Remote Sensing Scene Classification Method Based on Evolutionary Neural Network Architecture Search Algorithm
By using an evolutionary neural network architecture search algorithm, combined with particle swarm optimization and random forest, a convolutional neural network suitable for remote sensing scene classification is automatically designed, solving the problem of limited performance in remote sensing image classification and achieving automated architecture optimization with high accuracy and low parameters.
Patent Information
- Application Number
- CN202311597056.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-28
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2043-11-28
AI Technical Summary
Existing deep learning methods fail to fully utilize the specific spatial patterns of remote sensing images in remote sensing scene classification, resulting in limited classification performance. Furthermore, the manually designed network structures are complex and have poor interpretability.
An evolutionary neural network architecture search algorithm is adopted, which combines particle swarm optimization and random forest to automatically search for and design convolutional neural network models suitable for remote sensing scene classification. Individuals are optimized through crossover, mutation and adaptive strategies to quickly find the optimal neural network architecture.
We have achieved a high-precision, low-parameter automated neural network architecture design for remote sensing scene classification, which improves classification performance and reduces reliance on domain knowledge and the time required for manual tuning.
Smart Images

Figure CN117456275B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a remote sensing scene classification method, and more particularly to a remote sensing scene classification method based on an evolutionary neural network architecture search algorithm. Background Technology
[0002] With the rapid development of computer technology, remote sensing big data has become increasingly popular. Utilizing big data thinking and methods, remote sensing plays a vital role in many fields. For example, during disasters such as earthquakes and floods, remote sensing satellites can provide real-time images, which is crucial for disaster relief and rescue efforts. Currently, internet service platforms built upon remote sensing satellite systems have been widely applied in various scientific and technological fields. High-resolution remote sensing imagery contains a wealth of ground feature and texture information; how to correctly identify and classify this information is an urgent research topic.
[0003] Remote sensing scene classification involves correctly labeling a given remote sensing image using predefined semantic categories. With advancements in remote sensing imaging technology, the spatial resolution of remote sensing images has become increasingly higher, and categories increasingly influence each other. The same scene category can exhibit different color and texture features, and different objects also differ in their information. Similar individuals may exist within different scene categories; that is, two objects may share similar spectral characteristics. Furthermore, the imbalance in the distribution of various land cover types is prevalent in high-resolution remote sensing images, posing challenges to existing classification techniques.
[0004] Deep learning algorithms can learn hierarchical features at different levels of abstraction, achieving high accuracy in complex visual tasks. Deep convolutional neural networks (CNNs) are considered the most widely used deep learning algorithms and have been applied to scene classification tasks with excellent results. However, deep learning methods often require strong domain knowledge to design efficient network structures, and different network structures often lead to varying classification accuracies. For example, models with different structures, such as Googlenet, Mobilenet, and ResNet, can achieve different levels of accuracy in remote sensing image classification. Therefore, domain knowledge is crucial for designing efficient neural networks, and manually designed networks often have a large number of parameters, resulting in overly complex models with poor interpretability.
[0005] To address this limitation, in recent years, researchers have attempted to design automated convolutional neural networks (CNNs). This involves using optimization or search methods to automatically fine-tune the neural network structure, obtaining the optimal structure for the current task and achieving good performance. Several different neural network architecture methods have been proposed, with evolutionary computation algorithms being a mainstream approach due to their excellent global search performance. Evolutionary computation-based methods utilize a population search framework, allowing for the trial of many different neural network architectures during the search process, ultimately leading to the highest-performing neural network.
[0006] While numerous convolutional neural networks (CNNs) and network architecture search methods have emerged and achieved promising results, most are applied to object classification image data and are less suitable for remote sensing scene classification data. Significant and unique differences exist between remote sensing / aerial imagery and natural imagery. For example, remote sensing scene imagery is typically composed of local patches extracted from large satellite images of varying resolutions. These patches are composed of highly complex geometries, with significant differences in viewpoint, scale, and shape of ground features. Existing CNNs are almost entirely designed based on natural images and have not been modified to address the specific spatial patterns of remote sensing images. This limits the potential of deep learning methods and impacts classification performance. Therefore, there is an urgent need to design an automatic search method for CNN models specifically for scene classification to further improve classification performance. Summary of the Invention
[0007] To address the aforementioned issues, this invention proposes a novel evolutionary neural network architecture search method. Without any pre-training or data augmentation, it automatically searches for convolutional neural network architectures, obtaining a convolutional neural network model for scene classification without human intervention.
