Distributed remote sensing image classification method based on alternating direction multiplier method
Through the distributed remote sensing image classification method based on the alternating direction multiplier method, the convergence speed and generalization performance problems of the remote sensing classification model in large-scale and long-term sequencing geographic information extraction are solved, and faster model convergence and higher classification accuracy are achieved.
Patent Information
- Application Number
- CN202510581549.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-07
- Publication Date
- 2025-07-18
AI Technical Summary
When handling the extraction of large-scale and long-term geographic information, the existing remote sensing classification model faces the problems of limited convergence speed, strict assumptions and poor generalization performance. Especially in cross-large-scale classification tasks, the acquisition cost of multimodal remote sensing data is high and the data island phenomenon is serious.
The distributed remote sensing image classification method based on the alternating direction multiplier method is adopted. By distributed storage of remote sensing image data sets and model optimization is performed using the alternating direction multiplier method framework, the distributed decomposition of the global model and the global aggregation of the local model are realized, and the assumption requirements of convergence conditions are reduced.
The convergence speed and generalization performance of the model are improved, the applicability and robustness of the method are enhanced, and the problem of remote sensing image classification in non-independent and same-distribution environments is solved.
Smart Images
Figure CN120339721A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of remote sensing image classification, and particularly relates to a distributed remote sensing image classification method based on the alternating direction multiplier method. Background Art
[0002] In the context of the era of remote sensing big data, remote sensing classification, as a key technical means to explore the value of massive remote sensing data, has become an important way to achieve long-term dynamic monitoring and multi-scale ground object information extraction. However, affected by multiple factors such as the characteristics of the underlying surface, atmospheric conditions, solar radiation intensity, and sensor performance, remote sensing data at the same geographical location in different time phases or the same time phase or different geographical locations shows significant spatio-temporal heterogeneity, which restricts the generalization performance of remote sensing classification models and makes it difficult to effectively meet the requirements of large-scale and long-term ground object information extraction.
[0003] To improve the generalization ability and classification accuracy of remote sensing classifiers, the academic community has proposed a variety of improvement methods:
[0004] 1) Semi-supervised classification methods expand unlabeled samples with high confidence by using a small number of labeled samples. However, there is an error accumulation effect, and as the number of iterations increases, misclassified samples may cause the model performance to continue to decline;
[0005] 2) Active learning dynamically selects samples that are most valuable for model improvement through algorithms for annotation. However, the cost of sample annotation in active learning algorithms is high, and especially when querying and annotating high-dimensional data (such as hyperspectral remote sensing images), it still faces the "curse of dimensionality" problem;
[0006] 3) Transfer learning transfers the model trained in the source domain to the target domain, including homogeneous transfer learning and heterogeneous transfer learning.
[0007] Although the above methods reduce the dependence on labeled samples to a certain extent and improve the generalization ability of remote sensing classifiers, when dealing with cross-large-scale classification tasks (such as cross-climate zones, cross-decades), due to the significant variation of the spectral characteristics of the same ground object with climate and the underlying surface, etc., the generalization ability of existing methods still cannot meet the application requirements. In addition, remote sensing data also presents multi-modal characteristics, such as multi-spectral remote sensing images, synthetic aperture radar images, hyperspectral remote sensing images, and lidar point cloud data, etc. Multi-modal data fusion can extract complementary features of ground objects and enhance the diversity of sample data. However, for long-term or large-scale remote sensing classification tasks, the acquisition cost of multi-modal data is not only high, but also the willingness of sample data owners to share is low.
[0008] With the development of remote sensing technology and the increasing popularity of its applications, the scale of remote sensing image data has shown explosive growth, and sample data for various tasks is also accumulating rapidly. However, due to factors such as difficulties in cross-departmental collaboration, data security and privacy protection, the phenomenon of "data islands" is still prevalent in the remote sensing field. How to break through data barriers and efficiently integrate scattered sample data to improve the generalization ability of the model, thereby improving the accuracy of remote sensing classification, is a key issue that needs to be urgently addressed in this field.
