A forest change detection method and system based on remote sensing image fusion
Patent Information
- Application Number
- CN202611155305.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-31
- Publication Date
- 2026-10-27
AI Technical Summary
[0038]本发明在修正环节没有直接利用两通道组数据进行相互修正，而是利用二者的变化同步性来判断是真实发生了森林变化还是受到影响产生的误判，然后利用各自时间序列的数据进行修正，这样保证了不同高光谱遥感图像自己所在通道组的数据稳定性和一致性，保证了高光谱遥感图像各自的数据稳定性和一致性，还可以减少因为误变引起的同物异谱或同谱异物导致误检的概率。本发明还在交叉注意力机制环节中，针对双通道组特征彼此进行注意力处理，能够有效融合双通道组各自的光谱信息，从而进一步减少同物异谱、同谱异物导致的识别错误问题。
Smart Images
Figure CN122990476A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of remote sensing image detection technology, and relates to a method and system for detecting forest changes based on remote sensing image fusion. Background Technology
[0002] Forests and other resources are fundamental resources upon which human beings depend for survival. Scientifically determining their changing characteristics and spatial distribution patterns is of great significance for promoting the spatial optimization and protection of forest resources.
[0003] With the development of neural network technology, it has become possible to detect changes in ground features with good results. In particular, hyperspectral point cloud technology can effectively address issues such as different spectra for the same object and different objects with the same spectrum in remote sensing imagery. However, for forest change detection, due to the vast area covered and the limitations of current equipment, hyperspectral point cloud technology is not well-suited for large-scale forest change detection. Furthermore, large-area hyperspectral point cloud data introduces enormous computational demands, requiring sophisticated computing equipment and incurring significant computational overhead, greatly complicating implementation.
[0004] For detecting changes in land type and buildings over large areas, remote sensing image processing is a relatively economical and feasible solution. However, relying solely on remote sensing image processing is not only affected by issues such as different spectra of the same object and different objects with the same spectrum, which can easily lead to false detections, but these false detections are also caused by spectral information and are not easy to detect. In addition, hyperspectral remote sensing images are also easily affected by shadows and changes in illumination, which affects both the accuracy and stability of the detection results. Summary of the Invention
[0005] This invention aims to address the problem that current methods for detecting forest changes based on single remote sensing images are prone to false detections.
[0006] A forest change detection method based on remote sensing image fusion includes:
[0007] Hyperspectral remote sensing images of the corresponding area were acquired using different hyperspectral remote sensing image acquisition devices. and , j represents the data corresponding to time j; and Perform spatial alignment and normalization; then based on or The corrected position is determined by using the other party's changed position as a reference, and based on the corrected position, , Correct based on its own sequence data;
[0008] The two time phases Recorded as and ,Will Image blocks are divided to obtain multiple first time-phase blocks. Using the same method Multiple second-phase blocks were obtained by partitioning. ;right Perform the same processing to obtain the first time phase block. and two phase blocks The corresponding positions of the same division and and and As a set of dual-temporal remote sensing image patches;
[0009] against and Feature maps were obtained by extracting features using encoder branches respectively. and At the same time, targeting and Feature maps were obtained by extracting features using segmentation network branches respectively. and ;Will , and , Adjust to the same resolution, and and The first phase fusion feature is obtained by superimposing it along the channel dimension. ,Will and The second temporal fusion feature is obtained by superimposing it along the channel dimension. ;against and Conduct and and The same processing yields the corresponding first-phase fusion features. Second phase fusion features ;
[0010] against and Cross-attention mechanism is used to obtain cross-fusion features At the same time, targeting and Cross-attention mechanism is used to obtain cross-fusion features ;Will and The first phase feature is obtained by superimposing it along the channel dimension. ;against and Cross-fusion features were obtained using the same method. and cross-fusion features The second temporal features are then superimposed on the channel dimension. ;
[0011] Based on the characteristics of the first time phase Second phase characteristics Obtain the characteristics of change Ultimately, based on the characteristics of change Identify areas for forest change monitoring.
