Optical remote sensing image target change detection method and system combining depth features

Abstract

The invention provides a depth feature combined optical remote sensing image target change detection method and system. The method comprises the steps of performing feature extraction on high-resolution remote sensing images of two time phases; carrying out difference calculation to obtain a feature difference graph between the features of the two images; performing double-round depth feature supervision on the feature difference image to obtain multi-scale features of change information; introducing an attention mechanism module into the multi-scale features, and outputting features Y and a feature map formed by the features Y; activating the feature Y by using an activation function to generate a change graph, and inputting the feature graph optimized by the attention mechanism into a classifier to generate a target change detection result; a cross entropy loss function is used for measuring the difference between a model prediction result and a real result, and training and optimization of a target change detection model are guided. According to the technical scheme, the technical problem that in the prior art, when complex edge information and tiny changes are captured, boundary blur or detail missing often occurs is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to an optical remote sensing image target change detection method and system combining depth features. Background Technique

[0002] Optical remote sensing image target change detection is a core task in the field of remote sensing. Its main objective is to identify changes in land cover and land use by comparing remote sensing images of the same area at different times (Literature 1 [1].

[0003] A. Singh, “Review article digital change detection techniques using remotely-sensed data,”, Int. J. Remote Sens., vol. 10, no. 6, pp. 989 - 1003, jun. 1989, doi: 10.1080 / 01431168908903939.). Change detection methods are commonly used for urban land use dynamic monitoring (Literature 2 P.K. Mishra, A. Rai, and S.C. Rai, “Land use and land cover change detection using geospatial techniques in the Sikkim Himalaya, India,” Egypt. J. Remote Sens. Space Sci., vol. 23, no. 2, pp. 133 - 143, Aug. 2020, doi: 10 / 1016 / j.ejrs.2019.02.001.), urban planning (Literature 3 G. Xian and C. Homer, “Updating the 2001 National Land Cover Database impervious surface products to 2006 unsing Landsat imagery change detection methods,” Remote Sens. Environ., vol. 114, no. 8, pp. 1676 - 1686, Aug. 2010, doi: 10.1016 / j.rse.2010.02.018.), environmental monitoring (Literature 4 C. Song, B. Huang, L. ke, and K.S. Richards, “Remote sensing of alpine lake water environment changes on the Tibetan Plateau and surroundings: A view,” ISPRS J. Photogramm. Remote Sens., vol. 92, pp. 26 - 37, Jun. 2014, doi: 10.1016 / j.isprsjprs.2014.03.001.), disaster assessment and management (Literature 5 P. Lu, Y. Qin, Z. Li, A.C. Mondini, and N.Casagli, "Landslide mapping from multi-sensor data through improved change detection-based Markov random field," Remote Sens. Environ., vol. 231, Sep. 2019, Art. No. 111235, doi: 10.1016 / j.rse.2019.111235.) and other fields. Especially for sudden disaster events, it is very important to quickly and accurately assess the damage of building targets. With the continuous prosperity and development of domestic optical sensors in recent years, it provides an important guarantee for generating a large number of high-resolution remote sensing images with high temporal resolution and high spatial resolution characteristics, which also provides data support for the booming development of change detection methods.

[0004] Traditional change detection methods mainly include: 1) Difference-based change detection methods, which judge whether changes have occurred by calculating the spectral value differences of pixels at the same location, including image interpolation method, ratio method and change vector analysis; 2) Classification-based change detection methods, which independently classify multi-temporal images and then compare the classification results to determine the change areas, including maximum likelihood classification and support vector machine; 3) Object-based change detection methods, which segment remote sensing images into multiple objects and then analyze the geometric and spectral changes between objects; 4) Time series-based change detection methods, which analyze the change trend by using the time series information of multi-temporal images, including time series analysis method and change point detection method. Although traditional optical remote sensing image change detection methods meet the change detection requirements to a certain extent, their limitations are particularly obvious in complex scenarios: 1) Sensitive to noise, relying on spectral changes, easily affected by illumination conditions, sensor noise and image registration errors, resulting in a decrease in detection accuracy; 2) Limited feature expression, relying on manually designed features, difficult to fully capture multi-dimensional change information in complex scenarios, such as texture, geometry and semantic features; 3) Lack of robustness, poor adaptability in complex surface cover types, multi-source data fusion and multiple resolutions, prone to missed detection and false detection; 4) Computational complexity, object-based and time series-based methods require complex segmentation or model construction, with high requirements for computing resources and not suitable for large-scale applications.

