Ocean surface cloud detection method based on quantum convolutional neural network
By employing quantum convolutional neural networks, leveraging quantum parallelism and feature extraction capabilities, the problem of insufficient accuracy in detecting complex backgrounds and thin clouds in existing cloud detection methods is solved, achieving more efficient cloud detection results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF SCI & TECH OF CHINA
- Filing Date
- 2025-06-06
- Publication Date
- 2026-07-07
AI Technical Summary
Existing cloud detection methods are not accurate enough when dealing with complex backgrounds and thin clouds. Traditional methods have poor robustness, machine learning relies on manual features and has limited generalization ability, and deep learning requires large computing resources and is prone to overfitting, making it difficult to achieve efficient detection in large-scale marine remote sensing data.
We employ a quantum convolutional neural network (QCNN) approach, leveraging quantum parallelism and feature extraction capabilities. By utilizing qubit rotation gates and CZ gates to simulate classical convolution, we perform quantum pooling and measurement, dynamically adjusting feature response weights to reduce the number of parameters and improve generalization ability.
It performs better on unknown spectral data, has a small generalization error, high computational efficiency, adapts to data with different spatial and temporal distributions, reduces training parameters, and achieves more efficient cloud detection.
Smart Images
Figure CN120747728B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of cloud detection technology, specifically a method for detecting ocean surface clouds based on quantum convolutional neural networks. Background Technology
[0002] Currently, cloud detection is a crucial step in remote sensing data processing, widely used in fields such as weather forecasting, climate research, and environmental monitoring. Existing cloud detection methods mainly include traditional thresholding methods, machine learning methods, and deep learning methods, as detailed below:
[0003] A: Threshold-based cloud detection methods mainly distinguish between cloud and non-cloud areas by setting single-band or multi-band brightness temperature (BT) or reflectance (R) thresholds. For example, the commonly used MODIS cloud detection algorithm uses reflectance thresholds in the near-infrared and mid-infrared bands to identify clouds. These methods are simple to calculate, but they are sensitive to different types of clouds, changes in ground background and imaging conditions, and have poor robustness.
[0004] B: Regarding cloud detection methods based on traditional machine learning, in recent years, methods such as support vector machine (SVM), random forest (RF), and K-means clustering have been used for cloud detection. These methods classify clouds and non-clouds by constructing statistical feature models. However, since these methods rely on manual feature extraction, the selection of features has a significant impact on classification accuracy, and it is difficult to make full use of information from multispectral or high-dimensional remote sensing data.
[0005] C: Regarding deep learning-based cloud detection methods, in recent years, deep learning methods such as convolutional neural networks (CNNs) have made some progress in cloud detection tasks. For example, semantic segmentation networks such as U-Net and FCN are used to perform pixel-level classification of remote sensing images, which improves the accuracy of cloud detection. These methods can automatically extract multi-level features, but they require a large amount of data and have high computational resource requirements.
[0006] Considering the application of quantum computing in image processing, this is a relatively unexplored area. Currently, quantum computing technology is mainly applied to basic image classification tasks, such as simple pattern recognition based on quantum circuits. The technology is not mature enough: existing quantum neural networks are mostly still in the theoretical or simulation stage, lacking practical integration with remote sensing image analysis. The task adaptability is insufficient: quantum computing architectures have not been specifically designed for complex scenarios such as multispectral and multi-temporal data fusion for ocean cloud detection.
[0007] Based on the above, the limitations of existing cloud detection methods are as follows: traditional methods are insufficient for detecting complex backgrounds and thin clouds, and are easily affected by imaging conditions and threshold settings; machine learning methods rely on artificial features and have limited generalization ability; although deep learning methods have improved accuracy, they require large computational resources and may suffer from overfitting when data is insufficient; traditional methods and classic CNNs have difficulty balancing detection accuracy and real-time performance, and are particularly inefficient when processing large-scale marine remote sensing data.
[0008] Therefore, in view of the limitations of existing methods, this invention proposes a method for ocean cloud detection based on quantum convolutional neural network QCNN. Summary of the Invention
[0009] The present invention aims to solve at least one of the technical problems existing in the prior art, so as to improve the accuracy and computational efficiency of cloud detection by utilizing the parallelism and feature extraction capabilities of quantum computing; to this end, the present invention proposes a method for ocean surface cloud detection based on quantum convolutional neural networks.