[0012] Furthermore, based on spatiotemporal alignment or The process of determining the correction position by using the other party as a reference for the changed position includes:
[0013] Images at J consecutive time points Determine the pixel value at position i. If the rate of change of pixel value between two adjacent time points is greater than a set change threshold, then record the time j1 when the change occurs; this applies to images at J consecutive time points. Perform the same process and record the moment j2 when the change occurs;
[0014] If there is a record j2 with an alignment time in j1, times j1 and j2 are not used as correction positions; otherwise, times j1 and j2 are used as correction positions.
[0015] Furthermore, based on the corrected position, , The specific process of correcting based on its own sequence data includes:
[0016] When using time j1 as the correction position, the pixel value at the next time step is corrected using the pixel value at time j1-1. Correction Using time j2 as the correction position, the pixel value at time j2-1 is used to correct the pixel value at the next time step. Correction .
[0017] Furthermore, the aforementioned targeting and Feature maps were obtained by extracting features using encoder branches respectively. and The process includes:
[0018] against and First, band selection is performed, and pseudo-color images are synthesized based on the selected band pairs. Then, an encoder is used to extract features to obtain feature maps. and .
[0019] Furthermore, the encoder uses the YOLO V9 backbone network, and the segmentation network uses the DeepLabV3+ model.
[0020] Furthermore, cross-fusion characteristics were obtained. and The specific process includes:
[0021] against and ,based on The Q value is obtained based on The output obtained through cross-attention is the K and V values. ;
[0022] against and ,based on The Q value is obtained based on The output obtained through cross-attention is the K and V values. .
[0023] A forest change detection system based on remote sensing image fusion includes:
[0024] Preprocessing unit: Acquires hyperspectral remote sensing images of the corresponding area using different hyperspectral remote sensing image acquisition devices. and , j represents the data corresponding to time j; and Perform spatial alignment and normalization.
[0025] Correction unit: based on spatiotemporal pair alignment or The corrected position is determined by using the other party's changed position as a reference, and based on the corrected position, , Correct based on its own sequence data;
[0026] Phase block partitioning unit: dividing two consecutive phases Recorded as and ,Will Image blocks are divided to obtain multiple first time-phase blocks. Using the same method Multiple second-phase blocks were obtained by partitioning. ;right Perform the same processing to obtain the first time phase block. and two phase blocks The corresponding positions of the same division and and and As a set of dual-temporal remote sensing image patches;
[0027] Feature extraction and fusion unit: targeting and Feature maps were obtained by extracting features using encoder branches respectively. and At the same time, targeting and Feature maps were obtained by extracting features using segmentation network branches respectively. and ;Will , and , Adjust to the same resolution, and and The first phase fusion feature is obtained by superimposing it along the channel dimension. ,Will and The second temporal fusion feature is obtained by superimposing it along the channel dimension. ;against and Conduct and and The same processing yields the corresponding first-phase fusion features. Second phase fusion features ;
[0028] Cross-attention mechanism fusion unit: targeting and Cross-attention mechanism is used to obtain cross-fusion features At the same time, targeting and Cross-attention mechanism is used to obtain cross-fusion features ;Will and The first phase feature is obtained by superimposing it along the channel dimension. ;against and Cross-fusion features were obtained using the same method. and cross-fusion features The second temporal features are then superimposed on the channel dimension. ;
[0029] Change detection region detection unit: based on first phase features Second phase characteristics Obtain the characteristics of change Ultimately, based on the characteristics of change Identify areas for forest change monitoring.
[0030] Furthermore, based on spatiotemporal alignment or The process of determining the correction position by using the other party as a reference for the changed position includes:
[0031] Images at J consecutive time points Determine the pixel value at position i. If the rate of change of pixel value between two adjacent time points is greater than a set change threshold, then record the time j1 when the change occurs; this applies to images at J consecutive time points. Perform the same process and record the moment j2 when the change occurs;
[0032] If there is a record j2 with an alignment time in j1, times j1 and j2 are not used as correction positions; otherwise, times j1 and j2 are used as correction positions.
