Abnormity detection method for satellite remote sensing application system

Abstract

The invention discloses an anomaly detection method for a satellite remote sensing application system. The anomaly detection method comprises the following steps: acquiring spectral remote sensing data in real time by using a multispectral sensor and carrying out data preprocessing on the spectral remote sensing data; establishing a processing model based on a convolutional neural network, and performing feature extraction and fusion on the preprocessed spectral remote sensing data to generate spectral remote sensing feature representation; establishing a spectrum anomaly detection model based on a deep learning algorithm, performing learning training on the spectrum anomaly detection model, and deploying the detection model in a satellite remote sensing system after the training is finished; and taking the spectral remote sensing feature representation as input of a detection model, performing anomaly detection on the spectral remote sensing feature representation by the detection model, and outputting a detection result. According to the method, the processing processes of multispectral data acquisition, convolutional neural network feature extraction and deep learning anomaly detection are combined, many defects in anomaly detection in the prior art are effectively overcome, and the anomaly detection precision, efficiency and real-time performance in a satellite remote sensing application system are comprehensively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing data analysis and machine learning, and in particular to an anomaly detection method for a satellite remote sensing application system. Background Art

[0002] With the increasing demands for global environmental change, natural resource exploration and management, satellite remote sensing technology has been increasingly used in the field of earth observation. As an important part of remote sensing technology, multispectral remote sensing can effectively capture the spectral characteristics of objects by using spectral information in different bands, thereby realizing the identification and classification of objects.

[0003] Traditional multispectral remote sensing data processing methods mainly rely on rule-based or empirical statistical analysis methods, such as principal component analysis (PCA) and linear discriminant analysis (LDA). These methods can extract the global characteristics of the data to a certain extent. However, with the increase of data dimension and complexity, these methods have limited capabilities in data fusion and feature extraction, and are difficult to cope with the potential information interaction and nonlinear relationships in high-dimensional multispectral data. In addition, these methods have high requirements for data preprocessing and require manual feature selection and optimization. They are easily affected by human factors and cannot fully explore the deep features hidden in spectral data.

[0004] In recent years, with the development of computer vision and deep learning technology, especially the successful application of convolutional neural networks (CNN) in the field of image processing, deep learning algorithms have gradually been introduced into the field of remote sensing data processing. Deep learning models can automatically learn and extract high-level features in data through the construction of multi-layer neural network structures, greatly improving the ability to analyze and identify data. Compared with traditional statistical-based processing methods, deep learning has shown significant advantages in feature extraction, fusion and anomaly detection of remote sensing data.

[0005] For example, feature extraction of multispectral remote sensing data through convolutional neural networks can capture the complex relationship between different bands and achieve efficient data fusion and feature representation. This automated data processing method not only improves the accuracy and efficiency of feature extraction, but also demonstrates stronger robustness in a variety of complex scenarios. However, most of the current research on the application of deep learning models to remote sensing data analysis is still in the exploratory stage, mainly focusing on image classification and target detection applications. There is still a lack of systematic solutions and widespread applications for anomaly detection in multispectral data.

[0006] Existing technologies have many deficiencies in detecting anomalies in multispectral remote sensing data, including:

[0007] 1. Traditional anomaly detection methods, such as threshold-based detection or simple statistical feature-based detection, are difficult to deal with nonlinear features and complex background noise in the data. These methods are inefficient when processing high-dimensional, multi-band remote sensing data and are prone to false detection or missed detection;

[0008] 2. Although deep learning methods provide new ideas for remote sensing data analysis, most existing methods are not optimized for anomaly detection in multispectral data, and the generalization and real-time processing capabilities of the models still need to be improved. Summary of the invention

[0009] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract of the specification and the title of the invention of this application to avoid blurring the purpose of this section, the abstract of the specification and the title of the invention, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0010] In view of the above-mentioned existing problems, the present invention is proposed, which aims to improve the accuracy and efficiency of anomaly detection in remote sensing data by constructing a convolutional neural network processing model, extracting and fusing features of multispectral remote sensing data, and performing anomaly detection based on a deep learning algorithm.