[0005] Deep learning methods, with their powerful feature extraction capabilities and end-to-end learning frameworks, have overcome many deficiencies of traditional methods, mainly including: 1) Change detection methods based on convolutional neural networks (CNNs), which automatically extract multi-scale features from images by stacking convolutional layers and pooling layers, including FCN (Reference 6: Daudt R C, Saux B L, Boulch A, 2018. Fully Convolutional Siamese Networks for Change Detection [A]. arXiv.), two-stream CNN models, U-Net (Reference 7: Chen H, Shi Z, 2020. A Spatial-Temporal Attention-Based Method and a New Dataset for Remote Sensing Image Change Detection [J]. Remote Sensing, 12(10):1662., Reference 8: Fang S, Li K, Shao J, et al., 2022. SNUNet-CD: A Densely Connected Siamese Network for Change Detection of VHR Images [J]. IEEE Geoscience and Remote Sensing Letters, 19:1-5., Reference 9: Han C, Wu C, Guo H, et al., 2023. HANet: A Hierarchical Attention Network for Change Detection with bi-temporal very-high-resolution remote sensing images [J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing:1-17.), etc.; 2) Change detection methods based on generative adversarial networks (GANs), which enhance feature expression capabilities through game learning between generators and discriminators to generate high-quality change region maps; 3) Change detection methods based on recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), which use RNNs and LSTMs to capture temporal features in time series images and analyze change trends; 4) Change detection methods based on Transformers (Reference 10: Chen H, Qi Z, Shi Z, 2022.Remote Sensing Image Change Detection With Transformers[J]. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-14.) captures change features in long-distance contexts through self-attention mechanisms and exhibits high robustness in complex scenarios. Generally speaking, change detection methods based on deep learning have the following advantages: 1) Automatic feature extraction, which can extract multi-scale and multi-level features without relying on manual design, significantly improving the feature expression ability, especially in scenarios with complex textures and diverse ground objects; 2) Multi-modal fusion, which supports the joint input of optical images and other data such as radar and thermal infrared images to improve the detection effect through multi-modal information; 3) End-to-end learning framework, which integrates traditional data preprocessing, feature extraction, change detection and other steps into an end-to-end framework; 4) Ability to capture complex change patterns, which can capture non-linear complex changes and meet the detailed detection requirements of scenarios such as buildings and roads.

[0006] Although deep learning has made significant progress in the change detection of optical remote sensing image targets, its application still faces some challenges. For example, the ability to extract target edge details and minute changes still needs to be improved. For instance, boundary blurring or detail omission often occurs when capturing complex edge information and minute changes. Therefore, how to effectively improve the extraction of change information in the target edge area of remote sensing images and reduce the missed detection phenomenon is the first key difficulty in this field. How to effectively extract the change information inside the change information, especially inside building targets, and reduce the hole phenomenon is the second key difficulty in this field. In addition, the accuracy of overall target change detection still needs to be improved, which is the third key difficulty in this field. Summary of the Invention

[0007] The present invention provides a method and system for change detection of optical remote sensing image targets combining deep features, which can solve the technical problems of boundary blurring or detail omission that often occur in the prior art when capturing complex edge information and minute changes.

[0008] According to one aspect of the present invention, there is provided a method for detecting target changes in optical remote sensing images by combining depth features. The method for detecting target changes in optical remote sensing images by combining depth features includes: Step 1, extracting features from high-resolution remote sensing images of two time phases of the same area; Step 2, calculating the difference between the high-resolution remote sensing images and the extracted features to obtain a feature difference map between the features of the two images; Step 3, obtaining multi-scale features of change information from the feature difference map through two-round depth feature supervision; Step 4, introducing an attention mechanism module into the multi-scale features of the change information to enhance the feature expression of the change information edges and internal regions, and outputting feature Y and a feature map composed of feature Y; Step 5, using an activation function to activate the feature Y output by the attention mechanism module to generate a change map, and inputting the feature map optimized by the attention mechanism into a classifier to generate a target change detection result; Step 6, according to the target change detection result, using a cross-entropy loss function to measure the difference between the model prediction result and the true result, and guiding the training and optimization of the target change detection model.

