Ocean surface cloud detection method based on quantum convolutional neural network

Abstract

The invention discloses an ocean surface cloud detection method based on a quantum convolutional neural network, and relates to the technical field of cloud detection, and the method comprises the steps: enabling a reflectivity value of each channel to serve as an angle of a revolving door after normalization mapping processing, and inputting the angle into a coding layer; encoding two features for each quantum bit through Rx and Ry gate operations; spatial feature association of adjacent quantum bits is realized through a CZ gate, and local connection characteristics of classic convolution are simulated; performing pooling operation on the quantum bits; the states of two remaining quantum bits after convolution and pooling operation are measured for multiple times, measurement expectation values of the two quantum bits are obtained, cloud detection tasks are carried out by training the quantum convolutional neural network, and better generalization ability and less parameter quantity are achieved; under the driving of quantum parallelism, single quantum convolution operation can act on the whole spectrum dimension at the same time, and redundant steps of channel-by-channel serial calculation in a classical CNN are avoided.

Description

Technical Field

[0001] The present invention belongs to the field of cloud detection technology, and specifically is a method for detecting ocean clouds based on quantum convolutional neural networks. Background Art

[0002] Currently, cloud detection is a key step in remote sensing data processing and is widely used in fields such as meteorological forecasting, climate research, and environmental monitoring. Existing cloud detection methods mainly include traditional threshold methods, machine learning methods, and deep learning methods, as follows:

[0003] A: Threshold-based cloud detection methods primarily 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 for cloud discrimination. These methods are computationally simple but are sensitive to variations in cloud types, surface background, and imaging conditions, resulting in poor robustness.

[0004] B: Regarding traditional machine learning-based cloud detection methods, in recent years, methods such as support vector machines (SVMs), random forests (RFs), and K-means clustering have been used for cloud detection. These methods build statistical feature models to classify clouds and non-clouds. However, because these methods rely on manual feature extraction, feature selection significantly affects classification accuracy, and they also struggle to fully utilize the information from multispectral or high-dimensional remote sensing data.

[0005] C: Regarding deep learning-based cloud detection methods, deep learning methods such as convolutional neural networks (CNNs) have made some progress in cloud detection tasks in recent years. For example, the use of semantic segmentation networks such as U-Net and FCN for pixel-level classification of remote sensing images has improved the accuracy of cloud detection. These methods can automatically extract multi-level features, but they have high data requirements and large demands on computing resources.

[0006] However, considering the application of quantum computing in image processing, this is a relatively blank area. At present, quantum computing technology is mainly used in basic image classification tasks, such as simple pattern recognition based on quantum circuits; the technology maturity is low: existing quantum neural networks mostly remain in the theoretical or simulation stage, lacking practical integration with remote sensing image analysis; insufficient task adaptability: quantum computing architecture has not been specifically designed for complex scenarios such as multi-spectral and multi-temporal data fusion in ocean cloud detection.

[0007] Based on the above, the limitations of existing cloud detection methods are as follows: traditional methods are insufficient in 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 capabilities; although deep learning methods have improved accuracy, they require large computing resources and may suffer from overfitting problems 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 ocean remote sensing data.

[0008] Therefore, in response to the limitations of existing methods, the present invention proposes an ocean cloud detection method 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 utilize the parallelism and feature extraction capabilities of quantum computing to improve the accuracy and computational efficiency of cloud detection. 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 clouds based on a quantum convolutional neural network comprises the following steps:

[0011] In the quantum coding layer, the reflectivity values ​​of each channel are processed by normalization mapping and then input into the coding layer as the angle of the rotating gate; two features are encoded for each quantum bit through the Rx and Ry gate operations;

[0012] In the quantum convolution layer, adjacent quantum bits (such as q0 and q1) are spatially associated through CZ gates, simulating the local connection characteristics of classical convolution.

[0013] In the quantum pooling layer, the quantum bits are pooled;

[0014] By measuring the states of the two qubits remaining after the convolution and pooling operations multiple times, the expected measurement values ​​of the two qubits are obtained.

[0015] Furthermore, the parameter value range of the revolving door is (-π, π), and the reflectivity value of each channel is mapped to the interval (-π, π) through min-max normalization and then input into the coding layer as the angle of the revolving door.

