A binary detection method for remote sensing images by integrating feature differences and twin structures

Abstract

The present invention relates to a remote sensing image binary detection method that integrates feature differences and twin structures. The method includes: acquiring images; dividing regions; extracting features; determining changes; screening construction areas; adjusting thresholds; and binary detection fusion. The present invention ensures sampling uniformity and multi-scale adaptability through grid division. The coverage change rate and bare land change correspond to the natural coupling of vegetation-bare soil conversion. The registration residual reflects the coupling effect of image alignment error and surface disturbance. The length and curvature of the mechanical trajectory accurately depict the operating range and path regularity. The dynamic threshold adjustment responds to the construction concentration based on the coupling of regional aggregation density and historical frequency. Rule screening and twin network cross-validation combine data-driven characterization, effectively solving the problem of inaccurate identification of construction areas and high misjudgment rate caused by image registration error, illumination change, and spectral similarity of ground objects, which is difficult to accurately distinguish.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image change detection, and in particular to a remote sensing image binary detection method that integrates feature differences and twin structures. Background Art

[0002] Urban development zones often face a complex interplay of diverse land use changes: linear infrastructure construction such as road paving and pipeline installation, point-based construction of buildings and supporting facilities, and surface disturbances from large-scale site leveling and earthwork operations. Furthermore, temporal fluctuations in climate, seasons, and vegetation conditions further challenge the accurate identification of construction disturbance signals. While high-resolution remote sensing imaging technology can provide millisecond-level spatiotemporal observations, issues such as image registration errors, varying lighting conditions, and spectral similarity of ground objects often make it difficult for a single indicator to reliably distinguish between real construction activity and seasonal or sporadic disturbances. Accurately capturing weak local construction signals while ensuring real-time and high-frequency response for wide-area monitoring has become a critical issue facing urban planning, environmental monitoring, and smart engineering management.

[0003] Existing remote sensing change detection methods still face numerous limitations when applied to the complex context of urban construction. For one thing, they often employ static thresholds or fixed templates for regional delineation and change identification, making them ill-suited to the highly uneven spatial distribution and intensity of change within construction zones. Furthermore, they lack the ability to comprehensively characterize spatial and temporal evolutionary patterns, making them prone to misinterpreting short-term disturbances, seasonal vegetation changes, or image registration errors as construction activity. This can lead to false positives and missed detections in dynamic and complex urban development zones. Furthermore, traditional methods are inadequately responsive to fine-grained surface activity and often lack the accuracy to accurately detect subtle signals such as latent disturbances during the initial construction phase, mechanical micro-movements, or brief exposures of bare ground. This limits their practicality and timeliness in real-world urban supervision and dynamic development monitoring. Furthermore, existing frameworks have limited ability to integrate multi-source features, and change assessments often rely on single indicators or crude combinations, making it difficult to fundamentally disentangle differences in ground features caused by natural disturbances and human construction. Parameter setting and model tuning also suffer from a strong dependence on regional characteristics and weak generalization, resulting in low deployment efficiency and poorly stable results. Furthermore, most methods are relatively crude in their treatment of localized changes and lack a sensitive understanding of the continuity, density, and trends of change signals across space. This makes it difficult to accurately capture construction signs of practical engineering significance but limited spatial coverage in large-scale imagery. These shortcomings collectively hinder the further application of remote sensing change detection technology in complex urban development zones. Summary of the Invention

[0004] To this end, the present invention provides a remote sensing image binary detection method that integrates feature differences and twin structures, which is used to overcome the problems in the existing technology of inaccurate construction area identification and high misjudgment rate caused by image registration errors, illumination changes and spectral similarity of ground objects, which are difficult to accurately distinguish.