[0009] According to another aspect of the present invention, there is provided a model for detecting target changes in optical remote sensing images by combining depth features. The model for detecting target changes in optical remote sensing images by combining depth features uses the method for detecting target changes in optical remote sensing images by combining depth features as described above to detect target changes in optical remote sensing images. The model for detecting target changes in optical remote sensing images by combining depth features includes a feature extraction module, a two-round depth feature supervision module, an attention mechanism module, and a classification and change detection module. The feature extraction module is used to extract deep features from the input image based on a convolutional neural network. The two-round depth feature supervision module is used to achieve multi-scale change features through two-round depth feature fusion and supervision optimization. The attention mechanism module is used to optimize the feature expression of the edges and internal regions. The classification and change detection module is used to generate a final target change detection result by combining the feature fusion output.

[0010] According to yet another aspect of the present invention, there is provided a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the steps of the method for detecting target changes in optical remote sensing images by combining depth features as described above.

[0011] According to still another aspect of the present invention, there is provided a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, it implements the steps of the method for detecting target changes in optical remote sensing images by combining depth features as described above.

[0012] Applying the technical solution of the present invention, a method for detecting target changes in optical remote sensing images combined with depth features is provided. Through double-round depth feature supervision, feature fusion, and an attention mechanism module, this method gradually improves the ability to extract edge information and internal details of the target change area, and enhances the accuracy and robustness of target change detection. The main innovations are as follows: 1) A dual-depth feature supervised change detection network is proposed, which realizes the fusion of multi-scale change feature information through two rounds of depth feature supervision, thereby gradually improving the model's expression of the edge information and internal features of building targets; 2) An attention mechanism module for improving change features is proposed, which guides the model to focus on the edge and internal area information of the building through the change feature module, thereby making up for the incomplete extraction of the edge and internal information of building targets; 3) An efficient target change detection network structure is constructed, which adapts to multi-resolution remote sensing images, and its superiority is verified on two public datasets. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The accompanying drawings included are used to provide a further understanding of the embodiments of the present invention, which form a part of the specification, are used to illustrate the embodiments of the present invention, and are used to explain the principles of the present invention together with the text description. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0014] Figure 1 It shows a schematic diagram of a method for detecting target changes in optical remote sensing images combined with depth features provided according to a specific embodiment of the present invention;

[0015] Figure 2 It shows a schematic diagram of an attention mechanism module provided according to a specific embodiment of the present invention;

[0016] Figure 3 It shows a schematic diagram of the effect on the LEVIR-CD dataset provided according to a specific embodiment of the present invention;

[0017] Figure 4 It shows a schematic diagram of the effect on the WHU-CD dataset provided according to a specific embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and in no way limits the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0019] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0020] Unless otherwise specifically stated, the relative arrangements of components and steps, numerical expressions, and values set forth in these embodiments do not limit the scope of the present invention. At the same time, it should be understood that, for the sake of convenience of description, the dimensions of the various parts shown in the drawings are not drawn in actual proportional relationships. Technologies, methods, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and devices should be regarded as part of the authorization specification. In all the examples shown and discussed here, any specific value should be construed as merely exemplary and not as a limitation. Therefore, other examples of the exemplary embodiments may have different values. It should be noted that like reference numerals and letters denote like items in the following drawings, and thus, once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0021] As Figure 1 and Figure 2 shown, according to a specific embodiment of the present invention, an optical remote sensing image target change detection method combining depth features is provided. The optical remote sensing image target change detection method combining depth features includes: Step 1, extracting features from two-phase high-resolution remote sensing images X 1 and X 2 of the same area; Step 2, for the high-resolution remote sensing images X 1 and X 2Calculate the differences of the extracted features to obtain a feature difference map between the features of the two images; Step 3, obtain multi-scale features of the change information from the feature difference map through two rounds of deep feature supervision; Step 4, introduce an attention mechanism module into the multi-scale features of the change information to enhance the feature expressions of the edges and internal regions of the change information, and output the feature Y and the feature map ΔF composed of the feature Y (a) ; Step 5, use an activation function to activate the feature Y output by the attention mechanism module to generate a change map, and input the feature map ΔF (a) optimized by the attention mechanism into a classifier to generate the target change detection result; Step 6, according to the target change detection result, use the cross-entropy loss function to measure the difference between the model prediction result and the true result, and guide the training and optimization of the target change detection model.