[0010] A method for detecting ocean surface clouds based on quantum convolutional neural networks includes the following steps:
[0011] In the quantum coding layer, the reflectivity values of each channel are processed by normalization mapping and then used as the angle input of the rotation gate into the coding layer; two features are encoded for each quantum bit through Rx and Ry gate operations;
[0012] In quantum convolutional layers, adjacent qubits (such as q0 and q1) are spatially correlated through CZ gates, simulating the local connectivity characteristics of classical convolution.
[0013] In the quantum pooling layer, pooling operations are performed on the qubits;
[0014] The expected measurement values of the two qubits are obtained by measuring the state of the remaining two qubits after multiple convolution and pooling operations.
[0015] Furthermore, the parameter values of the revolving door are in the range of (-π, π). The reflectivity values of each channel are normalized to the interval (-π, π) through min-max and then used as the angle input of the revolving door into the encoding layer.
[0016] Furthermore, the relationship between the reflectivity values of each channel and the rotation angle is as follows:
[0017] θ=2π·(xx min ) / (x max -x min )-π.
[0018] Furthermore, methods for encoding two features using Rx and Ry gate operations include:
[0019] Eight features are encoded onto four qubits, q0, q1, q2, and q3, respectively. Rx and Ry gates are selected to encode the spatial distribution and spectral intensity differences of the band features, respectively.
[0020] Furthermore, the quantum convolutional layer also includes:
[0021] The parameterized rotation gate is optimized through gradient descent and dynamically adjusts the feature response weights. The parameters are equivalent to the learnable parameters of the classic convolutional kernel.
[0022] Furthermore, during gradient descent optimization, the parameters of the qubit gates outside the coding layer are optimized, the loss function is chosen as cross-entropy loss, and the optimizer is Adam; the model is then trained repeatedly and saved.
[0023] Furthermore, the saved model is used to average the results when using different test sets to reduce error.
[0024] Furthermore, the predictive ability of the model is evaluated based on the accuracy, precision, recall, and F1-score of the predictions on the test set.
[0025] Furthermore, when performing pooling operations on qubits, measurement operations are performed on only two qubits.
[0026] Furthermore, it also includes:
[0027] Compare the expected values of the two obtained qubits.
[0028] The classification result with a larger expected value is marked as 0, indicating no cloud cover;
[0029] Conversely, the classification result is marked as 1, indicating that there are clouds.
[0030] Compared with the prior art, the beneficial effects of the present invention are:
[0031] Compared to CNNs with similar structures, this invention exhibits stronger generalization ability and smaller generalization error during model training in cloud detection tasks with unknown spectral data. For data with different spatial and temporal distributions, QCNN performs better than CNN. Compared to CNNs with similar structures, QCNN requires fewer training parameters. By training a quantum convolutional neural network for cloud detection tasks, it has better generalization ability and fewer parameters. Driven by quantum parallelism, a single quantum convolution operation can be applied to the entire spectral dimension simultaneously, avoiding the redundant steps of channel-by-channel serial computation in classical CNNs. Attached Figure Description
[0032] Figure 1 The overall structure of a quantum convolutional neural network used for binary classification of cloud detection;
[0033] Figure 2 Quantum circuits used for the data encoding layer;
[0034] Figure 3 For parameterized quantum circuits used in convolutional layers;
[0035] Figure 4 For pooling layer quantum circuits;
[0036] Figure 5 For the training and testing process of QCNN and CNN models;
[0037] Figure 6 The loss and generalization error variations for the training and validation sets of the two models are the average values from ten repeated experiments. Detailed Implementation
[0038] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0039] Example 1:
[0040] This application provides a method for ocean cloud detection based on quantum convolutional neural networks. The quantum convolutional neural network, after being trained, helps to address the high computational cost of traditional CNNs and exhibits better generalization ability. Quantum computing can demonstrate exponential speedup in certain types of operations, theoretically improving the computational speed and processing power of convolutional neural networks, especially when processing high-dimensional data. Specifically:
[0041] This invention uses MODIS spectral data to train a quantum convolutional neural network to achieve the goal of cloud detection. The specific bands used in the data are shown in Table 1, which includes the albedo of three bands, the brightness temperature values of three bands, and the brightness temperature difference values of two bands.
[0042] Table 1 Selected bands
[0043]
[0044]
[0045] A schematic diagram of the quantum convolutional neural network used in this invention is shown below. Figure 1 It includes an encoding layer E that encodes classical data into quantum states, a convolutional layer C, and a pooling layer P.