[0009] Federated learning, as a privacy-preserving distributed machine learning paradigm, provides a feasible solution to this problem. This method allows multiple participants to collaboratively train a global model without sharing the original data, and improves classification performance through distributed aggregation of model parameters. In response to the common non-independent and identically distributed characteristics of remote sensing data, researchers have proposed a variety of federated learning algorithms. These algorithms are mainly divided into two categories: one is the optimization algorithm focusing on local training, which effectively constrains the deviation of local models from global goals by introducing neighboring regularization terms or gradient correction mechanisms; the other is the improvement method focusing on aggregation strategies, which uses techniques such as parameter normalization or personalized weighting to deal with data distribution differences. However, the existing methods are based on the traditional gradient descent framework and still face inherent defects such as limited convergence speed and strict assumptions. In particular, when dealing with land feature data with significant non-independent and identically distributed characteristics, there is still a bottleneck in improving the generalization performance of the global model. Summary of the invention
[0010] The purpose of the present invention is to provide a distributed remote sensing image classification method based on the alternating direction multiplier method to address the above-mentioned deficiencies in the prior art, so as to solve the problems faced by the prior art, such as limited convergence speed, strict assumptions and poor generalization performance of the global model.
[0011] In order to achieve the above object, the technical solution adopted by the present invention is:
[0012] A distributed remote sensing image classification method based on an alternating direction multiplier method comprises the following steps:
[0013] S1. Collect remote sensing image data and construct a remote sensing image dataset;
[0014] S2, distributing and storing the remote sensing image data set in each node;
[0015] S3, deploy information to the central server and local models of each node;
[0016] S4. The central server aggregates the parameter information of the local models on each node, and then constructs a global model, and sends the global model to each node in parallel;
[0017] S5. Each node trains and updates its local model based on the received global model and remote sensing image dataset, sends the parameter information of the locally trained and updated model to the central server, and executes S4 until the local models of all current nodes meet the convergence condition, then executes S6;
[0018] S6. Classify the newly acquired remote sensing image data using the global model output by the central server.
[0019] Furthermore, in S3, the loss function deployed by each node's local model is:
[0020]
[0021] In the formula, F(w i ; D i ) represents the loss function of the machine learning model deployed by the i-th node; w i is the local model variable of the i-th node; D i is the local training data stored by the i-th node; is the l-th remote sensing image in D i , is the corresponding label; L CE is the cross-entropy loss function; f represents the machine learning model for remote sensing image classification; := means definition;
[0022] The regularized expression of the loss function of each node's local model:
[0023]
[0024] In the formula, F R (w i ; D i ) is the loss function of the machine learning model deployed by the i-th node after regularization; μ is the regularization hyperparameter.
[0025] Furthermore, in S3, the objective function Q i (w i ) of each node's local model deployment is:
[0026]
[0027] In the formula, D = {D1, D2,..., D K}, represents the data set stored by each node; represents the gradient of the function F; w is the global model variable of the central server; λ i is the multiplier variable of the i-th node; β i , τ i represent model parameters, β i , τ i>0, i = 1, 2, … K, where K is the number of nodes.
[0028] Furthermore, in S3, the objective function deployed by the central server is:
[0029]
[0030] In the formula, Q(w) is the objective function deployed by the central server.
[0031] Furthermore, in S4, the central server aggregates the parameter information of the local models on each node, and then constructs a global model;
[0032] Among them, the global model is expressed as:
[0033]
[0034] In the formula, w t represents the global model of the central server; represents the multiplier variable of the i-th node in the (t - 1)-th round of iteration; represents w i represents the local model of the i-th node in the (t - 1)-th round of iteration.
[0035] Furthermore, in S5, each node trains and updates the local model according to the received global model and the remote sensing image dataset;
[0036] Among them, the updated local model after training is expressed as:
[0037]
[0038] In the formula, is the updated local model after training.
[0039] Furthermore, in S5, the convergence condition is that the change amount of the local models of each node in two consecutive rounds is less than a threshold, and this convergence condition is specifically expressed as:
[0040]
[0041] In the formula, ε represents the threshold.
[0042] Furthermore, in S6, the global model output by the central server is:
[0043]
[0044] In the formula, w * is the final global model output by the central server; represents the multiplier variable of the i-th node in the t-th round of iteration.
[0045] The distributed remote sensing image classification method based on the alternating direction method of multipliers provided by the present invention has the following
[0046] Advantages:
[0047] Based on the framework of the alternating direction method of multipliers, the present invention designs a collaborative computing mechanism between nodes to solve the problem of remote sensing image classification in the case of non-independent and identically distributed data and a distributed storage environment.
