and the corrected first moment estimate is The corrected second moment estimate The parameter update formula is

[0100] The exponential decay learning rate α t The calculation formula of is α t = α0γ t . Where the initial learning rate is α0, the decay rate is γ, and the current training round is t.

[0101] In the interaction module, in the user interaction interface, data import, export, update, and query are implemented through the data management module; users can configure training parameters and monitor the progress; call the model prediction according to the input conditions and generate a report, and finally visually display the data and results in the form of maps and charts.

[0102] Compare the prediction of the change trends of the areas of cultivated land, forest land, and construction land over time (from 2020 to 2030) by the traditional CA-GIS model, the unimproved PINN model, and the method used in the system of this embodiment.

[0103] As Figure 3 shown, the land use type distribution map is presented in the form of a map, where different colors represent different land use types. Green represents forest land, yellow represents cultivated land, and gray represents construction land, etc. The land use type distribution in 2030 predicted by the three methods is shown respectively. It can be visually seen from the figure that the traditional CA-GIS model has a relatively blurred division of land types in some border areas, and there is a problem of unclear transition areas; the unimproved PINN model is not accurate enough in distinguishing forest land and grassland in some complex terrain areas, such as mountainous areas; while the land use type distribution map drawn by the improved PINN network model has clear boundaries, and the division of different land types in areas such as complex terrain and urban fringes is more accurate.

[0104] As Figure 4 shown, the change in land type area is presented in the form of a line chart. The abscissa is from 2020 to 2030, and the ordinate is the land type area. The change trends of the areas of the main land types such as cultivated land, forest land, and construction land predicted by the three methods over time are shown. Blue represents the prediction result of the traditional CA-GIS model, green represents the result of the unimproved PINN model, and red represents the prediction result of the system of this embodiment. It can be seen from the line chart that the growth trend of construction land predicted by the traditional CA-GIS model is relatively gentle and fails to accurately reflect the explosive growth of construction land during the rapid urbanization process; although the unimproved PINN model captures the growth trend of construction land, the prediction of the reduction in cultivated land area fluctuates greatly and has a large deviation from the actual situation; the change trend of the area of each land type predicted by the improved PINN network model has the highest degree of fitting with the actual statistical data and can more accurately reflect the dynamic change of the land type area.

[0105] The above is only a preferred embodiment of the present invention and is not used to limit the protection scope of the present invention.

Claims

1. A land change prediction system based on an improved PINN network, characterized in that, Including: A collection and processing module for collecting land change data and performing preprocessing to obtain processed data; The land change data includes current land use data, topographic and geomorphic data, socio-economic data, climate data, and soil data; A model improvement module for improving the loss function based on the traditional PINN network to obtain a PINN network model optimized by the improved loss function. The improved loss function includes a data fitting loss and a physical constraint loss. The data fitting loss uses the mean squared error loss function to measure the difference between the predicted value of land use data and the current land use data, and the physical constraint loss uses the logic of land total conservation and land type conversion to measure; A model training module for using the Adam optimization algorithm and dynamically adjusting through an exponentially decaying learning rate strategy to train the PINN network model optimized by the improved loss function; A prediction module for using the trained model to predict the land to be predicted.

2. The land change prediction system based on the improved PINN network according to claim 1, wherein In the collection and processing module, the preprocessing includes data cleaning, filling missing values, and format standardization.

3. The land change prediction system based on the improved PINN network according to claim 2, characterized in that The data cleaning uses the 3-sigma principle to remove outliers, the filling of missing values uses the linear regression algorithm, and the format standardization uses the Pandas data analysis library and the UTM projection conversion algorithm.

4. The land change prediction system based on the improved PINN network according to claim 1, wherein, In the collection and processing module, the current land use data is obtained through remote sensing image interpretation, the topographic and geomorphic data is obtained from the digital elevation model (DEM) data, and the socio-economic data is obtained through statistical websites.

5. The land change prediction system based on the improved PINN network according to claim 1, characterized in that, The collection and processing module also includes extracting slope and aspect features from the topographic and geomorphic data. Specifically: The slope calculation formula is: Where Slope is the slope, z is the elevation value, and x and y are the plane coordinates, and are the partial derivatives of z with respect to x and y respectively, and are obtained by numerical difference method; The aspect calculation formula is: where Aspect is the aspect.

6. The land change prediction system based on the improved PINN network according to claim 1, wherein In the PINN network model optimized by the improved loss function, The mean square error loss function L data has the following formula: where y i is the i-th current land use data sample, is the predicted value of the i-th land use data, and the number of samples is N; The total land quantity conservation constraint loss L area is formulated as follows: where A j is the actual area of the j-th land type, is the predicted area of the j-th land type, the total number of land types is M, and the total area is S; Divide the area by a preset reference area S0 to obtain the relative area and Change the loss function of the PINN network to: Among them, The land type conversion logic constraint loss L trans is formulated as follows: Among them, P is the probability matrix, and P jk represents the probability of converting from land type j to land type k; n jk is the actually observed number of conversions from land type j to land type k, is the predicted value of the conversion quantity; Considering normalization, let N total be the total number of observed land conversions. Then the formula for the loss function is: Among them, Combining the data fitting loss and the physical constraint loss to obtain the total loss function L: L = L data + λ1L area + λ2L trans where λ1 and λ2 are weight coefficients used to balance the importance of different loss terms.

7. The land change prediction system based on the improved PINN network according to claim 1, characterized in that The specific implementation method of the model training module includes: Dividing the data processed by the collection and processing module into a training set, a validation set, and a test set according to a preset ratio; Inputting the training set into the PINN network model optimized by the improved loss function, using the Adam optimization algorithm and the exponentially decaying learning rate strategy, updating the model parameters according to the formula, and performing multiple rounds of training to minimize the loss function; During the training process, the performance of the model is evaluated using the validation set at a preset frequency, and the loss L on the validation set is calculated. val , accuracy, recall, precision, and root mean square error are calculated, and it is observed whether the model is overfitting or underfitting. The hyperparameters of the model are adjusted according to the validation results. After training is completed, based on the socio-economic data, climate data, and soil data, use the test set to evaluate the model.

8. The land change prediction system based on the improved PINN network according to claim 5, characterized in that When the prediction module uses the trained model to predict the land to be predicted, it combines the slope and aspect for prediction.

9. The land change prediction system based on the improved PINN network according to claim 1, wherein It also includes: An interaction module for providing a user interaction interface, displaying prediction results and data visualization, and implementing data import, export, update, and query.

10. The land change prediction system based on the improved PINN network according to claim 9, characterized in that, The interaction module, in the user interaction interface, implements data import, export, update, and query; provides users with the configuration of training parameters and monitors the training progress; displays prediction results for users and generates reports, presenting data and results in maps or charts.

Images