[0022] Applying this configuration method provides an optical remote sensing image target change detection method combining deep features. This method gradually improves the extraction ability of the edge information and internal details of the target change area through two rounds of deep feature supervision, feature fusion and an attention mechanism module, and improves the accuracy and robustness of target change detection. The main innovation points are as follows: 1) A dual deep feature supervision change detection network is proposed, and multi-scale change feature information fusion is realized through two rounds of deep feature supervision, so as to gradually improve the model's expression of the edge information and internal features of building targets; 2) An attention mechanism module for improving change features is proposed, and the change feature module is used to guide the model to pay attention to the information of the building edge and internal area, so as to make up for the incomplete extraction of the building target edge and internal information; 3) An efficient target change detection network structure is constructed, which is adapted to multi-resolution remote sensing images, and its superiority is verified on 2 public datasets.

[0023] Furthermore, in Step 1, the deep features of the high-resolution remote sensing images X 1 and X 2 at two time phases are extracted through a VGG-16 feature extraction network with shared parameters, where F 1 = f(X 1 ; θ), F 1 = f(X 1 ; θ), f(·; θ) represents the feature extraction network, is the feature map of the high-resolution remote sensing image X 1 , is the feature map of the high-resolution remote sensing image X 2 , W represents the width of the feature map, H represents the height of the feature map, D represents the number of channels of the feature map, and θ is the network parameter.

[0024] Further, in Step 2, the feature difference map ΔF between the two image features can be obtained according to ΔF = |F 1 - F 2 |.

[0025] Further, Step 3 specifically includes: performing the first round of supervision, performing a convolution operation on the feature difference map to extract preliminary change features; fusing the preliminary change features obtained from the first round of supervision with the original features of the two images extracted in Step 2 to form the result of the second round of deep supervision, and obtaining the multi-scale features of the second round of deep supervision.

[0026] Among them, in Step 3, the preliminary change feature ΔF (1) is calculated according to , and the multi-scale feature ΔF (2) of the second round of deep supervision is calculated according to where g (1) represents the first convolution operation, g (2) represents the second convolution operation, represents the parameters in the network, and + represents the concatenation operation.

[0027] Further, in the present invention, Step 4 specifically includes: using the multi-scale feature ΔF (2) of the change information as the input feature X of the attention mechanism; reducing the dimension of the input feature X through two different convolution operations to obtain the dimension-reduced features Q and K; reconstructing the dimension-reduced features Q and K, and transforming the dimension-reduced features Q and K into two-dimensional tensors Q' and K'; calculating the attention weight matrix A based on the two-dimensional tensors Q' and K'; using the attention weight matrix A to perform weighted summation on the input feature X to obtain the weighted feature map V; transforming the weighted feature map V into the original size to obtain the weighted feature V'; and fusing the input feature X and the weighted feature V' to obtain the output feature Y.

[0028] Among them, the dimension-reduced features Q and K are calculated according to , where Conv 1 and Conv 2 represent the convolution operation, the dimension-reduced features Q and K each have channels, C represents the number of channels of the input feature map X, W represents the width of the feature map X, H represents the height of the feature map X, represents the set of real numbers.

[0029] According to another aspect of the present invention, there is provided an optical remote sensing image target change detection model integrating depth features. The optical remote sensing image target change detection model integrating depth features uses the optical remote sensing image target change detection method integrating depth features as described above to perform optical remote sensing image target change detection. The optical remote sensing image target change detection model includes a feature extraction module, a two-round depth feature supervision module, an attention mechanism module, and a classification and change detection module. The feature extraction module is used to extract deep features from the input image based on a convolutional neural network. The two-round depth feature supervision module is used to achieve multi-scale change features through two-round depth feature fusion and supervision optimization. The attention mechanism module is used to optimize the feature expressions of the edge and internal regions. The classification and change detection module is used to generate the final target change detection result by combining the feature fusion output.

[0030] According to still another aspect of the present invention, there is provided a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the steps of the optical remote sensing image target change detection method integrating depth features as described above.

[0031] According to yet another aspect of the present invention, there is provided a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the steps of the optical remote sensing image target change detection method integrating depth features as described above are implemented.

[0032] For a further understanding of the present invention, the following combines Figures 1 to 4 to detail the optical remote sensing image target change detection method integrating depth features provided by the present invention.