[0016] Furthermore, the corresponding relationship between the reflectivity value of each channel and the rotation angle is:

[0017] θ=2π·(xx min ) / (x max -x min )-π.

[0018] Furthermore, the method of encoding two features through Rx and Ry gate operations includes:

[0019] The eight features are encoded onto four quantum bits, namely q0, q1, q2, and q3, and the Rx and Ry gates are selected to encode the spatial distribution and spectral intensity differences of the band features respectively.

[0020] Furthermore, in the quantum convolution layer, it also includes:

[0021] The parameterized revolving gate dynamically adjusts the feature response weights through gradient descent optimization, and its parameters are equivalent to the learnable parameters of the classic convolution kernel.

[0022] Furthermore, when optimizing through gradient descent, the parameters of the quantum bit gate outside the coding layer are optimized, the cross entropy loss is selected as the loss function, and the Adam optimizer is selected; repeat the training and save the model.

[0023] Furthermore, the saved model is used to average the results when using different test sets to reduce the error.

[0024] Furthermore, the prediction ability of the model is evaluated based on the accuracy, precision, recall and F1-score value of the test set predictions.

[0025] Furthermore, when performing a pooling operation on qubits, measurement operations are only performed on two qubits.

[0026] Furthermore, it also includes:

[0027] Compare the expected values ​​of the two qubits obtained;

[0028] The classification result with a larger expected value is marked as 0, indicating no cloud;

[0029] Otherwise, the classification result is marked as 1, indicating that there are clouds.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] Compared with CNNs with similar structures, the present invention has stronger generalization ability and smaller generalization error in model training in the task of cloud detection of unknown spectral data; QCNN has better performance than CNN for data with different spatial and temporal distributions; compared with CNNs with similar structures, QCNN has fewer training parameters, and by training quantum convolutional neural networks for cloud detection tasks, it has better generalization ability and fewer parameters; driven by quantum parallelism, a single quantum convolution operation can act on the entire spectral dimension simultaneously, avoiding the redundant steps of channel-by-channel serial calculations in classic CNNs. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The overall structure of the quantum convolutional neural network for cloud detection binary classification;

[0033] Figure 2 quantum circuits used for the data encoding layer;

[0034] Figure 3 Parameterized quantum circuits used in convolutional layers;

[0035] Figure 4 is the pooling layer quantum circuit;

[0036] Figure 5 Training and testing process for QCNN and CNN models;

[0037] Figure 6 The training set and validation set loss and generalization error changes of the two models are the average of ten repeated experiments. DETAILED DESCRIPTION

[0038] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts 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 a quantum convolutional neural network. The application of trained quantum convolutional neural networks to cloud detection helps address the high computational overhead of traditional CNNs and improves generalization capabilities. Quantum computing can demonstrate exponential acceleration in certain types of operations and, in theory, can improve the computational speed and processing power of convolutional neural networks, especially when processing high-dimensional data. Specifically:

[0041] The present invention uses MODIS spectral data to train the quantum convolutional neural network to achieve the goal of cloud detection. The specific bands used by the data are shown in Table 1, including the albedo of three bands, the brightness temperature values ​​of three bands, and two brightness temperature differences.

[0042] Table 1 Selected bands

[0043]

[0044]

[0045] The schematic diagram of the quantum convolutional neural network used in the present invention is shown in Figure 1 , which includes an encoding layer E (Encoding) that encodes classical data into quantum states, a convolution layer C (Convolution) and a pooling layer P (Pooling).

[0046] A method for detecting ocean clouds based on quantum convolutional neural networks is as follows:

[0047] First, the revolving door parameter value range is (-π, π), and the reflectivity value of each channel is mapped to the interval (-π, π) through min-max normalization and then input as the angle θ of the revolving door into Figure 2 In the coding layer shown, the corresponding relationship between it and the rotation angle is:

[0048] θ=2π·(xx min ) / (x max -x min )-π;

[0049] Figure 2 The quantum circuit used for the data encoding layer, Ri(θ) is a single-qubit rotation gate, i can be x or y, representing the rotation of the original state around the x or y axis on the Bloch sphere by an angle of θ.