[0005] To achieve the above objectives, the present invention provides a remote sensing image binary detection method that integrates feature differences and twin structures, comprising:

[0006] Real-time acquisition of dual-temporal remote sensing images of the detection area within the urban development zone;

[0007] Dividing the dual-temporal remote sensing image into a plurality of unit areas;

[0008] Extracting the surface coverage change rate, image registration residual mean, displacement track length and displacement track curvature of the construction machinery, and bare land area change rate within each unit area;

[0009] determining, based on the surface cover change rate, the bare land area change rate, and a preset change threshold, that the surface structure in the unit area has changed, and forming a plurality of temporary areas;

[0010] Selecting a number of new construction areas based on the displacement trajectory length and the displacement trajectory curvature of each temporary area;

[0011] Determine a number of significant change areas according to the image registration residual mean and the coverage change rate of each newly constructed area;

[0012] Adjusting the preset change threshold according to the position coordinates of each of the significantly changed areas within a preset adjustment period to obtain an adjusted change threshold;

[0013] Inputting the dual-temporal remote sensing image into a preset twin residual network model to obtain a number of ideal change areas;

[0014] A plurality of target areas are determined according to the ideal change area and the significant change area obtained based on the adjusted change threshold.

[0015] Furthermore, the process of determining that the surface structure in the unit area has changed based on the surface cover change rate, the bare land area change rate, and a preset change threshold, and forming a plurality of temporary areas includes:

[0016] Determining the change correlation based on all of the surface cover change rates and all of the bare land area change rates within a preset temporary determined time period;

[0017] According to the comparison result of the change correlation and the preset change threshold, it is determined that the surface structure in the unit area has changed, and a plurality of temporary areas are formed.

[0018] Furthermore, determining the change correlation coefficient according to all the surface cover change rates and all the bare land area change rates within the preset temporary determination time period includes:

[0019] Calculating the standard deviation of all the surface cover change rates between the initial moment and each moment within the preset temporarily determined time period to obtain a plurality of cover change fluctuation values;

[0020] Calculating the standard deviation of all the bare land area change rates between the initial moment and each moment within the preset temporarily determined time period to obtain a plurality of bare land change fluctuation values;

[0021] A change correlation coefficient is determined based on all of the coverage change fluctuation values ​​and all of the bare land change fluctuation values.

[0022] Furthermore, the process of selecting a plurality of new construction areas according to the displacement trajectory length and the displacement trajectory curvature of each temporary area includes:

[0023] Screening out a number of areas of interest based on a comparison result of the displacement trajectory length and a preset length threshold;

[0024] A number of new construction areas are selected based on the curvature of the displacement trajectory within all the areas of interest.

[0025] Furthermore, the process of selecting a number of new construction areas based on the curvature of the displacement trajectory in all the areas of interest includes:

[0026] Determining the change consistency according to the curvature of the displacement trajectory of any two of the regions of interest;

[0027] According to the comparison result of the change consistency and the preset consistency threshold, it is determined that the two focus areas are both the new construction areas, so as to screen out several new construction areas.

[0028] Furthermore, the process of determining the change consistency according to the curvature of the displacement trajectory of any two of the regions of interest includes:

[0029] Obtaining the displacement trajectory curvature of each of the focus areas within a preset screening time to obtain a trajectory curvature sequence;

[0030] The Pearson correlation coefficient of the trajectory curvature sequences of any two of the regions of interest is calculated to determine the change consistency.

[0031] Furthermore, the process of determining a number of significant change areas according to the image registration residual mean and the coverage change rate of each newly constructed area includes:

[0032] Calculating the standard deviation of the mean of all image registration residuals from the initial moment to each moment within a preset significant determination time length to obtain a plurality of residual fluctuation values;

[0033] Calculating the time derivatives of all the coverage change rates within the preset significant determination time length to obtain a coverage change coefficient;

[0034] determining a joint variation index according to the residual fluctuation value and the coverage variation coefficient;

[0035] Based on the comparison result of the combined change index and the preset standard index, the newly constructed area is determined to be the significant change area, so as to determine several significant change areas.