[0046] A method for detecting ocean surface clouds based on quantum convolutional neural networks is described below:
[0047] First, the revolving door parameter values range from (-π, π). The reflectivity values of each channel are normalized to the interval (-π, π) using a min-max normalization process and then used as the input angle θ of the revolving door. Figure 2 In the coding layer shown, the correspondence between it and the rotation angle is as follows:
[0048] θ=2π·(xx min ) / (x max -x min -π;
[0049] Figure 2 For the quantum circuit used in the data encoding layer, Ri(θ) is a single-qubit rotation gate, where i can be x or y, representing rotating the original state around the x or y axis on the Bloch sphere by an angle θ.
[0050] Two features are encoded for each qubit through Rx and Ry operations, and the eight features are encoded onto four qubits (q0, q1, q2, q3). Rx and Ry gates are selected to encode the spatial distribution and spectral intensity differences of the band features respectively. The loss of encoded information by a single rotation axis is avoided by using a compound rotation operation, such as the equatorial blind zone caused by using only the Rz gate.
[0051] Figure 3 For the parameterized quantum circuit used in the convolutional layer, U3(θ,φ,λ) is an arbitrary single-qubit gate. After passing through... Figure 3 After the convolutional layer C of the structure shown, in the quantum convolutional layer, adjacent qubits are spatially correlated through CZ gates, simulating the local connectivity characteristics of classical convolution.
[0052] In the quantum pooling layer, pooling operations are performed on the qubits; by measuring the states of the two remaining qubits after multiple convolution and pooling operations, the expected measurement values of the two qubits are obtained.
[0053] As an embodiment of the present invention, preferably, the selection of adjacent qubits is as follows: Figure 3 The spatial features of q0 and q1 shown are correlated through the CZ gate.
[0054] As an embodiment of the present invention, preferably, the parameterized rotating gate is optimized by gradient descent and dynamically adjusts the feature response weights, and its parameters can be equivalent to the learnable parameters of the classic convolutional kernel.
[0055] As an embodiment of the present invention, preferably, the quantum bits are subjected to the following... Figure 4 The pooling operation P shown consists of two controlled rotating gates, with the aim of reducing the number of bits that need to be measured. This applies only to operations such as... Figure 1 The qubits numbered q1 and q2 shown are subjected to measurement operations to reduce measurement overhead. By measuring the states of the two remaining qubits q1 and q2 after multiple convolution and pooling operations, their expected measurement values can be obtained.
[0056] As an embodiment of the present invention, preferably, if the expected value of q1 is larger, the classification result is 0 (no cloud); otherwise, the classification result is 1 (cloudy).
[0057] The final model can be obtained by optimizing the parameters of the qubit gates in the Variational Quantum Circuit (VQC) excluding the coding layer.
[0058] As an embodiment of the present invention, preferably, the loss function is cross-entropy loss, and the optimizer is Adam. The model is trained ten times and saved for averaging on different test sets to reduce error in the next step.
[0059] Example 2:
[0060] The quantum convolutional neural network QCNN model provided in Example 1 was implemented on the isQ simulator and the Tianyan-176-II quantum computer of Guodun respectively. At the same time, a classical convolutional neural network model with a similar structure, namely one convolutional layer and one pooling layer, was constructed and the same learning rate, loss function and optimizer were used to compare their prediction capabilities.
[0061] Furthermore, the model's predictive ability is evaluated based on its accuracy, precision, recall, and F1 score on the test set, using the following formulas:
[0062]
[0063] In the formula, TP (True Positives) are true positive examples, which are predicted to be positive and are actually positive examples; FP (False Positives) are false positive examples, which are predicted to be positive but are actually negative examples; FN (False Negatives) are false negative examples, which are predicted to be negative but are actually positive examples; and TN (True Negatives) are true negative examples, which are predicted to be negative and are actually negative examples.
[0064] The training and testing data used in this invention came from the Moderate Resolution Imaging Spectroradiometer (MODIS) on NASA Terra and Aqua satellites.
[0065] As an embodiment of the present invention, MODIS L1A products (MOD03, MYD03), MODIS L1B products (MOD02, MYD02), and MODIS L2 products (MOD35, MYD35) were preferably used. MODIS data, with its wide band coverage and high spatiotemporal resolution, is widely used in Earth observation and environmental monitoring. Ocean surface data with a resolution of 1 km was extracted and used for training. Validation was performed using data from different times and regions.
[0066] The training and validation sets consist of 150 samples of MOD data from 02:10 UTC on April 21, 2023, split in a 7:3 ratio. Two different time-based datasets from the East China Sea and the South China Sea were used for testing to explore the differences in its spatial and temporal generalization capabilities.