[0033] As Figures 1 to 4 shown, according to a specific embodiment of the present invention, there is provided an optical remote sensing image target change detection method. Through two-round depth feature supervision, feature fusion, and an attention mechanism module, the ability to extract edge information and internal details of the change region is gradually improved, and the accuracy and robustness of target change detection are enhanced. The main innovation points are as follows:

[0034] 1) A dual-depth feature supervision change detection network is proposed, and multi-scale change feature information fusion is achieved through two-round depth feature supervision, thereby gradually improving the model's expression of the edge information and internal features of building targets;

[0035] 2) An attention mechanism module for improving change features is proposed. The change feature module is used to guide the model to focus on the edge and internal region information of buildings, thereby making up for the incomplete extraction of the edge and internal information of building targets;

[0036] 3) An efficient target change detection network structure was constructed, which is adapted to multi-resolution remote sensing images, and its superiority was verified on two public datasets.

[0037] The main process is as follows:

[0038] The proposed optical remote sensing image target change detection model framework combining deep features mainly consists of the following modules:

[0039] 1) Feature extraction module: Extract deep features from the input image based on the convolutional neural network (CNN);

[0040] 2) Dual-wheel deep feature supervision module: Optimize the multi-scale change features of the soil enterprise through dual-wheel deep feature fusion and supervision;

[0041] 3) Attention mechanism module: Introduce a change feature enhancement module to optimize the feature expression of the edge and internal regions;

[0042] 4) Classification and change detection: Combine the feature fusion output to generate the final target change detection result.

[0043] The specific process of the algorithm is as follows:

[0044] Step 1, let the high-resolution remote sensing images of two time phases in the same area be X 1 and X 2 (The purpose of setting the images of two time phases is to find the changed parts between the two images), and their size is H×W. X 1 and X 2 are in a one-to-one correspondence relationship at the same position pixels, and the value of each pixel is 0 (no change) or 1 (change);

[0045] Perform feature extraction on X 1 and X 2 images. The feature extraction steps are as follows:

[0046] Extract the deep features of the dual-time phase images through the VGG-16 feature extraction network with shared parameters. The meaning of shared parameters is that each convolutional layer in VGG-16 has multiple convolutional kernels, and these convolutional kernels are invariant when processing the input data, that is, they will slide over the entire input image to extract local features, but the values of the convolutional kernels will not change. The advantage of shared parameters is that it can reduce the computational amount, extract general features and improve the generalization ability. Its specific mathematical expression is as follows:

[0047] F 1 = f(X 1 ; θ), F 2 = f(X 2 ; θ)

[0048] Among them, f(·; θ) represents the feature extraction network, F is the feature map, θ is the network parameter, W represents the width of the feature map, H represents the height of the feature map, and D represents the number of channels of the feature map.

[0049] Step 2: Calculate the difference between the features extracted from X 1 and X 2 images, and calculate the difference feature map between the features of the two images. The formula is:

[0050] ΔF = |F 1 - F 2 |

[0051] Step 3: After extracting the feature difference map ΔF, the multi-scale features of the change information can be obtained through two rounds of deep feature supervision. The two rounds of deep feature supervision steps are as follows:

[0052] The first round of supervision: Perform a convolution operation on the feature difference map ΔF to extract the preliminary change features. The calculation process is:

[0053]

[0054] The second round of supervision: Fuse the preliminary change features in the first round with the original features to form the result of the second round of deep supervision. The calculation process is:

[0055]

[0056] Among them, ΔF (1) is the multi-scale feature of the first round of deep supervision, ΔF (2) is the multi-scale feature of the second round of deep supervision, g (1) and g (2) represent the convolution operation, and represent the parameters in the network, and + represents the splicing operation.

[0057] Step 4: After two rounds of deep supervision, an attention mechanism module is introduced to enhance the feature expression of the edges and internal regions of the change information. The feature calculation process of the attention mechanism is:

[0058] ΔF (a) = Attention(ΔF (2) )

[0059] Among them, Attention represents the attention mechanism operation, and its specific calculation steps are as follows:

[0060] 4.1 Feature dimensionality reduction: Use the multi-scale feature ΔF (2) obtained above as the input feature X of the attention mechanism, and perform dimensionality reduction on it through two different convolution operations:

[0061]

[0062] Among them, Conv 1 and Conv 2 represent the convolution operation. The output features Q and K respectively have channels. C represents the number of channels of the input feature map X, W represents the width of the feature map X, and H represents the height of the feature map X. represents the set of real numbers.