[0036] Furthermore, the preset change threshold is adjusted according to the position coordinates of each of the significantly changed areas within the preset adjustment period, and the process of obtaining the adjusted change threshold includes:

[0037] Obtaining the center position coordinates of all the regions with significant changes within the preset adjustment period to form a position coordinate set;

[0038] Performing clustering processing on the position coordinate set, and extracting a number of spatially varying cluster areas based on a preset spatial cluster radius and a preset minimum cluster number threshold;

[0039] Based on the number and spatial distribution of significant change areas within each of the change clustering areas, the corresponding change spatial density value is calculated to obtain a spatial density sequence;

[0040] The preset change threshold is adjusted based on the spatial density sequence and the change frequency information within a preset historical adjustment period to obtain the adjusted change threshold.

[0041] Furthermore, the preset change threshold is adjusted based on the spatial density sequence and the change frequency information within a preset historical adjustment period. The process of obtaining the adjusted change threshold includes:

[0042] Extracting the number of times the significant change region is determined in each of the change cluster regions to obtain a historical frequency sequence;

[0043] Generating a local adjustment coefficient based on the spatial density sequence, the preset density weight, the historical frequency sequence, and the preset frequency weight to obtain an adjustment coefficient set;

[0044] Modifying the preset change threshold corresponding to each of the change concentration areas based on the adjustment coefficient set to form a regional adjustment change threshold set;

[0045] Performing spatial interpolation and smoothing processing on the region adjustment change threshold set to obtain the adjustment change threshold.

[0046] Furthermore, the process of determining a plurality of target areas according to the ideal change area and the significant change area obtained based on the adjusted change threshold includes:

[0047] The overlapping areas of all the significant change areas and all the target areas are taken as the target areas to determine a plurality of target areas.

[0048] Compared with the existing technology, the beneficial effect of the present invention is that the uniformity and multi-scale adaptability of sampling are ensured through grid division, and the coverage change rate and bare land change logically correspond to the natural coupling of vegetation-bare soil conversion; the registration residual reflects the coupling effect of image alignment error and surface disturbance, and the mechanical trajectory length and curvature accurately depict the operating range and path regularity, which conforms to the intrinsic relationship between kinematics and engineering execution behavior; the dynamic threshold adjustment responds to the continuity and concentration of construction based on the coupling of regional aggregation density and historical frequency, maintaining the model's adaptability to spatiotemporal heterogeneity; finally, rule screening and twin network cross-validation combine interpretable physical thresholds with powerful data-driven representation, which not only ensures high detection accuracy, but also enhances the robustness and interpretability of small construction disturbances in complex environments.

[0049] Furthermore, by introducing two highly correlated surface disturbance indicators, the surface cover change rate and the bare land area change rate, a change determination mechanism that conforms to the logic of natural geographical processes was constructed. The surface cover change rate reflects the degree of change in land feature types, such as a decrease in vegetation and an increase in hard ground, while the bare land area change rate directly indicates the increasing trend in the degree of human disturbance. There is a highly correlated natural coupling relationship between the two: during construction and excavation processes, the surface cover type typically shifts from high-coverage features such as vegetation to bare land, which in turn causes changes in both indicators. By setting a change threshold, it is possible to adaptively screen out areas that truly show signs of construction activity, effectively filtering out minor changes caused by noise, seasonal differences, or temporary disturbances, and improving the accuracy and robustness of identifying areas of significant change.

[0050] Furthermore, by quantifying the standard deviation fluctuations of the rates of change in land cover and bare land area, we not only capture the synchronous signals of reduced vegetation cover and increased bare soil exposure, but also utilize vectorization and cosine similarity to transform temporal fluctuations into angular similarity metrics in high-dimensional space, accurately reflecting the co-evolutionary characteristics of the two during dynamic changes. The directional analysis based on standard deviation fluctuation sensitivity and cosine similarity not only conforms to the natural coupling of vegetation and bare soil transformations, which indicates that vegetation loss is inevitably accompanied by increased bare soil exposure due to human activities such as construction or excavation, but also mathematically reveals the consistency and strength of the changing trends, avoiding the temporal bias caused by single threshold judgments and effectively improving the reliability and robustness of identifying areas of significant change.