[0008] The technical solution adopted in this invention is as follows:
[0009] The remote sensing scene classification method based on the evolutionary neural network architecture search algorithm includes the following steps:
[0010] S1, Initialize the population;
[0011] S2, evaluate the fitness of the population and record the historical best individual and the global best individual for each population;
[0012] S3 uses the particle swarm optimization algorithm to evolve the population;
[0013] S4: Determine if the algebra is greater than 5. If it is less than 5, go to S1; if it is greater than 5, go to S5.
[0014] S5 uses data from the previous 5 generations to build a random forest;
[0015] S6 uses crossover, mutation, and adaptive strategies to generate 20 to 60 individuals;
[0016] S7: Use random forest to select individuals, perform real-world evaluations, and update the random forest model;
[0017] S8, after the evolution is complete, select the best individuals to test on the test set.
[0018] Furthermore, the specific steps of S2 are as follows:
[0019] S21, Input the image that needs to be classified;
[0020] S22, the images are divided into an evolutionary training set, an evolutionary test set, and a test set, with the test set images used as the data for training the final optimal architecture;
[0021] S23, use an early stopping strategy to train the decoded neural network model;
[0022] S24 records the parameters of the architecture and the corresponding classification accuracy.
[0023] Furthermore, the specific steps of S23 are as follows:
[0024] S231, The neural network model is trained on an evolutionary training set;
[0025] S232, The neural network model is evaluated on an evolutionary test set;
[0026] S233: If the number of model iterations or the classification accuracy reaches a preset value, then training is terminated.
[0027] Furthermore, the specific steps of S5 are as follows:
[0028] S51, load the codes of the first 5 generations of individuals as data, and the classification accuracy is the label;
[0029] S52, construct a random forest by building decision trees;
[0030] S53, construct multiple decision trees, and train each decision tree;
[0031] S54 integrates multiple decision trees into a single random forest model;
[0032] S55: Use the average of multiple decision trees in the random forest model as the final prediction result.
[0033] Furthermore, the specific steps of S6 are as follows:
[0034] S61, the particle swarm evolution produces 20 individuals;
[0035] S62, 20 individuals are generated by cross-pollinating the parent population with the best historical individual;
[0036] S63, 20 new individuals were produced by mutation in the parent population;
[0037] S64, the parent population uses an adaptive particle swarm evolution strategy to produce 20 individuals.
[0038] Furthermore, the specific steps of S62 are as follows:
[0039] S621, obtain the minimum value 'a' of the encoding length of the current individual and the encoding length of the historical best individual;
[0040] S622, randomly generate a random number between 0 and a as the position to be swapped;
[0041] S623, swaps the configuration particles and connection particles of two individuals.
[0042] Furthermore, the specific steps of S63 are as follows:
[0043] S631: Get the encoding length b of the current individual;
[0044] S632: Generate a random number between 0 and b as the mutation location;
[0045] S633: Based on the configuration of particles and the connection of particles, a different number is generated at the position of mutation.
[0046] Furthermore, in S64, the adaptive particle swarm optimization algorithm is an improvement based on the standard particle swarm optimization algorithm; the standard particle swarm optimization algorithm guides the particles to the vicinity of the global optimum by continuously updating their positions and velocities; a typical update rate formula is as follows:
[0047] V i (t+1) =ω·V i (t) +c1·r1·(Pbest i -X i (t) )+c2·r2·(Gbest-X i (t) )
[0048] In the formula, V i (t+1) ω is the velocity of particle i at the next moment; ω is the inertia weight, usually a positive number less than 1, used to balance the global and local searches; c1 and c2 are acceleration coefficients, usually positive numbers, used to control the interaction between individuals and the group; r1 and r2 are randomly generated values in the range [0,1], introducing randomness; Pbesti It is the best historical position of particle i itself; X i (t) Gbest is the position of particle i at the current moment; Gbest is the historical best position of the entire particle swarm.
[0049] The position update formula is as follows:
[0050]
[0051] The adaptive particle swarm optimization algorithm modifies the update formula for ω based on the standard particle swarm optimization algorithm. The update formula for ω is as follows:
[0052]
[0053] Furthermore, the specific steps of S7 include:
[0054] S71, Read the codes of the individuals generated by S6 and add them to the dataset;
[0055] S72, input the data into the random forest model generated by S5;
[0056] S73 outputs the classification accuracy generated by the model;
[0057] S74, select the 20 with the best classification accuracy as the offspring population;
[0058] S75, accurately assess the offspring population and record the corresponding classification accuracy;
[0059] S76 adds the actual assessed offspring population and corresponding classification accuracy to the random forest dataset to update the random forest model.