[0063] 4.2, Feature Reconstruction: Reconstruct the downsampled features Q and K, transform Q and K into two-dimensional tensors, and the calculation process is as follows:

[0064]

[0065] 4.3, Attention Weight Calculation: Calculate the attention weight matrix A based on the two-dimensional tensors Q' and K':

[0066]

[0067] Among them, the softmax function is the activation function.

[0068] 4.4, Attention-Weighted Feature: Weighted feature calculation, use the attention weight matrix A to perform weighted summation on the input feature X to obtain the weighted feature map V:

[0069]

[0070] 4.5, Transform the weighted feature map V into the original size:

[0071]

[0072] 4.6, Output Feature Y: Fuse the input and weighted features, fuse the input feature X and the weighted feature V' (for example, by element-wise addition), and the output feature Y constitutes the feature map ΔF (a) .

[0073] Y = X + V'

[0074] Step Five, Change Map Generation: The final change map is generated through a specific operation (such as the activation function σ):

[0075] Change Map = σ(Y)

[0076] Classification and Change Detection: Input the optimized feature map ΔF (a) into the classifier to generate the target change detection result:

[0077]

[0078] Among them, represents the classification network, is the classification network parameter, and σ is the activation function (such as Softmax or Sigmoid).

[0079] Step six, finally, after the network obtains the change detection result Y′ each time, the cross-entropy loss function is used to measure the difference between the model prediction result and the true result, so as to guide the training and optimization of the model (that is, training and optimizing the feature extraction module, the double-round deep feature supervision module, the attention mechanism module, and the classification and change detection module). The following are the specific definitions of these two loss functions:

[0080] Cross-entropy loss: Usually used in classification tasks, it measures the difference between the probability distribution predicted by the model and the true label distribution. For binary classification tasks, the definition of cross-entropy loss is:

[0081]

[0082] Among them, the total number of samples in the above target change detection result Y′ is N, and y i is the true label (0 or 1) of the i-th sample; p i is the probability that the i-th sample is predicted as the positive class. The smaller the cross-entropy loss, the closer the probability distribution predicted by the model is to the true label distribution.

[0083] The following verifies the effect of the present invention through experiments.

[0084] This experiment was written in the Python language and used the most mainstream PyTorch framework to reproduce the proposed model. The experimental results were trained and tested on a single NVIDIA RTX 3090 GPU.

[0085] The datasets used in the experiment are four open-source datasets, LEVIR-CD, WHU-CD. For fair comparison, all images were cropped to a size of 256×256. The details are as follows:

[0086] Table 1

[0087]

[0088] Evaluation metrics: Evaluations are performed using F1, Recall (simply Rec.), Precision (simply Pre.), and IoU (Intersection over Union). All of them can be calculated from TP (True Positive), TN (True Negative), FP (False Positive), and FN (False Negative). The following are the calculation methods for each metric:

[0089]

[0090] Table 2

[0091]

[0092]

[0093] On the LEVIR-CD dataset, compared with other methods, the present invention has the highest values in these metrics, indicating that the method optimizes Precision and Recall more balancedly, thus improving IoU and the overall detection performance, demonstrating the superior overall performance of the present invention on this dataset. On the WHU-CD dataset, it further proves its excellent edge information extraction ability and robustness in complex scenarios. In addition, the present invention leads comprehensively in terms of Precision, Recall, F1-score, and IoU, indicating that it achieves a good balance in reducing false positives and false negatives. In terms of the ability to extract edge information and internal details, the performance of the present invention shows that it has a significant advantage in accurately locating boundaries. The ability to capture internal details of buildings has also been optimized, further improving IoU and Recall. In terms of the robustness and generality of the model, it performs excellently on both datasets, indicating that the model is not only applicable to a single scenario but also has strong generality and robustness. In summary, the present invention is the method with the best performance in the table, fully demonstrating the effectiveness of double deep feature supervision and the attention mechanism in remote sensing image target change detection. Its significant improvements in edge extraction and internal detail capture provide new ideas for target change detection in complex scenarios and also lay a technical foundation for practical applications in related fields.