[0051] Furthermore, a displacement length threshold filtering method can eliminate sporadic, short-distance machine movements and focus on high-intensity, large-scale construction activities, in line with the fundamental physical law in engineering that "operation distance and scale are positively correlated." Trajectory curvature analysis, combined with the kinematic principle that "straight or regular curved motion is a typical characteristic of machinery executing a predetermined trajectory," effectively distinguishes the orderly movement of construction machinery along roads or building boundaries from random jitter or detours. The synergistic effect of these two methods not only reflects the inherent coupling relationship between "working condition intensity and path shape" in spatial motion, but also significantly improves the accuracy and robustness of identifying new construction areas.

[0052] Furthermore, the trajectory curvature consistency is used to quantify the morphological similarity of the motion paths of construction machinery in different areas. This is consistent with the principle of "similar curvature indicates similar motion patterns" in kinematics, and is also in line with the law that the round-trip operation paths of machinery in urban construction remain consistent and repeatable. By setting the consistency threshold, occasional jitter and path deviation can be effectively filtered out, thereby enhancing the accuracy and stability of identifying new construction areas.

[0053] Furthermore, by quantifying the synchronous fluctuations of the curvature sequence through the Pearson correlation coefficient, it is possible to capture the consistency of the path morphology of construction machinery in different areas, which is in line with the basic physical principle of "similar curvature means similar motion pattern" in kinematics, and eliminates occasional noise and short-term jitter from a statistical perspective; at the same time, curvature, as a measure of the degree of path bending, together with dynamic parameters such as trajectory length and motion speed, reflects the spatial distribution law of mechanical operations, embodies the natural science logic of "the underlying coupling of path geometry and construction behavior", thereby significantly improving the accuracy, robustness and interpretability of the identification of new construction areas.

[0054] Furthermore, the volatility of the registration error is measured by the standard deviation to reflect the impact of surface structure changes on image alignment; the time derivative is used to capture the instantaneous rate of coverage change, revealing the rapid changes in vegetation or surface caused by construction activities; the two are coupled in a joint change index, which not only conforms to the synchronization enhancement effect, but also follows the natural science logic of the interaction of "error fluctuation-change rate" in physical space, thereby jointly verifying the authenticity of the change signal at the mathematical and physical levels, greatly improving the accuracy and interpretability of identifying areas with significant changes.

[0055] Furthermore, the intensity of localized changes is measured through "spatial density," with the concentration of significant spatially changing areas reflecting the concentration of construction activity. Historical change frequency is also introduced to capture persistent trends, enabling adaptive adjustment of change thresholds. This adjustment logic is based on the underlying positive correlation between "density, frequency, and intensity," meaning that greater spatial change density and higher change frequency indicate a greater intensity of the underlying disturbance source (new construction). This approach improves sensitivity and enhances the accuracy and robustness of change detection results.

[0056] Furthermore, by integrating spatial density (reflecting the concentration of change) with historical frequency (reflecting the persistence of change), an adjustment factor model was constructed that conforms to the spatial-temporal coupling characteristics of natural phenomena. This reflects the scientific logic that higher density and more frequent changes should increase the sensitivity of change detection. At the same time, a combination of normalization and weighting ensures the universality and comparability of local adjustment coefficients at different scales. Spatial interpolation and smoothing further enhance the continuity and stability of the threshold field in geographic space, effectively reducing the impact of local anomalies on overall detection results, and enhancing the model's adaptability to complex feature evolution and detection accuracy.