[0011] Therefore, the technical problem solved by the present invention is that the existing multispectral remote sensing data anomaly detection technology has the problems of low detection accuracy, low computational efficiency and insufficient generalization ability.

[0012] In order to solve the above technical problems, the present invention provides the following technical solutions: using a multispectral sensor to collect spectral remote sensing data in real time and perform data preprocessing on it;

[0013] Establishing a processing model based on a convolutional neural network, extracting and fusing features of the preprocessed spectral remote sensing data, and generating spectral remote sensing feature representation;

[0014] Establishing a spectral anomaly detection model based on a deep learning algorithm and training it, and after the training, deploying the detection model in a satellite remote sensing system;

[0015] The spectral remote sensing feature representation is used as the input of the detection model, and the detection model performs anomaly detection on it and outputs the detection result.

[0016] As a preferred solution of the anomaly detection method for a satellite remote sensing application system of the present invention, the acquisition is performed by scanning the target area with the multispectral sensor to obtain spectral data of different bands;

[0017] Among them, the data sampling frequency and spatial resolution of each band are set according to the satellite orbit and sensor configuration.

[0018] As a preferred solution of the anomaly detection method for a satellite remote sensing application system described in the present invention, the spectral data in different bands at least include red light R, green light G, blue light B, and near infrared NIR data.

[0019] As a preferred solution of the anomaly detection method for a satellite remote sensing application system described in the present invention, the data preprocessing includes radiation correction, geometric correction and atmospheric correction.

[0020] As a preferred solution of the anomaly detection method for a satellite remote sensing application system according to the present invention, generating the spectral remote sensing feature representation includes:

[0021] Extract feature maps for each band;

[0022] The feature maps of each band are weighted and fused to form a multi-spectral fusion feature representation, and its mathematical expression formula is:

[0023]

[0024] Among them, F 融合 is the fused feature representation, F i is the feature map of the i-th band, α i is the fusion weight, and n is the total number of bands.

[0025] As a preferred solution of the anomaly detection method for a satellite remote sensing application system described in the present invention, the spectral anomaly detection model established based on the deep learning algorithm is an autoencoder model, including an encoder and a decoder, wherein:

[0026] The encoder compresses the input feature representation into a low-dimensional implicit representation. The encoder formula is:

[0027] h=σ(W e ·F 融合 +b e )

[0028] Among them, h is an implicit representation, W e is the weight matrix of the encoder, b e is the bias and σ is the activation function.

[0029] As a preferred solution of the anomaly detection method for satellite remote sensing application system of the present invention, the decoder attempts to reconstruct the input features, and the decoder formula is:

[0030]

[0031] in, is the reconstructed input feature, W dis the weight matrix of the decoder, b d For bias.

[0032] As a preferred solution of the anomaly detection method for a satellite remote sensing application system according to the present invention, the anomaly detection is performed, including:

[0033] The spectral features acquired in real time are expressed as F 融合 Input into the deployed detection model;

[0034] Calculate the anomaly score of each input sample and determine whether it is abnormal based on the anomaly score. The mathematical expression formula of the anomaly score is:

[0035]

[0036] Among them, S(F 融合 ) is the abnormality score, is the fusion feature reconstructed by the autoencoder, and ||·|| represents the L2 norm (Euclidean distance).

[0037] As a preferred embodiment of the anomaly detection method for a satellite remote sensing application system according to the present invention, S(F 融合 ) has a range of non-negative real numbers, where:

[0038] When the value is within the interval [0, 0.05], it is judged to be normal, that is, the input feature is considered normal and consistent with the feature distribution of the model training data;

[0039] When the value is greater than 0.05, it is judged as abnormal, that is, the input feature deviates from the normal feature distribution and there is an abnormality.