[0094] From the target change detection result graphs of the given LEVIR-CD and WHU-CD datasets, we can observe the detection performance of different methods for the changed areas in remote sensing images. 1) The present invention is excellent in the ability to extract edge information: the edge detection results of some methods appear incoherent or overly smooth, while the present invention can accurately capture the changes in the building edges, with clear and continuous boundaries. Compared with other methods, it effectively avoids the blurring and breakage phenomena during edge extraction, especially showing strong robustness in complex backgrounds. 2) The present invention has higher integrity in the details within the changed information: it performs excellently in detecting the changed areas inside buildings, without obvious "hole" phenomena. Compared with the possible internal omissions (such as local changes being ignored) in other methods, the results of the present invention are closer to the true labels. 3) The present invention has fewer false detection areas: in the detection results, some methods (such as the middle few rows) have significant false detection phenomena in the background, while the present invention shows stronger noise suppression ability. For complex backgrounds or areas without changes, there are almost no redundant false detection points. This indicates that the feature supervision mechanism of the present invention effectively guides the model to focus on the real changed areas. 4) The present invention has more precise coverage of the changed areas: the detected changed areas have a very high degree of matching with the true labels. Whether it is small-scale changes or large-area regions, the present invention can accurately mark them, demonstrating excellent adaptability to changes at different scales. In summary, the target change detection results of the present invention on the LEVIR-CD and WHU-CD datasets show the following advantages: precise edge extraction, clearer and more accurate capture of edge regions; complete internal detection, less internal information loss in the changed areas; low false detection rate, significant noise suppression effect; strong robustness, capable of adapting to complex backgrounds and multi-scale changes.

[0095] As can be seen from the above, although deep learning has made significant progress in the target change detection of optical remote sensing images, its application still faces some challenges. For example, the ability to extract target edge details and minute changes still needs to be improved. For instance, boundary blurring or detail omission often occurs when capturing complex edge information and minute changes. Therefore, how to effectively improve the extraction of change information in the target edge regions of remote sensing images and reduce the missed detection phenomenon is the first key difficulty faced by this field. How to effectively extract the change information inside the changed information, especially inside building targets, and reduce the hole phenomenon is the second key difficulty faced by this field. In addition, the accuracy of the overall target change detection still needs to be improved, which is the third key difficulty faced by this field.

[0096] To solve the above problems, the present invention proposes an optical remote sensing image target change detection method combining deep features. Through double-round deep feature supervision, feature fusion, and attention mechanism modules, it gradually improves the ability to extract edge information and internal details of the target changed areas, and enhances the accuracy and robustness of target change detection. The main innovation points are as follows:

[0097] 1) A dual-depth feature supervised change detection network is proposed, which realizes the multi-scale change feature information fusion through two rounds of depth feature supervision, thereby gradually improving the model's expression of the edge information and internal features of building targets;

[0098] 2) A change feature improved attention mechanism module is proposed, which guides the model to focus on the edge and internal area information of buildings through the change feature module, thereby making up for the incomplete extraction of the edge and internal information of building targets;

[0099] 3) An efficient target change detection network structure is constructed, which adapts to multi-resolution remote sensing images and verifies its excellent performance on two public datasets.

[0100] For ease of description, spatial relative terms such as "above", "over", "on the upper surface", "upper" etc. can be used here to describe the spatial position relationship between a device or feature shown in the figure and other devices or features. It should be understood that the spatial relative terms are intended to include different orientations in use or operation in addition to the orientation described in the figure for the device. For example, if the device in the drawing is inverted, the device described as "above" or "over" other devices or structures will then be positioned "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both the orientations of "above" and "below". The device can also be positioned in other different ways (rotated 90 degrees or in other orientations), and the corresponding explanations for the spatial relative descriptions used here will be made.

[0101] In addition, it should be noted that using words such as "first", "second" etc. to limit components is only for the convenience of distinguishing the corresponding components. Without further statement, the above words have no special meaning, so they cannot be understood as a limitation on the protection scope of the present invention.