[0057] Furthermore, a confidence enhancement mechanism is introduced through spatial overlap operations, making the identification of target areas more robust. Significantly changed areas are derived from surface feature change parameters such as the bare land area change rate and the cover change rate. These parameters are driven by surface physical processes and have a clear spatiotemporal logic. Ideally changed areas, on the other hand, are derived from the high-dimensional representation of image residual signals by the twin network, which is sensitive to subtle changes and structural anomalies. The overlapping areas of these two represent the intersection of "significant physical change" and "high-confidence model identification areas." This conforms to the natural science principle of "multi-source information fusion and discrimination" in remote sensing change detection, and embodies the logic of "feature-driven + model-aware." This improves the accuracy and stability of target area identification, and has excellent promotional and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of the remote sensing image binary detection method that integrates feature differences and twin structures in this embodiment;

[0059] Figure 2 This is a logic diagram for determining changes in the surface structure in the unit area in this embodiment;

[0060] Figure 3 This is a logic diagram for determining the area of ​​interest in this embodiment;

[0061] Figure 4 This is the decision logic diagram for determining the new construction area in this embodiment. DETAILED DESCRIPTION

[0062] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention.

[0063] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0064] See also Figure 1 As shown, it is a flow chart of the remote sensing image binary detection method of this embodiment that integrates feature differences and twin structures;

[0065] This embodiment provides a remote sensing image binary detection method that integrates feature differences and twin structures, including:

[0066] Real-time acquisition of dual-temporal remote sensing images of the detection area within the urban development zone;

[0067] Dividing the dual-temporal remote sensing image into a plurality of unit areas;

[0068] Extracting the surface coverage change rate, image registration residual mean, displacement track length and displacement track curvature of the construction machinery, and bare land area change rate within each unit area;

[0069] determining, based on the surface cover change rate, the bare land area change rate, and a preset change threshold, that the surface structure in the unit area has changed, and forming a plurality of temporary areas;

[0070] Selecting a number of new construction areas based on the displacement trajectory length and the displacement trajectory curvature of each temporary area;

[0071] Determine a number of significant change areas according to the image registration residual mean and the coverage change rate of each newly constructed area;

[0072] Adjusting the preset change threshold according to the position coordinates of each of the significantly changed areas within a preset adjustment period to obtain an adjusted change threshold;

[0073] Inputting the dual-temporal remote sensing image into a preset twin residual network model to obtain a number of ideal change areas;

[0074] A plurality of target areas are determined according to the ideal change area and the significant change area obtained based on the adjusted change threshold.

[0075] The process of dividing the dual-temporal remote sensing image into a number of unit areas includes:

[0076] Grid division is performed in the detection area with a preset grid side length to form a number of non-overlapping grid areas, which are divided into a number of unit areas.

[0077] In this embodiment, the surface cover change rate in each unit area is obtained by calculating the differences in indicators such as the Normalized Difference Vegetation Index (NDVI) and surface reflectance in the dual-temporal remote sensing images. The image registration residual mean is calculated using a feature point matching and geometric correction algorithm. After the dual-temporal images are registered, the registration error distribution is statistically analyzed and its regional mean is calculated. The displacement trajectory length and curvature of the construction machinery are extracted by performing motion target detection and tracking on high-resolution remote sensing images of continuous phases in the image sequence, and the total length and curvature change of the trajectory are obtained based on spatial geometry calculations. The bare land area change rate is combined with image classification results and change detection algorithms to identify the area change ratio of the bare land category in each unit area, thereby constructing a multidimensional change feature parameter set for urban construction activity identification.

[0078] The preset grid side length refers to the fixed grid size used when dividing the remote sensing image into unit areas. It depends on the spatial resolution of the image, the target detection scale, and the efficiency requirements of subsequent feature extraction and calculation. It is usually set between 5m and 100m. In this embodiment, it is set to 30m. It can take into account the regional analysis requirements of computational efficiency and change patterns while ensuring the expression of spatial details.