[0040] Beneficial effects of the invention: The invention effectively solves many deficiencies of the prior art in anomaly detection by combining the processing of multispectral data acquisition, convolutional neural network feature extraction and deep learning anomaly detection, and achieves a comprehensive improvement in the accuracy, efficiency and real-time performance of anomaly detection in satellite remote sensing application systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0042] Figure 1 The figure is a flow chart of the anomaly detection method for a satellite remote sensing application system shown in the present invention. DETAILED DESCRIPTION

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0044] Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without making any creative work should fall within the scope of protection of the present invention.

[0045] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0046] According to an embodiment of the present invention, Figure 1 The flowchart shown in the figure is a method for detecting anomalies in a satellite remote sensing application system, which specifically includes the following steps:

[0047] S1. Use a multispectral sensor to collect spectral remote sensing data in real time and perform data preprocessing. Among them, what needs to be explained in this step is:

[0048] The multispectral sensor scans the target area and obtains spectral data of different bands (e.g., red light R, green light G, blue light B, near infrared NIR);

[0049] The data sampling frequency and spatial resolution (e.g., 10-meter resolution, one sample per second) of each band are set according to the satellite orbit and sensor configuration.

[0050] Specifically, data preprocessing includes at least radiation correction, geometric correction and atmospheric correction, among which:

[0051] (1) Radiation correction: used to eliminate the influence of sensor response and environmental conditions;

[0052] The radiation correction formula converts the digital quantization value of the sensor into a radiation brightness value in physical units by subtracting the offset and dividing by the gain for subsequent spectral analysis and processing. The mathematical expression formula is as follows:

[0053]

[0054] Where L(λ) is the radiant brightness, DN is the digital quantization value output by the sensor, Offset is the offset, and Gain is the gain;

[0055] (2) Geometric correction: used to correct geometric distortion caused by changes in sensor viewing angle, satellite motion, and terrain undulations, ensuring the accuracy of image spatial coordinates;

[0056] Select several Ground Control Points (GCPs) to align with the reference map;

[0057] Using the polynomial geometric correction model, the image is transformed according to GCPs. The mathematical expression of the geometric correction formula (quadratic polynomial transformation) is as follows:

[0058] x′=a0+a1x+a2y+a3xy+a4x 2 +a5y 2

[0059] y′=b0+b1x+b2y+b3xy+b4x 2 +b5y 2

[0060] Among them, (x, y) is the original image coordinate, (x′, y′) is the corrected image coordinate, and a i 、b i is the geometric transformation coefficient, obtained by fitting the ground control point data;

[0061] (3) Atmospheric correction: used to eliminate the influence of aerosols and gases in the atmosphere on the spectral signal and restore the true spectral reflectance characteristics of the surface;

[0062] The 6S (Second Simulation of a Satellite Signal in the Solar Spectrum) model is used to simulate and correct the atmospheric effects;

[0063] According to the sensor position, observation angle, ground reflectivity and solar radiation parameters, the surface reflectivity R(λ) is calculated, and its mathematical expression formula is:

[0064]

[0065] Where R(λ) is the surface reflectivity, L(λ) is the radiance after radiation correction, d is the distance from the Earth to the Sun (in astronomical units), and E s (λ) is the solar radiation, θ s is the solar zenith angle, and T(λ) is the atmospheric transmittance.

[0066] Specifically, in satellite remote sensing systems, multispectral sensors are used to collect spectral data of the earth's surface in real time. This sensor can collect data in multiple bands (such as red light, green light, blue light, near infrared, etc.), and the information in each band provides a reflection of the different spectral characteristics of the target surface.