[0067] Figure 5 For the training and testing process of QCNN and CNN models, see [link to training process description]. The accuracy and loss changes during training are shown below. Figure 5 ,exist Figure 6 In the diagram, we plotted the loss difference between the training and validation sets, i.e., the generalization error, as a function of the number of iterations. The generalization error refers to the difference in loss between the training dataset and the test dataset. Figure 6 The loss and generalization error variations for the training and validation sets of the two models are the average values from ten repeated experiments.
[0068] The results from test sets at different times and locations are shown in Table 2;
[0069] Table 2 shows the accuracy, precision, recall, and F1 score for different datasets. Cells are formatted as "QCNN (simulator) / QCNN (physical machine) / CNN".
[0070]
[0071]
[0072] Based on Embodiments 1 and 2, this invention, tested using an ISQ simulator and a quantum computer, demonstrates stronger generalization ability compared to CNNs with similar structures in unknown spectral data cloud detection tasks. The generalization error during model training is also smaller. (See...) Figure 6 The accuracy and F1 score of four sets of data with different spatial and temporal distributions were analyzed. The specific test results are shown in Table 2. QCNN has better performance than CNN. Compared with CNN with similar structure, QCNN has fewer training parameters. For the two networks used in the experiment of this invention, QCNN has 31 parameters and CNN has about 160 parameters.
[0073] As an embodiment of the present invention, preferably, if the network structure is changed, for example, by using different data encoding schemes, different structured convolutional layers, and pooling layers, it should still have better performance than traditional CNNs with similar structures. This should be understood as a modification or equivalent substitution of the technical method of the present invention without departing from the spirit and scope of the technical method of the present invention.
[0074] As an embodiment of the present invention, it is preferred that the loss function and optimizer be changed, which can be understood as modifying or equivalently replacing the technical method of the present invention without departing from the spirit and scope of the technical method of the present invention.
[0075] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.
Claims
1. A method for detecting ocean surface clouds based on quantum convolutional neural networks, characterized in that, The steps include the following: In the quantum coding layer, the reflectivity values of each channel are processed by normalization mapping and then used as the angle input of the rotation gate into the coding layer; two features are encoded for each quantum bit through Rx and Ry gate operations; In quantum convolutional layers, adjacent qubits are spatially correlated through CZ gates, simulating the local connectivity characteristics of classical convolution. In the quantum pooling layer, pooling operations are performed on the qubits; The expected measurement values of the two qubits are obtained by measuring the state of the two remaining qubits after multiple convolution and pooling operations. The parameter value range of the revolving door is: The reflectivity values of each channel are mapped to the interval using min-max normalization. This angle is then used as the input to the encoding layer for the revolving door. The relationship between the reflectivity values of each channel and the rotation angle is as follows: θ=2π (xx) min ) / (x max -x min )-π, where X represents the reflectivity, brightness temperature or brightness temperature difference of the input channel, and Xmax and Xmin represent its maximum and minimum values, respectively; Methods that encode two features using Rx and Ry gate operations include: Eight features are encoded onto four qubits. The eight features include the albedo of three bands, the brightness temperature of three bands, and the brightness temperature difference of two bands. The four qubits are q0, q1, q2, and q3, respectively. Rx and Ry gates are selected to encode the spatial distribution and spectral intensity difference of the band features, respectively. When performing a pooling operation on a qubit, the pooling operation consists of two controlled rotating gates, and only the two qubits are measured. Compare the expected values of the two obtained qubits. The classification result with a larger expected value is marked as 0, indicating no cloud cover; Conversely, the classification result is marked as 1, indicating that there are clouds.
2. The ocean cloud detection method based on quantum convolutional neural networks according to claim 1, characterized in that, Quantum convolutional layers also include: The parameterized rotation gate is optimized through gradient descent and dynamically adjusts the feature response weights. The parameters are equivalent to the learnable parameters of the classic convolutional kernel.
3. The ocean cloud detection method based on quantum convolutional neural networks according to claim 2, characterized in that, When optimizing using gradient descent, the parameters of the qubit gates outside the coding layer are optimized, the loss function is cross-entropy loss, and the optimizer is Adam; the model is trained repeatedly and saved.
4. The ocean cloud detection method based on quantum convolutional neural networks according to claim 3, characterized in that, The saved model is used to average the results when using different test sets to reduce error.
5. The ocean cloud detection method based on quantum convolutional neural networks according to claim 4, characterized in that, The predictive ability of a model is evaluated based on its accuracy, precision, recall, and F1-score on the test set.