[0079] The preset change threshold is a criterion for determining whether a region in a remote sensing image has undergone significant changes. It depends on the spatial resolution of the image, the type of ground object, the monitoring period, and the characteristic strength of the target change behavior. It is usually set between 0.15 and 0.35 to balance the missed detection rate and the false detection rate. In this embodiment, it is set to 0.25, which can effectively identify real changes in new construction areas in urban development zones, while suppressing pseudo-changes caused by registration errors or temporary interference, thereby improving the accuracy and stability of change detection.

[0080] The preset adjustment period is a time window used to statistically analyze the spatial distribution of areas with significant changes and dynamically update the change threshold. It depends on the frequency of remote sensing image acquisition and the urban construction progress cycle. It is usually set between 30 and 90 days. In this embodiment, it is set to 60 days, which can strike a balance between maintaining the timeliness of threshold response and avoiding short-term noise interference.

[0081] Specifically, the process of inputting the dual-temporal remote sensing image into a preset twin residual network model to obtain several ideal change regions includes:

[0082] The deep and shallow features of the preprocessed unit area are extracted through the preset twin residual network model;

[0083] The deep features are input into the dilated spatial pyramid module for fusion to obtain multi-scale spatial information;

[0084] The multi-scale spatial information is input into a semantic enhancement module embedded in a multi-scale pooling fusion structure, and enhanced semantic features are output by modeling long-range context dependencies;

[0085] The enhanced semantic features of the two-phase are cascaded in the channel dimension and input into a binary change detection module composed of a stack of several residual structures to generate an ideal change region.

[0086] The process of extracting the deep features and shallow features of the preprocessed unit area through the preset Siamese residual network model includes:

[0087] The preset twin residual network model is obtained by improving the public Resnet34 network. An improved twin Resnet34 network with weight sharing is built to obtain the twin residual network, which includes 5 layers, namely Layer0, Layer1, Layer2, Layer3 and Layer4, where Layer0 consists of a 7×7 convolution layer, a BN layer, a Relu nonlinear activation layer and a maximum pooling layer. The structures of Layer1 and Layer2 are consistent with those in the public Resnet34 network, and the stride of all convolution layers in Layer3 and Layer4 is 1.

[0088] The unit area in the enhanced training sample data is used as the input image, input into the built Resnet34 network, the semantic feature map output by Layer2 is used as the shallow feature, and the semantic feature map output by Layer4 is used as the deep feature. The specific process is expressed as follows:

[0089]

[0090] in, Represents shallow features, Represents deep features, Conv7 represents a convolution operation with a convolution kernel size of 7×7, BN represents batch normalization, Relu represents the Relu nonlinear activation function, maxpool represents the maximum pooling operation, x1 represents the unit area of ​​the first phase, x2 represents the unit area of ​​the second phase, F1 represents the feature map of the first phase after Layer0 (i.e., 7x7 convolution, BN, Relu, and maximum pooling), and F2 represents the feature map of the second phase after Layer0 (same operation);

[0091] While maintaining the powerful feature extraction capabilities of ResNet34, the improved twin Resnet34 network has undergone structural optimization for remote sensing change detection tasks, particularly in two aspects: First, parameter sharing through the twin structure ensures that the dual-phase images follow a consistent feature mapping path and convolution kernel response during feature extraction, avoiding the introduction of additional non-semantic differences due to differences in network structure, in line with the basic logic of the "controlled variable method", that is, ensuring that input differences are only caused by changes in the image itself. Second, the convolutional layer stride of Layer3 and Layer4 is set to 1, avoiding the destruction of spatial details by the downsampling operation in the standard ResNet and retaining high-resolution feature maps. This is consistent with the distribution pattern in remote sensing images that "changing target areas are often small-scale and complex in shape."