[0067] As an example, if the digital quantization value DNR of the red light band (R) obtained by multispectral data acquisition is 1500, the sensor's Offset is 50, and Gain is 10, then the radiation brightness obtained by radiation correction is:

[0068]

[0069] In the further geometric correction and atmospheric correction process, GCPs are used for coordinate alignment and 6S model is used to eliminate the atmospheric effect;

[0070] Select the distance from the Earth to the Sun d = 1AU (mean Earth-Sun distance): in astronomical units (AU);

[0071] Select solar radiation Es(λ): For the typical value of the red light band, take it from a typical solar radiation table, such as:

[0072] E s (R)=2000W / m 2 / μm

[0073] Select the solar zenith angle θ s : If it is 45 degrees, then:

[0074]

[0075] Select the atmospheric transmittance T(λ): define T(R) = 0.1 (typical atmospheric correction parameter);

[0076] The final surface reflectivity can be calculated by the following formula:

[0077]

[0078] R R =0.222, this value represents the surface reflectance of the red light band after radiation correction, geometric correction and atmospheric correction.

[0079] It can be seen from the above examples that this step realizes the comprehensive and accurate acquisition of multi-band spectral information of the target area by using multi-spectral sensors to collect spectral remote sensing data in real time. Multi-spectral data can capture the reflection and absorption characteristics of different bands and reveal the spectral characteristics of the ground objects, thereby providing a rich source of information for subsequent feature extraction and anomaly detection.

[0080] Preferably, through data preprocessing, the quality of data is improved, measurement errors and environmental influences are eliminated, and the accuracy and consistency of subsequent data processing are ensured, thereby achieving the beneficial effect of obtaining high-quality remote sensing data in real time and reliably.

[0081] S2. Establish a processing model based on a convolutional neural network to extract and fuse the preprocessed spectral remote sensing data to generate spectral remote sensing feature representation. Among them, the following points need to be explained in this step:

[0082] Extract feature maps for each band;

[0083] The feature maps of each band are weighted and fused to form a multi-spectral fusion feature representation, and its mathematical expression formula is:

[0084]

[0085] Among them, F 融合 is the fused feature representation, F i is the feature map of the i-th band, α i is the fusion weight, and n is the total number of bands.

[0086] Preferably, the processing model combines the feature information of multiple bands into a unified feature representation through weighted averaging, thereby enhancing the perception of spectral features by the anomaly detection model.

[0087] As an example, the extracted feature map is F R ,F G ,F B ,F NIR , corresponding to the red light, green light, blue light and near-infrared bands respectively.

[0088] As an example, if the feature maps extracted from the red (R), green (G), and blue (B) bands are F R =0.3, F G =0.4F, F B =0.5, and the fusion weights are α R =0.2, α G =0.3, α B =0.5;

[0089] The fused feature representation F 融合 =0.2×0.3+0.3×0.4+0.5×0.5=0.38.

[0090] It should be noted that this step uses a convolutional neural network (CNN) model to perform multi-level and multi-scale spectral feature extraction, thereby realizing deep feature mining of remote sensing data. The convolutional neural network can automatically learn the spatial and spectral features in the data and capture the complex relationship between the bands in the multispectral data through convolution operations.

[0091] Preferably, data fusion further integrates multi-band information to generate a more representative spectral remote sensing feature representation, reduce information redundancy, strengthen key features, and effectively improve the accuracy and reliability of anomaly detection. Through feature extraction and fusion, the beneficial effect of constructing a feature representation with strong robustness and wide applicability is achieved, which provides a guarantee for the efficiency and accuracy of subsequent anomaly detection.

[0092] S3: Establish a spectral anomaly detection model based on a deep learning algorithm and train it. After the training is completed, the detection model is deployed in the satellite remote sensing system. The following points need to be explained in this step:

[0093] The spectral anomaly detection model established based on the deep learning algorithm is an autoencoder model, which consists of two parts: an encoder and a decoder. The encoder compresses the input feature representation into a low-dimensional implicit representation, and the decoder attempts to reconstruct the input features and uses the L2 norm as the loss function to minimize the error between the input features and the reconstructed features.

[0094] Specifically, the encoder formula is:

[0095] h=σ(W e ·F 融合 +b e )

[0096] Among them, h is an implicit representation, W e is the weight matrix of the encoder, b e is the bias, and σ is the activation function (e.g., ReLU).