[0033] Furthermore, based on the corrected position, , The specific process of correcting based on its own sequence data includes:
[0034] When using time j1 as the correction position, the pixel value at the next time step is corrected using the pixel value at time j1-1. Correction Using time j2 as the correction position, the pixel value at time j2-1 is used to correct the pixel value at the next time step. Correction .
[0035] Furthermore, the aforementioned targeting and Feature maps were obtained by extracting features using encoder branches respectively. and The process includes:
[0036] against and First, band selection is performed, and pseudo-color images are synthesized based on the selected band pairs. Then, an encoder is used to extract features to obtain feature maps. and .
[0037] Beneficial effects:
[0038] This invention does not directly use the data from the two channel groups for mutual correction in the correction stage. Instead, it utilizes the synchronicity of changes in both groups to determine whether a genuine forest change has occurred or a misjudgment is due to influence. Then, it uses the time-series data of each group for correction. This ensures the stability and consistency of the data within each channel group of different hyperspectral remote sensing images, and also reduces the probability of false detections caused by erroneous changes leading to different spectra of the same object or different objects within the same spectrum. Furthermore, in the cross-attention mechanism stage, this invention performs attention processing on the features of the two channel groups, effectively fusing the spectral information of each channel group, thereby further reducing identification errors caused by different spectra of the same object or different objects within the same spectrum.
[0039] Furthermore, this invention also captures texture-focused information through an encoder. This process can select features suitable for forest detection based on prior knowledge, which can effectively improve the targeting and effectiveness of forest change detection. On this basis, this invention also sets up a segmentation network branch to extract features, which can expand the "expression space" of features at different levels as much as possible, thereby better cooperating with the texture extraction branch and improving the detection effect. Attached Figure Description
[0040] Figure 1 for and A schematic diagram illustrating the cross-attention mechanism for obtaining cross-fusion features;
[0041] Figure 2 for and A schematic diagram illustrating the cross-attention mechanism for obtaining cross-fusion features;
[0042] Figure 3 This is a diagram showing the detection results of an embodiment of the present invention. Detailed Implementation
[0043] Specific Implementation Method 1: This implementation method is a forest change detection method based on remote sensing image fusion, including:
[0044] S1. For the forest detection area, acquire hyperspectral remote sensing images of the corresponding area using different hyperspectral remote sensing image acquisition devices. and , j represents the data corresponding to time j; based on geographic information technology, and Spatial alignment and normalization are performed. Hyperspectral remote sensing images acquired by two different hyperspectral remote sensing acquisition devices are used as data for two channels. To distinguish them from the "channels" in the feature map, the hyperspectral remote sensing images acquired by the two different devices are referred to as hyperspectral remote sensing images of two channel groups.
[0045] When acquiring hyperspectral remote sensing images of a forest detection area, if the forest detection area is too large, it is not necessary to acquire hyperspectral remote sensing images covering the entire area at once. Instead, images can be acquired gradually using a sliding window method, based on the actual area and data processing volume. This is as long as subsequent processing can effectively mitigate the impact of the forest's size and size. and Spatial alignment / geometric registration can then be performed to resolve the issue. It should also be noted that: and Spatial alignment and normalization processes can be performed using existing technologies, and will not be described in detail here. In addition to the above-mentioned processes, this invention may also include other processes, such as radiometric correction, which can also be implemented using existing technologies, and will not be described in detail here either.
[0046] S2, based on spatiotemporal alignment or The corrected position is determined by using the other party's changed position as a reference, and based on the corrected position, , Correct based on its own sequence data:
[0047] Images at J consecutive time points Determine the pixel value at position i. Whether a significant change has occurred at J consecutive time points, i.e. whether the rate of change of pixel value between two adjacent time points is greater than a set change threshold, which is set to 5% in this embodiment; if it is greater than the change threshold, the time j1 at which the change occurs is recorded.
[0048] Images at J consecutive time points Perform the same process and record the moment j2 when the change occurs.