[0092] The process of inputting the deep features into the dilated spatial pyramid module for fusion to obtain multi-scale spatial information includes:

[0093] Build a feature fusion module. The feature fusion module is a dilated spatial convolution pooling pyramid module consisting of five branches, namely the first branch, the second branch, the third branch, the fourth branch and the fifth branch; the first branch is a global average pooling layer, which consists of a two-dimensional adaptive average pooling layer, a 1×1 convolution, a BN layer and a Relu nonlinear activation layer; the second branch consists of a 3×3 dilated convolution with a dilation factor of 6, a BN layer and a Relu nonlinear activation layer; the third branch consists of a 3×3 dilated convolution with a dilation factor of 12, a BN layer and a Relu nonlinear activation layer; the fourth branch consists of a 3×3 dilated convolution with a dilation factor of 18, a BN layer and a Relu nonlinear activation layer; the fifth branch consists of a 3×3 dilated convolution with a dilation factor of 24, a BN layer and a Relu nonlinear activation layer;

[0094] The deep features output by the feature extraction network are used as the input image and input into the constructed dilated spatial convolution pooling pyramid module. Five features of different scales are extracted through five branches. These are cascaded in dimension one and passed through a 1×1 convolution, a BN layer, a Relu nonlinear activation layer, and a Dropout operation to obtain multi-scale spatial information. The specific process is expressed as follows:

[0095]

[0096] in, represents the bilinear interpolation upsampling operation, Represents the convolution operation with a convolution kernel size of 1×1, They are the dilated convolutions of the second branch, the third branch, the fourth branch, and the fifth branch, respectively. They are the output features of the first branch, the second branch, the third branch, the fourth branch, and the fifth branch respectively. cat is the concat cascade method. It is multi-scale spatial information;

[0097] Deep features are fed into a pre-configured dilated spatial pyramid module for multi-scale spatial information fusion. This module consists of five parallel branches: the first branch uses global average pooling to extract the overall image context, followed by feature compression and nonlinear transformation via 1×1 convolution, batch normalization (BN) layers, and ReLU activation. The remaining four branches employ 3×3 dilated convolutions with dilation factors of 6, 12, 18, and 24, respectively, to extract local semantic information at different scales with varying receptive fields. Each branch is also followed by BN and ReLU layers to maintain numerical stability and nonlinear expressiveness. The features output by these five branches are concatenated along the channel dimension, then subjected to a 1×1 convolution to unify the number of channels. BN and ReLU activation functions, along with dropout, mitigate overfitting, ultimately yield a fused multi-scale spatial information feature map for subsequent semantic enhancement.

[0098] The process of embedding the multi-scale spatial information input into a semantic enhancement module of a multi-scale pooling fusion structure and outputting enhanced semantic features by modeling long-range context dependencies includes:

[0099] In view of the large-scale changes in the target scale in remote sensing images, it is necessary to capture long-range and multi-scale contextual features to give the network a wider receptive field and more robust long-range context modeling capabilities. Therefore, a semantic enhancement module is built. This semantic enhancement module is obtained by embedding a multi-scale pooling fusion module on the basis of the non-local module to fuse channel information with spatial information. The non-local module consists of three branches: Q, K, and V. The number of channels of each branch is adjusted through 1×1 convolution.

[0100] Four pooling kernels of different scales are used for fusion. The specific calculation process is as follows:

[0101]

[0102] in, It is the enhanced semantic features of the two phases output by the semantic enhancement module. The superscript T represents the matrix transposition operation. Represents the matrix dimension transformation operation; MPS represents the multi-scale pooling fusion operation. The specific MPS operation is expressed as follows:

[0103]

[0104] in, It is a mean pooling operation with different step lengths, and Output is the output of the MPS operation;

[0105] Build a binary change detection module, which consists of a residual structure 1 and four residual structures 2 stacked in series. The residual structure 1 consists of a 3×3 convolutional layer, a batch normalization layer, a Relu nonlinear activation layer, and a 1×1 convolutional layer and a batch normalization layer on the shortcut branch. The residual structure 2 consists of a 3×3 convolutional layer, a batch normalization layer, and a Relu nonlinear activation layer.