[0097] The decoder formula is:

[0098]

[0099] in, is the reconstructed input feature, W d is the weight matrix of the decoder, b d For bias.

[0100] The model learns the compressed representation of input data through encoding and decoding structures, so that it can identify abnormal samples that deviate from the normal data distribution in anomaly detection.

[0101] In an optional embodiment, an unsupervised learning algorithm is used to train the autoencoder (i.e., the spectral anomaly detection model), the goal of which is to minimize the mean square error (MSE) between the input features and the reconstructed features. The mathematical expression formula of the loss function is:

[0102]

[0103] Among them, L is the reconstruction error (loss function), N is the number of training samples, and F 融合 ,i is the input feature of the i-th sample, is the reconstructed input feature.

[0104] As an example, if the input fusion feature is F 融合 =[0.3,0.4,0.5], the autoencoder reconstructs the feature

[0105] The reconstruction error

[0106] In an optional implementation, the trained model (autoencoder) is deployed to a ground processing station or cloud server of a satellite remote sensing system to start running, ensuring that the data input stream matches the model input format, and setting a real-time data stream interface.

[0107] It should be noted that through deep learning training, the detection model can capture the potential abnormal features and complex nonlinear relationships in spectral data, effectively improving the sensitivity and accuracy of detection.

[0108] Preferably, the trained detection model is deployed in a satellite remote sensing system, which can realize automated anomaly detection without human intervention, greatly improving detection efficiency and real-time performance.

[0109] S4: The spectral remote sensing feature representation is used as the input of the detection model, and the detection model performs anomaly detection on it and outputs the detection result. Among them, what needs to be explained in this step is:

[0110] The spectral features acquired in real time are expressed as F 融合 Input to the deployed detection model;

[0111] Calculate the anomaly score of each input sample and determine whether it is abnormal based on the anomaly score. The mathematical expression formula of the anomaly score is:

[0112]

[0113] Among them, S(F 融合 ) is the abnormality score, is the fusion feature reconstructed by the autoencoder, and ||·|| represents the L2 norm (Euclidean distance).

[0114] It should be noted that S(F 融合 ) has a range of non-negative real numbers, that is, S(F 融合 )≥0, when the input features and the reconstructed features are exactly the same, the anomaly score is 0 if and only if When , the input feature completely conforms to the normal data distribution and is judged to be normal. Among them:

[0115] When the value is within the interval [0, 0.05], it is judged to be normal, that is, the input feature is considered normal and consistent with the feature distribution of the model training data;

[0116] When the value is greater than 0.05, it is judged as abnormal, that is, the input feature deviates from the normal feature distribution and there is an abnormality.

[0117] It can be seen that the anomaly score formula evaluates the degree of abnormality of the sample by calculating the difference between the input features and the reconstructed features. The higher the anomaly score, the greater the deviation and the higher the possibility of abnormality.

[0118] As an example, if the input feature is F 融合 =[0.35,0.45,0.55], the autoencoder reconstructed feature is Then the abnormal score S(F 融合 )=0.042, and compare it with the above threshold range, that is, 0.042<0.0.5, it is determined to be normal.

[0119] In an optional implementation, the application system records the spatial location, timestamp, and abnormal feature description of the data sample determined to be abnormal, generates an anomaly detection report including the coordinates, time, and band feature information of the abnormal point, and provides a visual anomaly distribution map to assist user decision-making.

[0120] As an example, the visualized anomaly distribution map generates a graphical representation of the anomaly detection results through a visualization formula to help users quickly locate the abnormal area and conduct subsequent analysis. The visualization formula is:

[0121]

[0122] Among them, (x, y) is the spatial coordinate of the sample, and Threshold is the threshold for anomaly detection.

[0123] As an example, in a certain area monitored by satellite remote sensing, if the anomaly detection results show multiple abnormal points with scores of 0.07, 0.09 and 0.06 (exceeding the threshold of 0.05) respectively, the system will generate an anomaly report, listing the specific coordinates and time of these abnormal points, and display them visually on a map to help users understand and deal with these abnormal situations in a timely manner.