[0049] If j1 has a record j2 at the same time, meaning that at a certain time j, both j1 and j2 were detected and recorded (in fact, j1 = j2), then it indicates that synchronous changes occurred within the "dual-channel group (also dual-view)". In this case, the pixel value change is highly likely to be a real change in ground features, and no correction is needed for each pixel. Otherwise (j1 does not have a record j2 at the same time, or j2 does not have a record j1 at the same time), it may be due to the influence of different meteorological factors within their respective fields of view. At this point, the high probability is not a real forest change, therefore it is necessary to... and Make corrections; taking time j1 as the correction position, use the pixel value at time j1-1 to correct the pixel value at the next time step, which can be expressed as using... Correction Using time j2 as the correction position, and using the pixel value at time j2-1 to correct the pixel value at the next time step, can be expressed as: Correction .
[0050] It should be noted that this does not directly utilize... and The data are cross-corrected, relying solely on the synchronicity of changes between the two images to determine whether forest changes have actually occurred or are misjudged due to influence. Corrections are then made using their respective time-series data. This ensures the stability and consistency of data within the same channel group of different hyperspectral remote sensing images. While this method cannot completely resolve the issue of different spectra for the same object or different objects within the same spectrum, it guarantees the stability and consistency of data within each hyperspectral remote sensing image. Data analysis and verification have shown that this can, to some extent, avoid the problem of different spectra for the same object or different objects within the same spectrum caused by erroneous changes. Based on this, due to the hyperspectral remote sensing images... and Each feature is more consistent and accurate, which also ensures the credibility and accuracy of subsequent feature fusion, thereby improving the accuracy of subsequent forest change detection.
[0051] S3, Regarding the revised and Extract dual-temporal remote sensing image patches separately:
[0052] The two time phases Recorded as and The superscript 1 indicates the first channel group, i.e. In the corresponding image, subscripts 1 and 2 represent two time phases; Image blocks are divided to obtain multiple first time-phase blocks. Using the same method Multiple second-phase blocks were obtained by partitioning. ; to divide the corresponding positions and As a pair of dual-temporal remote sensing image patches.
[0053] right Perform the same processing to obtain the first time phase block. and two phase blocks The superscript 2 indicates the second channel group; the corresponding channels at the same division position are... and As a pair of dual-temporal remote sensing image patches, and The corresponding pair of dual-temporal remote sensing image patches and the same location and Matching the corresponding pair of dual-temporal remote sensing image patches yields a set of dual-temporal remote sensing image patches.
[0054] S4, targeting and Feature maps were obtained by extracting features using encoder branches respectively. and .
[0055] To better preserve the spatial contours and texture features of objects, this invention sets up a texture extraction branch; targeting and First, band selection is performed. This band selection process can be determined based on prior information, but it should be noted that... and The selected wavebands are the same. Then, pseudo-color images are synthesized based on the selected wave pairs, and then feature maps are obtained by extracting features using the YOLOv9 backbone network. and It should be noted that this implementation uses the YOLOv9 backbone network to extract features, which can better preserve... and The overall texture features are presented, but this invention is not limited to the backbone network of YOLO V9. In fact, other feature extraction networks can also be used, as long as the detection effect can be guaranteed, they can be used for the texture extraction branch.
[0056] At the same time, targeting and Feature maps were obtained by extracting features using segmentation network branches respectively. and .
[0057] This invention sets up segmentation network branches, thereby enabling subsequent processing of feature maps. and The fusion method preserves more detailed information and improves the model's detection capability. In this embodiment, the segmentation network uses the DeepLabV3+ model. This model can better classify each pixel for segmentation, better preserve detailed features, and better cooperate with the feature maps extracted by the texture extraction branch to improve change detection accuracy. It should be noted that this invention is not limited to the DeepLabV3+ model; other segmentation networks can also be used. It is worth noting that networks that focus on pixel segmentation should be selected as much as possible to expand the "expression space" of features at different levels, thereby better cooperating with the texture extraction branch.
[0058] Will , and , Adjust to the same resolution, and and The first phase fusion feature is obtained by superimposing it along the channel dimension. ,Will and The second temporal fusion feature is obtained by superimposing it along the channel dimension. .