[0050] Two features are encoded for each quantum bit through Rx and Ry operations, and 8 features are encoded on 4 quantum bits (q0, q1, q2, q3). Rx and Ry gates are selected to encode the spatial distribution and spectral intensity differences of the band characteristics respectively. The loss of encoded information of a single rotation axis is avoided through compound rotation operations, such as the equatorial plane information blind spot caused by using only the Rz gate.

[0051] Figure 3 is the parameterized quantum circuit used in the convolutional layer, and U3(θ,φ,λ) is an arbitrary single-qubit gate. Figure 3 After the convolution layer C of the structure shown, in the quantum convolution layer, adjacent quantum bits realize spatial feature association through CZ gates, simulating the local connection characteristics of classical convolution.

[0052] In the quantum pooling layer, the quantum bits are pooled; by measuring the states of the two quantum bits remaining after the convolution and pooling operations multiple times, the expected values ​​of the two quantum bits are obtained.

[0053] As an embodiment provided by the present invention, preferably, adjacent quantum bits are selected as follows: Figure 3 As shown, q0 and q1 are spatially associated through the CZ gate.

[0054] As an embodiment provided by the present invention, preferably, the parameterized revolving gate dynamically adjusts the feature response weights through gradient descent optimization, and its parameters are equivalent to the learnable parameters of the classic convolution kernel.

[0055] As an embodiment provided by the present invention, preferably, the quantum bit is processed as follows Figure 4 The pooling operation P shown in FIG, P is composed of two controlled rotating gates, the purpose of which is to reduce the number of bits that need to be measured. Figure 1 The quantum bits numbered q1 and q2 shown in the figure are measured to reduce the measurement overhead. By measuring the states of the two quantum bits q1 and q2 remaining after the convolution and pooling operations multiple times, their measurement expectation values ​​can be obtained.

[0056] As an embodiment provided by the present invention, preferably, if the expected value of q1 is greater, the classification result is 0 (no cloud); otherwise, the classification result is 1 (cloud).

[0057] The final model can be obtained by optimizing the parameters of the quantum bit gates except the coding layer in the variable quantum circuit (VQC).

[0058] As an embodiment provided by the present invention, preferably, the loss function is cross entropy loss, and the optimizer is Adam. The training is repeated ten times and the model is saved for averaging to reduce the error when using different test sets 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 Guodun's Tianyan-176-II quantum computer, respectively. At the same time, a classical convolutional neural network model with a similar structure, i.e., one convolution layer and one pooling layer, was constructed, and the same learning rate, loss function, and optimizer were used to compare their predictive capabilities.

[0061] In addition, the prediction ability of the model is evaluated based on the accuracy, precision, recall and F1-score of the test set prediction. The specific formula is as follows:

[0062]

[0063] Where TP (True Positives) is a true positive example, which is predicted as a positive example and is actually a positive example; FP (False Positives) is a false positive example, which is predicted as a positive example but is actually a negative example; FN (false Negatives) is a false negative example, which is predicted as a negative example but is actually a positive example; TN (True Negatives) is a true negative example, which is predicted as a negative example and is actually a negative example.

[0064] The training and testing data used in this paper are from the Moderate Resolution Imaging Spectroradiometer (MODIS) on NASA's Terra and Aqua satellites.

[0065] As an embodiment provided by the present invention, MODIS L1A products (MOD03, MYD03), MODIS L1B products (MOD02, MYD02), and MODIS L2 products (MOD35, MYD35) are preferably used. MODIS data, with its wide band coverage and high temporal and spatial resolution, is widely used for Earth observation and environmental monitoring. Data from an ocean area with a resolution of 1 km was extracted for training. Data from different time periods and regions were used for verification.

[0066] The training and validation sets consist of 150 samples of MOD data at 02:10 UTC on April 21, 2023, split 7:3. Testing was performed using two samples from the East China Sea and two from the South China Sea at different times to explore differences in generalization capabilities across space and time.