[0102] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for detecting target changes in optical remote sensing images combined with depth features, characterized in that: The optical remote sensing image target change detection method combined with depth features includes: Step 1: Extract features from high-resolution remote sensing images X1 and X2 of two phases in the same area; Step 2: performing difference calculation on the features extracted from the high-resolution remote sensing images X1 and X2 to obtain a feature difference map between the features of the two images; Step 3, obtaining multi-scale features of change information through double-round deep feature supervision on the feature difference map; Step 4: Introduce an attention mechanism module into the multi-scale features of the change information to enhance the feature expression of the edge and internal areas of the change information, and output feature Y and a feature map ΔF composed of feature Y (a) ; Step 5: Use the activation function to activate the feature Y output by the attention mechanism module to generate a change map, and convert the feature map ΔF optimized by the attention mechanism into (a) Input to the classifier to generate target change detection results; Step six, based on the target change detection result, use the cross entropy loss function to measure the difference between the model prediction result and the actual result, so as to guide the training and optimization of the target change detection model.

2. The method for detecting target changes in optical remote sensing images combined with depth features according to claim 1, characterized in that: In the step 1, the deep features of the high-resolution remote sensing images X1 and X2 of two phases are extracted through the VGG-16 feature extraction network with shared parameters, where F1 = f(X1; θ), F2 = f(X2; θ), f(; θ) represents the feature extraction network, is the feature map of the high-resolution remote sensing image X1, is the feature map of the high-resolution remote sensing image X2, W represents the width of the feature map, H represents the height of the feature map, D represents the number of channels of the feature map, and θ is the network parameter.

3. The method for detecting target changes in optical remote sensing images combined with depth features according to claim 1, characterized in that: In the step 2, the feature difference map ΔF between the features of the two images may be calculated according to ΔF=|F1-F2|.

4. The method for detecting target changes in optical remote sensing images combined with depth features according to claim 3, characterized in that: The step three specifically includes: Performing a first round of supervision, performing a convolution operation on the feature difference map to extract preliminary change features; The preliminary change features obtained from the first round of supervision are fused with the original features of the two images extracted in step 2 to form the result of the second round of deep supervision, and the multi-scale features of the second round of deep supervision are obtained.

5. The method for detecting target changes in optical remote sensing images combined with depth features according to claim 4, characterized in that: In step 3, the initial change characteristic ΔF (1) According to ΔF (1) =g (1) (ΔF; φ1) is calculated to obtain the multi-scale feature ΔF of the second round of deep supervision. (2) According to ΔF (2) =g (2) (ΔF (2) +F1+F2;φ2), where g (1) represents the first convolution operation, g (2) represents the second convolution operation, φ1 and φ2 represent the parameters in the network, and + represents the concatenation operation.

6. The method for detecting target changes in optical remote sensing images combined with depth features according to claim 5, characterized in that: The step 4 specifically includes: The multi-scale features of the change information ΔF (2) As the input feature X of the attention mechanism; Performing dimension reduction on the input feature X through two different convolution operations to obtain dimension-reduced features Q and K; Reconstruct the reduced-dimensional features Q and K, and transform the reduced-dimensional features Q and K into two-dimensional tensors Q' and K'; Calculate the attention weight matrix A based on the two-dimensional tensors Q' and K'; Use the attention weight matrix A to perform weighted summation on the input feature X to obtain the weighted feature map V; Transform the weighted feature map V to its original size to obtain the weighted feature V'; The input feature X and the weighted feature V' are fused to obtain the output feature Y.

7. The method for detecting target changes in optical remote sensing images combined with depth features according to claim 6, characterized in that: The features Q and K after dimension reduction are based on Calculation is obtained, where Conv1 and Conv2 represent convolution operations, and the features Q and K after dimensionality reduction have channels, C represents the number of channels of the input feature map X, W represents the width of the feature map X, and H represents the height of the feature map X. Represents the set of real numbers.

8. An optical remote sensing image target change detection model combined with depth features, characterized in that: The optical remote sensing image target change detection model combined with depth features uses the optical remote sensing image target change detection method combined with depth features as described in any one of claims 1 to 7 to perform optical remote sensing image target change detection. The optical remote sensing image target change detection model includes a feature extraction module, a dual-round deep feature supervision module, an attention mechanism module, and a classification and change detection module. The feature extraction module is used to extract deep features from the input image based on a convolutional neural network, the dual-round deep feature supervision module is used to realize multi-scale change features through dual-round deep feature fusion and supervised optimization, the attention mechanism module is used to optimize the feature expression of edges and internal areas, and the classification and change detection module is used to combine the feature fusion output to generate the final target change detection result.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes the computer program to implement the steps of the method for detecting target changes in optical remote sensing images combined with depth features according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for detecting target changes in optical remote sensing images combined with depth features as described in any one of claims 1 to 7 are implemented.

Images