[0059] S5, targeting and Conduct and and The same processing yields the corresponding first-phase fusion features. Second phase fusion features ;
[0060] S6, targeting and Cross-attention mechanism is used to obtain cross-fusion features At the same time, targeting and Cross-attention mechanism is used to obtain cross-fusion features ;Will and The first phase feature is obtained by superimposing it along the channel dimension. .
[0061] against and Cross-attention mechanism is used to obtain cross-fusion features During the process, such as Figure 1 As shown, based on The Q value is obtained based on The output obtained through cross-attention is the K and V values. ; and targeting and Cross-attention mechanism is used to obtain cross-fusion features During the process, such as Figure 2 As shown, based on The Q value is obtained based on The output obtained through cross-attention is the K and V values. .
[0062] By employing a cross-attention mechanism on the first temporal fusion feature map of the dual-channel groups (corresponding to different sensors), the feature information expressed from the first channel group to the second channel group and from the second channel group to the first channel group can be fully extracted. This effectively fuses the spectral information of each dual-channel group, thereby reducing the identification errors caused by different spectra of the same object or different objects with the same spectrum.
[0063] S7, targeting and Cross-fusion features are obtained in the same manner as in step S6. and cross-fusion features The second temporal features are then superimposed on the channel dimension. .
[0064] S8, Based on the characteristics of the first time phase Second phase characteristics Obtain the characteristics of change Ultimately, based on the characteristics of change Identify areas for forest change monitoring.
[0065] The detection effect of the present invention is as follows Figure 3 As shown, the first row is the original image before transformation, the second row is the original image after transformation, and the third row is the detection effect of this invention. The area circled in red in the figure represents the forest change area, and the yellow area represents the corresponding non-forest change area. Since this invention detects forest change, the band selection is based on prior information about the forest. Therefore, it detects forest change areas in the third column, while other non-forest types can be selectively ignored. Thus, this invention can detect forest changes according to actual needs, resulting in better detection performance for forest changes.
[0066] The solution of this invention is used for forest change detection. In fact, this invention can also be used for land cover change detection in other fields, such as large-area farmland detection. However, it should be noted that the band selection needs to be adjusted according to the corresponding prior information. Specific Implementation Method Two:
[0068] This embodiment is a forest change detection system based on remote sensing image fusion. It is a system corresponding to a forest change detection method based on remote sensing image fusion, and it mainly includes:
[0069] Preprocessing unit: Acquires hyperspectral remote sensing images of the corresponding area using different hyperspectral remote sensing image acquisition devices. and , j represents the data corresponding to time j; and Perform spatial alignment and normalization.
[0070] Correction unit: based on spatiotemporal pair alignment or The corrected position is determined by using the other party's changed position as a reference, and based on the corrected position, , Correction is made based on its own sequence data.
[0071] Among them, based on spatiotemporal alignment or The process of determining the correction position by using the other party as a reference for the changed position includes:
[0072] Images at J consecutive time points Determine the pixel value at position i. If the rate of change of pixel value between two adjacent time points is greater than a set change threshold, then record the time j1 when the change occurs; this applies to images at J consecutive time points. Perform the same process and record the moment j2 when the change occurs;
[0073] If there is a record j2 with an alignment time in j1, times j1 and j2 are not used as correction positions; otherwise, times j1 and j2 are used as correction positions.
[0074] Based on the corrected position, , The specific process of correcting based on its own sequence data includes:
[0075] When using time j1 as the correction position, the pixel value at the next time step is corrected using the pixel value at time j1-1. Correction Using time j2 as the correction position, the pixel value at time j2-1 is used to correct the pixel value at the next time step. Correction .
[0076] Phase block partitioning unit: dividing two consecutive phases Recorded as and ,Will Image blocks are divided to obtain multiple first time-phase blocks. Using the same method Multiple second-phase blocks were obtained by partitioning. ;right Perform the same processing to obtain the first time phase block. and two phase blocks The corresponding positions of the same division and and and As a set of dual-temporal remote sensing image patches.