[0060] Furthermore, the specific steps of S8 are as follows:
[0061] S81, Input the test set divided by S22;
[0062] S82 reads the parameters required for the test and the globally optimal individual;
[0063] S83, obtain the final classification accuracy.
[0064] The beneficial effects of this invention are:
[0065] 1. The introduction of crossover and mutation factors makes the offspring more superior;
[0066] 2. It has the advantages of simple implementation, fast speed, high classification accuracy and few adjustable parameters;
[0067] 3. An adaptive particle swarm optimization algorithm was introduced, which enables the search to take into account both global and local search capabilities.
[0068] 4. In remote sensing scene classification, the architecture is automatically designed, and the model parameters are relatively small. Attached Figure Description
[0069] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0070] The invention will now be further described with reference to the accompanying drawings.
[0071] like Figure 1 As shown, this invention is a remote sensing scene classification method based on an evolutionary neural network architecture search algorithm, comprising the following steps:
[0072] S1, Initialize the population;
[0073] S2, assess the fitness of the population and record the historical best individual and the global best individual for each population. The specific steps are as follows:
[0074] S21: Input the image that needs to be classified;
[0075] S22: Divide the images into an evolutionary training set, an evolutionary test set, and a test set. The test set images are used as the data for training the final optimal architecture.
[0076] S23: Train the decoded neural network model using an early stopping strategy. The specific steps of the strategy are as follows:
[0077] S231: The neural network model is trained on an evolutionary training set;
[0078] S232: The neural network model is evaluated on an evolutionary test set;
[0079] S233: If the number of model iterations or the classification accuracy reaches a preset value, then training is terminated;
[0080] S234: Record the parameters of the architecture and the corresponding classification accuracy;
[0081] S3 uses the particle swarm optimization algorithm to evolve the population;
[0082] S4: Determine if the algebra is greater than 5. If it is less than 5, go to S1; if it is greater than 5, go to S5.
[0083] S5 uses data from the first 5 generations to build a random forest. The specific steps are as follows:
[0084] S51: Load the codes of the first 5 generations of individuals as data, and the classification precision is the label;
[0085] S52: Construct a random forest from decision trees;
[0086] S53: Construct multiple decision trees, and train each decision tree;
[0087] S54: Integrate multiple decision trees into a single random forest model;
[0088] S55: Use the average of multiple decision trees in the random forest model as the final prediction result;
[0089] S6 generates 20 individuals using crossover, mutation, and adaptive strategies, with the following steps:
[0090] S61: Particle swarm evolution produces 20 individuals;
[0091] S62: Crossing the parent population with the historical best individual generates 20 individuals. The crossover factor introduced in S62 of this invention ensures that the offspring population interacts more with the historical best individual, producing elite individuals and resulting in superior performance. The crossover operation promotes information exchange and mixing among different individuals, helping to maintain population diversity, preventing the population from prematurely falling into local optima, and increasing global exploration of the problem space. Crossover can merge the characteristics of two parent individuals, generating offspring individuals with new characteristics, which may help generate solutions not present in the parent population, potentially providing better opportunities for problem optimization. Through crossover, offspring individuals can inherit some good characteristics and combinations of excellent solutions from their parents, thus approaching the global optimum more quickly and accelerating the algorithm's convergence speed. The specific steps are as follows:
[0092] S621: Obtain the minimum value 'a' between the current individual's encoding length and the historical best individual's encoding length;
[0093] S622: Randomly generate a random number between 0 and a as the position to be swapped;
[0094] S623: Swap the configuration particles and connection particles of two individuals.
[0095] S63: Parental population mutation produces 20 new individuals; the specific steps are as follows:
[0096] S631: Get the encoding length b of the current individual;
[0097] S632: Generate a random number between 0 and b as the mutation location;
[0098] S633: Based on the configuration of particles and the connection of particles, a different number is generated at the position of mutation.