[0048] The method proposed by the present invention is the first to introduce the alternating direction method of multipliers into the field of remote sensing image classification. By equivalently transforming the original distributed image classification problem, the optimization of the global model is automatically decomposed into the solution of multiple local sub-problems, and the aggregation of local models is achieved by dynamically updating the multiplier variables. Different from the traditional heuristic use of the gradient descent algorithm to minimize the local model, this method uses the framework of the alternating direction method of multipliers, and simultaneously considers the optimization of the original problem and the dual problem - the solution of the original problem realizes the distributed decomposition of the global model, while the solution of the dual problem completes the global aggregation of local model information. Therefore, the method of the present invention can not only improve the convergence speed of the model, but also reduce the assumption requirements of the convergence conditions, thereby enhancing the applicability and robustness of the method. Description of the Drawings
[0049] Figure 1 It is a flowchart of a distributed remote sensing image classification method based on the alternating direction method of multipliers. Specific Embodiments
[0050] The following describes the specific embodiments of the present invention to facilitate those skilled in the art of the present technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.
[0051] Embodiment 1
[0052] The distributed remote sensing image classification method based on the alternating direction method of multipliers in this embodiment solves the distributed remote sensing image classification problem in the form of primal-dual based on the alternating direction method of multipliers. The solution of the original problem realizes the distributed decomposition of the global model, and the solution of the dual problem realizes the global aggregation of local models. Refer to Figure 1 , and it specifically includes the following contents:
[0053] S1. Collect remote sensing image data and construct a remote sensing image data set;
[0054] S2. Distributively store the remote sensing image data set in each node;
[0055] S3. Deploy information to the central server and each node's local model;
[0056] In this embodiment, the i-th (1 ≤ i ≤ K) node stores local training data where the local training data is sourced from a remote sensing image dataset; M i is the number of remote sensing images in D i , is the l-th remote sensing image in D i , and is the corresponding label;
[0057] Specifically, the loss function for the deployment of each node's local model is:
[0058]
[0059] In the formula, F(w i ; D i ) represents the loss function of the machine learning model deployed by the i-th node; w i is the local model variable of the i-th node; D i is the local training data stored by the i-th node; is the l-th remote sensing image in D i , is the corresponding label; L CE is the cross-entropy loss function; f represents the machine learning model for remote sensing image classification; := represents definition;
[0060] In this embodiment, to prevent model overfitting, the loss function of each node's local model is regularized as:
[0061]
[0062] In the formula, F R (w; D i ) is the loss function of the machine learning model deployed by the i-th node after regularization; μ is the regularization hyperparameter used to balance the loss term and the regularization term.
[0063] The objective function Q i (w i ) for the deployment of each node's local model in this embodiment is:
[0064]
[0065] In the formula, D = {D1, D2, …, D K} represents the data set stored by each node; represents the gradient of the function F; w is the global model variable of the central server; λ i is the multiplier variable of the i-th node; β i, τ i denotes the model parameter, β i , τ i > 0, i = 1, 2, … K, where K is the number of nodes.
[0066] In addition, each node initializes its local model multiplier variable and the parameter and
[0067] The objective function Q(w) deployed by the central server in this embodiment is:
[0068]
[0069] In the formula, w is the global model variable of the central server.
[0070] S4. The central server aggregates the parameter information of the local models on each node, then constructs a global model, and distributes this global model to each node in parallel;
[0071] Specifically, in the t-th round of loop in this embodiment, the central server aggregates the local models of each node in the (t - 1)-th round of loop and the multiplier variables and other information, and then constructs a global model;
[0072] Among them, the global model is expressed as:
[0073]
[0074] In the formula, w t represents the global model of the central server; represents the multiplier variable of the i-th node in the (t - 1)-th round of loop; represents w i represents the local model of the i-th node in the (t - 1)-th round of loop.
[0075] S5. Each node trains and updates its local model according to the received global model and the remote sensing image dataset, and sends the parameter information of the trained and updated local model to the central server, and executes S4 until the local models of the current nodes meet the convergence condition, then execute S6;
[0076] During the model training and updating process, in this embodiment, after node i receives the global model w t from the central server, it uses the stored local data (from the remote sensing image dataset) to train and update the model;
[0077] Specifically, the trained and updated local model is expressed as:
[0078]
[0079] In the formula, is the updated local model after training.