[0077] Feature extraction and fusion unit: targeting and Feature maps were obtained by extracting features using encoder branches respectively. and At the same time, targeting and Feature maps were obtained by extracting features using segmentation network branches respectively. and ;Will , and , Adjust to the same resolution, and and The first phase fusion feature is obtained by superimposing it along the channel dimension. ,Will and The second temporal fusion feature is obtained by superimposing it along the channel dimension. ;against and Conduct and and The same processing yields the corresponding first-phase fusion features. Second phase fusion features .
[0078] Preferably, the encoder uses the YOLO V9 backbone network, and the segmentation network uses the DeepLabV3+ model.
[0079] Among them, targeting and Feature maps were obtained by extracting features using encoder branches respectively. and The process includes:
[0080] against and First, band selection is performed, and pseudo-color images are synthesized based on the selected band pairs. Then, an encoder is used to extract features to obtain feature maps. and .
[0081] Cross-attention mechanism fusion unit: targeting and Cross-attention mechanism is used to obtain cross-fusion features At the same time, targeting and Cross-attention mechanism is used to obtain cross-fusion features ;Will and The first phase feature is obtained by superimposing it along the channel dimension. ;against and Cross-fusion features were obtained using the same method. and cross-fusion features The second temporal features are then superimposed on the channel dimension. ;
[0082] The cross-fusion characteristics were obtained. and The specific process includes:
[0083] against and ,based on The Q value is obtained based on The output obtained through cross-attention is the K and V values. ;
[0084] against and ,based on The Q value is obtained based on The output obtained through cross-attention is the K and V values. .
[0085] Change detection region detection unit: based on first phase features Second phase characteristics Obtain the characteristics of change Ultimately, based on the characteristics of change Identify areas for forest change monitoring.
[0086] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A method for detecting forest change based on remote sensing image fusion, characterized in that, include: Hyperspectral remote sensing images of the corresponding area were acquired using different hyperspectral remote sensing image acquisition devices. and , j represents the data corresponding to time j; and Perform spatial alignment and normalization; then based on or The corrected position is determined by using the other party's changed position as a reference, and based on the corrected position, , Correct based on its own sequence data; The two time phases Recorded as and ,Will Image blocks are divided to obtain multiple first time-phase blocks. Using the same method Multiple second-phase blocks were obtained by partitioning. ;right Perform the same processing to obtain the first time phase block. and two phase blocks The corresponding positions of the same division and and and As a set of dual-temporal remote sensing image patches; against and Feature maps were obtained by extracting features using encoder branches respectively. and At the same time, targeting and Feature maps were obtained by extracting features using segmentation network branches respectively. and ;Will , and , Adjust to the same resolution, and and The first phase fusion feature is obtained by superimposing it along the channel dimension. ,Will and The second temporal fusion feature is obtained by superimposing it along the channel dimension. ;against and Conduct and and The same processing yields the corresponding first-phase fusion features. Second phase fusion features ; against and Cross-attention mechanism is used to obtain cross-fusion features At the same time, targeting and Cross-attention mechanism is used to obtain cross-fusion features ;Will and The first phase feature is obtained by superimposing it along the channel dimension. ;against and Cross-fusion features were obtained using the same method. and cross-fusion features The second temporal features are then superimposed on the channel dimension. ; Based on the characteristics of the first time phase Second phase characteristics Obtain the characteristics of change Ultimately, based on the characteristics of change Identify areas for forest change monitoring.
2. The forest change detection method based on remote sensing image fusion according to claim 1, characterized in that, Based on spatiotemporal pair alignment or The process of determining the correction position by using the other party as a reference for the changed position includes: Images at J consecutive time points Determine the pixel value at position i. If the rate of change of pixel value between two adjacent time points is greater than a set change threshold, then record the time j1 when the change occurs; this applies to images at J consecutive time points. Perform the same process and record the moment j2 when the change occurs; If there is a record j2 with an alignment time in j1, times j1 and j2 are not used as correction positions; otherwise, times j1 and j2 are used as correction positions.
3. The forest change detection method based on remote sensing image fusion according to claim 2, characterized in that, Based on the corrected position, , The specific process of correcting based on its own sequence data includes: When using time j1 as the correction position, the pixel value at the next time step is corrected using the pixel value at time j1-1. Correction Using time j2 as the correction position, the pixel value at time j2-1 is used to correct the pixel value at the next time step. Correction .