[0099] S64: The parent population uses an adaptive particle swarm evolution strategy to produce 20 individuals. The standard Particle Swarm Optimization (PSO) algorithm is a heuristic optimization algorithm used to find optimal or near-optimal solutions to complex problems. The basic idea of PSO is to simulate the group behavior of a flock of birds. In PSO, the problem is defined as a multi-dimensional search space, and each potential solution is represented as a particle. These particles move in the search space, searching for the global optimum. Particles guide their movement through communication and cooperation. The core idea of the algorithm is to continuously update the position and velocity of the particles, guiding them to the vicinity of the global optimum. A typical update rate formula is as follows:
[0100] V i (t+1) =ω·V i (t) +c1·r1·(Pbest i -X i (t) )+c2·r2·(Gbest-X i (t) )
[0101] In the formula, V i (t+1) ω is the velocity of particle i at the next moment; ω is the inertia weight, usually a positive number less than 1, used to balance the global and local searches; c1 and c2 are acceleration coefficients, usually positive numbers, used to control the interaction between individuals and the group; r1 and r2 are randomly generated values in the range [0,1], introducing randomness; Pbest i It is the best historical position of particle i itself; X i (t) Gbest is the position of particle i at the current moment; Gbest is the historical best position of the entire particle swarm; the position update formula is as follows:
[0102]
[0103] In the Adaptive Particle Swarm Optimization (APO) algorithm, a larger ω at the beginning encourages greater randomness in the particles, prompting them to explore more extensively in search of new regions in the potential solution space. This helps overcome the trap of local optima. As the algorithm progresses, gradually decreasing ω slows down the particles, causing them to focus more on the vicinity of known, more promising solutions, thus accelerating the convergence of the global optimum. The update formula for ω is as follows:
[0104]
[0105] This adaptive strategy, which uses a large ω value initially and then gradually decreases the ω value, helps PSO find a good trade-off between global and local search, thus improving the algorithm's performance.
[0106] S7, use random forest to select individuals, perform real-world evaluation, and update the random forest model; the specific steps are as follows:
[0107] S71: Read some codes of the individuals generated by S6 and add them to the dataset;
[0108] S72: Input the data into the random forest model generated by S5;
[0109] S73: Outputs the classification accuracy generated by the model;
[0110] S74: Select the 20 species with the best classification accuracy as the offspring population;
[0111] S75: Accurately assess the offspring population and record the corresponding taxonomic accuracy;
[0112] S76: Add the actual assessed offspring population and corresponding classification accuracy to the random forest dataset to update the random forest model.
[0113] S8, after the evolution is complete, select the best individuals to test on the test set.
[0114] To further illustrate the superiority of the present invention in the remote sensing scene classification problem, Table 1 shows the classification accuracy of the method of the present invention compared with some typical artificially designed networks ResNet34, MobileNet, and GoogleNet on the typical UCM remote sensing scene classification dataset.
[0115] The examples present experimental results on a typical remote sensing scene classification dataset. The UCM dataset has 21 categories, with 100 images per category, each 256×256 pixels. The experiments were conducted on an RTX 6000 GPU with a population size of 20 and 10 generations. Without any pre-training, the manually designed network was run 10 times independently, and our method was run 5 times independently. Table 1 shows the maximum accuracy, average accuracy, number of model parameters, and runtime. Each block in the table displays the experimental results on the UCM dataset. The comparison shows that the proposed method exhibits superior classification performance in remote sensing scene classification. Specifically, the proposed method outperforms other methods in both maximum and average accuracy. These results also demonstrate that the proposed method can find a good architecture for remote sensing scene classification, proving its effectiveness in classification tasks.
[0116] In summary, the remote sensing scene classification method based on the evolutionary neural network architecture search algorithm proposed in this invention can effectively process remote sensing datasets and find a good architecture.
[0117] Table 1 Comparison of experimental results on the UCM dataset
[0118]
[0119] Neural network architecture search can reduce the need for domain expert knowledge and the time spent on manual network tuning. By automatically optimizing the hyperparameters of the neural network and trying more novel network architectures, it can find the optimal neural network architecture. Evolutionary algorithms have been widely proven to have strong global search capabilities, capable of finding globally optimal or near-optimal solutions. Combining the two allows for a faster and more accurate discovery of the best neural network architecture.
[0120] This invention introduces crossover factors, mutation factors, and adaptive factors to generate combinations that inherit desirable characteristics and excellent solutions from their parent generations, thus producing elite solutions. The inclusion of random forests, as a surrogate model, allows for more accurate and rapid searching for the best solution among the generated offspring. In summary, this invention provides a novel remote sensing scene classification method based on evolutionary neural network architecture search. This method can effectively process remote sensing images without pre-training, automatically finding the optimal neural network architecture, and exhibits significant advantages over other manually designed neural networks.