[0080] As a preference of this embodiment, the convergence condition of this embodiment is that the change amount of the local models of each node in the previous and next rounds is less than a threshold value, and this convergence condition is specifically expressed as:
[0081]
[0082] In the formula, ε represents the threshold value.
[0083] S6. Classify the newly collected remote sensing image data by using the global model output by the central server.
[0084] When the convergence condition is satisfied in the t-th round of loop, the global model output by the central server is:
[0085]
[0086] In the formula, w * is the global model finally output by the central server; represents the multiplier variable of the i-th node in the t-th round of loop.
[0087] Although the specific implementation manners of the invention are described in detail with reference to the drawings, it should not be construed as a limitation on the protection scope of this patent. Within the scope described in the claims, various modifications and deformations that can be made by those skilled in the art without creative efforts still fall within the protection scope of this patent.
Claims
1. A distributed remote sensing image classification method based on the alternating direction method of multipliers, characterized in that, It includes the following steps: S1. Collect remote sensing image data and construct a remote sensing image data set; S2. Distributively store the remote sensing image data set in each node; S3. Deploy information for the central server and local models of each node; S4. The central server aggregates the parameter information of the local models on each node, and then constructs a global model, and concurrently sends the global model to each node; S5. Each node trains and updates the local model according to the received global model and the remote sensing image data set, and sends the parameter information of the trained and updated local model to the central server, and executes S4. When the local models of each node currently meet the convergence condition, execute S6; S6. Classify the newly collected remote sensing image data by using the global model output by the central server.
2. The distributed remote sensing image classification method based on the alternating direction multiplier method according to claim 1, wherein In S3, the loss function for the deployment of the local model of each node is: where, F(w i ; D i ) represents the loss function of the machine learning model deployed at the i-th node; w i is the local model variable of the i-th node; D i is the local training data stored at the i-th node; is the l-th remote sensing image in D i ; is the corresponding label; L CE is the cross-entropy loss function; f represents the machine learning model for remote sensing image classification; := means definition; The regularization expression of the loss function for the local model of each node: Where, F R (w i ; D i ) is the loss function of the machine learning model deployed at the i-th node after regularization processing; μ is the regularization hyperparameter.
3. The distributed remote sensing image classification method based on the alternating direction multiplier method according to claim 2, wherein In S3, the objective function Q for the deployment of each node's local model i (w i ) is as follows: where D = {D1, D2, …, D K}, representing the data set stored by each node; represents the gradient of the function F; w is the global model variable of the central server; λ i is the multiplier variable of the i-th node; β i , τ i represent model parameters, β i , τ i > 0, i = 1, 2, … K, where K is the number of nodes.
4. The distributed remote sensing image classification method based on the alternating direction multiplier method according to claim 3, wherein In S3, the objective function deployed by the central server is: In the formula, Q(w) is the objective function deployed by the central server.
5. The distributed remote sensing image classification method based on the alternating direction multiplier method according to claim 3, wherein In S4, the central server aggregates the parameter information of the local models on each node, and then constructs a global model; Among them, the global model is expressed as: where, w t represents the global model of the central server; represents the multiplier variable of the i-th node in the (t - 1)-th round of iteration; represents w i represents the local model of the i-th node in the (t - 1)-th round of iteration.
6. The distributed remote sensing image classification method based on the alternating direction multiplier method according to claim 5, characterized in that In S5, each node trains and updates the local model according to the received global model and the remote sensing image data set; Among them, the trained and updated local model is expressed as: In the formula, is the updated local model after training.
7. The distributed remote sensing image classification method based on the alternating direction multiplier method according to claim 6, wherein In S5, the convergence condition is that the change amount of the local models of each node in two consecutive rounds is less than the threshold, and this convergence condition is specifically expressed as: In the formula, ε represents the threshold.
8. The distributed remote sensing image classification method based on the alternating direction multiplier method according to claim 6, characterized in that In S6, the global model output by the central server is: where, w * is the global model finally output by the central server; represents the multiplier variable of the i-th node in the t-th round of iteration.