4. The forest change detection method based on remote sensing image fusion according to claim 1, characterized in that, The aforementioned targeting and Feature maps were obtained by extracting features using encoder branches respectively. and The process includes: against and First, band selection is performed, and pseudo-color images are synthesized based on the selected band pairs. Then, an encoder is used to extract features to obtain feature maps. and .
5. The forest change detection method based on remote sensing image fusion according to claim 4, characterized in that, The encoder uses the YOLO V9 backbone network, and the segmentation network uses the DeepLabV3+ model.
6. The forest change detection method based on remote sensing image fusion according to claim 1, characterized in that, Obtain cross-fusion features and The specific process includes: against and ,based on The Q value is obtained based on The output obtained through cross-attention is the K and V values. ; against and ,based on The Q value is obtained based on The output obtained through cross-attention is the K and V values. .
7. A forest change detection system based on remote sensing image fusion, characterized in that, include: Preprocessing unit: Acquires hyperspectral remote sensing images of the corresponding area using different hyperspectral remote sensing image acquisition devices. and , j represents the data corresponding to time j; and Perform spatial alignment and normalization. Correction unit: based on spatiotemporal pair alignment or The corrected position is determined by using the other party's changed position as a reference, and based on the corrected position, , Correct based on its own sequence data; Phase block partitioning unit: dividing two consecutive phases Recorded as and ,Will Image blocks are divided to obtain multiple first time-phase blocks. Using the same method Multiple second-phase blocks were obtained by partitioning. ;right Perform the same processing to obtain the first time phase block. and two phase blocks The corresponding positions of the same division and and and As a set of dual-temporal remote sensing image patches; Feature extraction and fusion unit: targeting and Feature maps were obtained by extracting features using encoder branches respectively. and At the same time, targeting and Feature maps were obtained by extracting features using segmentation network branches respectively. and ;Will , and , Adjust to the same resolution, and and The first phase fusion feature is obtained by superimposing it along the channel dimension. ,Will and The second temporal fusion feature is obtained by superimposing it along the channel dimension. ;against and Conduct and and The same processing yields the corresponding first-phase fusion features. Second phase fusion features ; Cross-attention mechanism fusion unit: targeting and Cross-attention mechanism is used to obtain cross-fusion features At the same time, targeting and Cross-attention mechanism is used to obtain cross-fusion features ;Will and The first phase feature is obtained by superimposing it along the channel dimension. ;against and Cross-fusion features were obtained using the same method. and cross-fusion features The second temporal features are then superimposed on the channel dimension. ; Change detection region detection unit: based on first phase features Second phase characteristics Obtain the characteristics of change Ultimately, based on the characteristics of change Identify areas for forest change monitoring.
8. A forest change detection system based on remote sensing image fusion according to claim 7, characterized in that, Based on spatiotemporal pair alignment or The process of determining the correction position by using the other party as a reference for the changed position includes: Images at J consecutive time points Determine the pixel value at position i. If the rate of change of pixel value between two adjacent time points is greater than a set change threshold, then record the time j1 when the change occurs; this applies to images at J consecutive time points. Perform the same process and record the moment j2 when the change occurs; If there is a record j2 with an alignment time in j1, times j1 and j2 are not used as correction positions; otherwise, times j1 and j2 are used as correction positions.
9. A forest change detection system based on remote sensing image fusion according to claim 8, characterized in that, Based on the corrected position, , The specific process of correcting based on its own sequence data includes: When using time j1 as the correction position, the pixel value at the next time step is corrected using the pixel value at time j1-1. Correction Using time j2 as the correction position, the pixel value at time j2-1 is used to correct the pixel value at the next time step. Correction .
10. A forest change detection system based on remote sensing image fusion according to claim 7, characterized in that, The aforementioned targeting and Feature maps were obtained by extracting features using encoder branches respectively. and The process includes: against and First, band selection is performed, and pseudo-color images are synthesized based on the selected band pairs. Then, an encoder is used to extract features to obtain feature maps. and .