[0124] It is also necessary to reiterate that the embodiment of the present invention uses the spectral remote sensing feature representation generated in the aforementioned steps as input, and the anomaly detection model analyzes and detects it. The model can directly use the optimized feature data to quickly identify abnormal signals in the data and output the detection results, making the detection process more targeted and efficient, reducing the redundancy of data processing and unnecessary calculations, and ultimately achieving the beneficial effect of improving the accuracy and real-time performance of anomaly detection, ensuring that abnormal changes in the target area can be identified and responded to in a timely and accurate manner.

[0125] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. 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 solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An anomaly detection method for a satellite remote sensing application system, characterized in that: include: Use multispectral sensors to collect spectral remote sensing data in real time and perform data preprocessing; Establishing a processing model based on a convolutional neural network, extracting and fusing features of the preprocessed spectral remote sensing data, and generating spectral remote sensing feature representation; Establishing a spectral anomaly detection model based on a deep learning algorithm and training it, and after the training, deploying the detection model in a satellite remote sensing system; The spectral remote sensing feature representation is used as the input of the detection model, and the detection model performs anomaly detection on it and outputs the detection result.

2. The anomaly detection method for a satellite remote sensing application system according to claim 1, characterized in that: The acquisition scans the target area through the multispectral sensor to obtain spectral data of different bands; Among them, the data sampling frequency and spatial resolution of each band are set according to the satellite orbit and sensor configuration.

3. The anomaly detection method for a satellite remote sensing application system according to claim 2, characterized in that: The spectral data of different wavebands at least include red light R, green light G, blue light B, and near infrared NIR data.

4. The anomaly detection method for a satellite remote sensing application system according to claim 1, characterized in that: The data preprocessing includes radiation correction, geometric correction and atmospheric correction.

5. The anomaly detection method for a satellite remote sensing application system according to claim 1, characterized in that: Generating the spectral remote sensing feature representation includes: Extract feature maps for each band; The feature maps of each band are weighted and fused to form a multi-spectral fusion feature representation, and its mathematical expression formula is: Among them, F 融合 is the fused feature representation, F i is the feature map of the i-th band, α i is the fusion weight, and n is the total number of bands.

6. The anomaly detection method for a satellite remote sensing application system according to claim 1, characterized in that: The spectral anomaly detection model established based on the deep learning algorithm is an autoencoder model, which includes an encoder and a decoder, wherein: The encoder compresses the input feature representation into a low-dimensional implicit representation. The encoder formula is: h=σ(W e ·F 融合 +b e ) Among them, h is an implicit representation, W e is the weight matrix of the encoder, b e is the bias and σ is the activation function.

7. The anomaly detection method for a satellite remote sensing application system according to claim 6, characterized in that: The decoder attempts to reconstruct the input features. The decoder formula is: in, is the reconstructed input feature, W d is the weight matrix of the decoder, b d For bias.

8. The anomaly detection method for a satellite remote sensing application system according to claim 1, characterized in that: The abnormality detection includes: The spectral features acquired in real time are expressed as F 融合 Input into the deployed detection model; Calculate the anomaly score of each input sample and determine whether it is abnormal based on the anomaly score. The mathematical expression formula of the anomaly score is: Among them, S(F 融合 ) is the abnormality score, is the fusion feature reconstructed by the autoencoder, and ||·|| represents the L2 norm (Euclidean distance).

9. The anomaly detection method for a satellite remote sensing application system according to claim 8, characterized in that: S(F 融合 ) has a range of non-negative real numbers, where: When the value is in the interval [0, 0.05], it is judged to be normal, that is, the input feature is considered normal and consistent with the feature distribution of the model training data; When the value is greater than 0.05, it is judged as abnormal, that is, the input feature deviates from the normal feature distribution and there is an abnormality.

Images