[0121] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A remote sensing scene classification method based on evolutionary neural network architecture search algorithm, characterized in that, Includes the following steps: S1, Initialize the population; S2, assess the fitness of the population and record the historical best individual and the global best individual for each population; the specific steps are as follows: S21, Input the image that needs to be classified; S22, the images are divided into an evolutionary training set, an evolutionary test set, and a test set, with the test set images used as the data for training the final optimal architecture; S23, The early stopping strategy is used to train the decoded neural network model; the specific steps are as follows: S231, The neural network model is trained on an evolutionary training set; S232, The neural network model is evaluated on an evolutionary test set; S233, If the number of model iterations or the classification accuracy reaches the preset value, then training is terminated; S24 records the architecture parameters and the corresponding classification accuracy; S3 uses the particle swarm optimization algorithm to evolve the population; S4: Determine if the algebra is greater than 5. If it is less than 5, go to S1; if it is greater than 5, go to S5. S5, construct a random forest prediction model using data from the first 5 generations; the specific steps are as follows: S51, load the codes of the first 5 generations of individuals as data, and the classification accuracy is the label; S52, construct a random forest by building decision trees; S53, construct multiple decision trees, and train each decision tree; S54 integrates multiple decision trees into a single random forest model; S55, the average of multiple decision trees in the random forest model is used as the final prediction result; S6 uses crossover, mutation, and adaptive strategies to generate 20-60 individuals; S7, use random forest to select individuals, perform real-world evaluation, and update the random forest model; the specific steps are as follows: S71, Read the codes of the individuals generated by S6 and add them to the dataset; S72, input the data into the random forest model generated by S5; S73 outputs the classification accuracy generated by the model; S74, select the 20 with the best classification accuracy as the offspring population; S75, accurately assess the offspring population and record the corresponding classification accuracy; S76, add the actual assessed offspring population and corresponding classification accuracy to the random forest dataset to update the random forest model; S8, after the evolution is complete, select the best individuals to test on the test set.
2. The remote sensing scene classification method based on the evolutionary neural network architecture search algorithm according to claim 1, characterized in that, The specific steps of S6 are as follows: S61, the particle swarm evolution produces 20 individuals; S62, 20 individuals are generated by cross-pollinating the parent population with the best historical individual; S63, 20 new individuals were produced by mutation in the parent population; S64 uses an adaptive particle swarm optimization algorithm to generate 20 individuals in the parent population.
3. The remote sensing scene classification method based on the evolutionary neural network architecture search algorithm according to claim 2, characterized in that, The specific steps of S62 are as follows: S621, obtain the minimum value 'a' of the current individual's encoding length and the encoding length of the historical best individual; S622, randomly generate a random number between 0 and a as the position to be swapped; S623, swaps the configuration particles and connection particles of two individuals.
4. The remote sensing scene classification method based on the evolutionary neural network architecture search algorithm according to claim 2, characterized in that, The specific steps of S63 are as follows: S631, obtain the encoding length b of the current individual; S632, randomly generate a random number between 0 and b, as the mutation position; S633, based on the configuration of particles and the connection of particles, produces a different number at the position of mutation.
5. The remote sensing scene classification method based on the evolutionary neural network architecture search algorithm according to claim 2, wherein in S64, the following is characterized: The adaptive particle swarm optimization algorithm is an improvement on the standard particle swarm optimization algorithm. The standard particle swarm optimization algorithm guides particles to the vicinity of the global optimum by continuously updating their positions and velocities. A typical update rate formula is as follows: In the formula, It is the velocity of particle i at the next moment; It is the inertia weight, usually a positive number less than 1, used to balance global search and local search; and It is the acceleration factor, usually a positive number, used to control the interaction between individuals and groups; and It is a value randomly generated within the range [0,1], introducing randomness; It is the best historical position of particle i itself; It is the position of particle i at the current moment; It is the best position in the history of the entire particle swarm. The position update formula is as follows: The adaptive particle swarm optimization algorithm is based on the standard particle swarm optimization algorithm, but with modifications. The update formula, its The update formula is as follows: 。 6. The remote sensing scene classification method based on the evolutionary neural network architecture search algorithm according to claim 1, characterized in that, The specific steps of S8 are as follows: S81, Input the test set divided by S22; S82 reads the parameters required for the test and the globally optimal individual; S83, obtain the final classification accuracy.
Citation Information
Patent Citations
High-resolution remote sensing image classification method and system based on deep learning
CN114694038A
Crop classification method and device based on improved random forest algorithm, and computer equipment
CN114724026A