[0067] Figure 5 For the training and testing process of QCNN and CNN models, the accuracy and loss changes during the training process are shown in Figure 5 ,exist Figure 6 In the figure, we plotted the difference in loss between the training and validation sets, i.e., the generalization error, as the number of iterations changes. The generalization error refers to the difference in loss between the training dataset and the test dataset. Figure 6 The training set and validation set loss and generalization error changes of the two models are the average of ten repeated experiments.

[0068] The results from different time and location test sets are shown in Table 2;

[0069] Table 2 Accuracy, precision, recall, and F1 values ​​for different datasets. The cell format is "QCNN (Simulator) / QCNN (Physical Machine) / CNN"

[0070]

[0071]

[0072] Based on the first and second embodiments, the present invention has been tested by an IQ simulator and a quantum computer. Compared with CNNs with similar structures, it has a stronger generalization ability in the task of detecting unknown spectral data clouds. The generalization error in model training is smaller, see Figure 6 ; Four sets of data with different spatial and temporal distributions were tested, and their accuracy and F1 values ​​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, the QCNN parameter is 31 and the CNN parameter is about 160.

[0073] As an embodiment provided by the present invention, preferably, if the network structure is changed, for example, using different data encoding schemes, different structured convolutional layers, and pooling layers, it should also have better performance than the traditional CNN with similar structure, which is understood as a modification or equivalent replacement 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 provided by the present invention, preferably, changing the loss function and the optimizer can also 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 method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A method for detecting ocean clouds based on quantum convolutional neural networks, characterized in that: The steps include: In the quantum coding layer, the reflectivity values ​​of each channel are processed by normalization mapping and then input into the coding layer as the angle of the rotating gate; two features are encoded for each quantum bit through the Rx and Ry gate operations; In the quantum convolution layer, adjacent qubits realize spatial feature association through CZ gates, simulating the local connection characteristics of classical convolution; In the quantum pooling layer, the quantum bits are pooled; By measuring the states of the two qubits remaining after the convolution and pooling operations multiple times, the expected measurement values ​​of the two qubits are obtained.

2. The ocean cloud detection method based on quantum convolutional neural network according to claim 1 is characterized in that: The parameter value range of the revolving door is (-π, π). The reflectivity value of each channel is mapped to the interval (-π, π) through min-max normalization and then input into the coding layer as the angle of the revolving door.

3. The ocean cloud detection method based on quantum convolutional neural network according to claim 2 is characterized in that: The corresponding relationship between the reflectivity value of each channel and the rotation angle is: 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 their maximum and minimum values ​​respectively.

4. The ocean cloud detection method based on quantum convolutional neural network according to claim 1 is characterized in that: Methods for encoding two features through Rx and Ry gate operations include: The eight features are encoded onto four quantum bits, namely q0, q1, q2, and q3, and the Rx and Ry gates are selected to encode the spatial distribution and spectral intensity differences of the band features respectively.

5. The ocean cloud detection method based on quantum convolutional neural network according to claim 1, characterized in that: In the quantum convolution layer, it also includes: The parameterized revolving gate dynamically adjusts the feature response weights through gradient descent optimization, and its parameters are equivalent to the learnable parameters of the classic convolution kernel.

6. The ocean cloud detection method based on quantum convolutional neural network according to claim 5, characterized in that: When optimizing through gradient descent, optimize the parameters of the qubit gates outside the coding layer, select the cross-entropy loss as the loss function, and select Adam as the optimizer; repeat the training and save the model.

7. The ocean cloud detection method based on quantum convolutional neural network according to claim 6, characterized in that: The saved model is used to average the error when using different test sets.

8. The ocean cloud detection method based on quantum convolutional neural network according to claim 7, characterized in that: The predictive ability of the model is evaluated based on the accuracy, precision, recall and F1-score value of the test set predictions.

9. The ocean cloud detection method based on quantum convolutional neural network according to claim 1, characterized in that: When performing a pooling operation on qubits, only two qubits are measured.

10. The ocean cloud detection method based on quantum convolutional neural network according to claim 9, characterized in that: Also includes: Compare the expected values ​​of the two qubits obtained; mark the classification result with the larger expected value as 0, indicating no cloud; Otherwise, the classification result is marked as 1, indicating that there are clouds.

Images