[0106] The enhanced semantic features of the two phases output by the semantic enhancement module are cascaded on dimension one. The cascaded feature map is then passed to the binary change detection module and upsampled to the size of the input image to obtain a binary change map, i.e., the ideal change area. The specific calculation process is as follows:

[0107]

[0108] in, Four consecutive times operate, represents a convolution operation with a convolution kernel size of 3×3, It is an ideal change area.

[0109] Real-time dual-temporal remote sensing images are acquired within the urban development zone, and the detection area is gridded with a preset grid side length to obtain several non-overlapping unit areas. Then, multi-source features such as the surface cover change rate, bare land area change rate, image registration residual mean, and displacement trajectory length and curvature of construction machinery are extracted in each unit area. Combined with the preset threshold, temporary change areas, new construction areas, and areas with significant changes are screened out layer by layer. The change threshold is dynamically updated within a preset adjustment period based on the spatial coordinates of the significant area. Next, the original dual-temporal image is input into the twin residual network to obtain the ideal change area determined by the network. Finally, the final target change area is comprehensively determined through cross-validation of the ideal area and the threshold-significant area.

[0110] Grid division ensures sampling uniformity and multi-scale adaptability. The coverage change rate and bare land change logically correspond to the natural coupling of vegetation-bare soil conversion; the registration residual reflects the coupling effect of image alignment error and surface disturbance, and the mechanical trajectory length and curvature accurately depict the operating range and path regularity, which conforms to the intrinsic relationship between kinematics and engineering execution behavior; dynamic threshold adjustment responds to the continuity and concentration of construction based on the coupling of regional aggregation density and historical frequency, maintaining the model's adaptability to spatiotemporal heterogeneity; finally, rule screening and twin network cross-validation combine interpretable physical thresholds with powerful data-driven representation, which not only ensures high detection accuracy, but also enhances the robustness and interpretability of small construction disturbances in complex environments.

[0111] Please continue reading Figure 2 As shown, it is a logic diagram for determining changes in the surface structure in the unit area in this embodiment;

[0112] The process of determining that the surface structure in the unit area has changed according to the surface cover change rate, the bare land area change rate, and a preset change threshold, and forming a plurality of temporary areas includes:

[0113] Determining the change correlation based on all of the surface cover change rates and all of the bare land area change rates within a preset temporary determined time period;

[0114] When the change correlation is greater than the preset change threshold, it is determined that the surface structure in the unit area has changed, and a number of temporary areas are formed.

[0115] The preset temporary determination time period refers to the time window used to count the surface cover change rate and the bare land area change rate to calculate the change correlation. It depends on the acquisition frequency of remote sensing images and the typical cycle of urban construction activities. It is usually set between 7 days and 30 days. In this embodiment, it is set to 14 days, which can balance the continuity of the change trend and the timeliness of the detection response.

[0116] First, within a preset temporary time period, the surface cover change rate and bare land area change rate corresponding to each unit area are extracted. Through statistical analysis of the two, the correlation between the changes is calculated. The correlation, using the Pearson correlation coefficient as an indicator, quantifies the linkage between the two within the same time period, reflecting their coordinated change trend. When the correlation is higher than the preset change threshold, it indicates that the unit area has undergone consistent changes in both surface use and bare land expansion. Therefore, it can be identified as a temporary area and included in the temporary area set for further analysis and screening.

[0117] By introducing two highly correlated surface disturbance indicators, the land cover change rate and the bare land area change rate, a change determination mechanism consistent with the logic of natural geographical processes was constructed. The land cover change rate reflects the degree of change in land feature type, such as a decrease in vegetation and an increase in hard ground, while the bare land area change rate directly indicates the increasing trend in the degree of human disturbance. There is a highly correlated natural coupling relationship between the two: during construction and excavation processes, the land cover type typically shifts from high-coverage features such as vegetation to bare land, leading to simultaneous changes in both indicators. By setting a change threshold, it is possible to adaptively screen out areas showing true signs of construction activity, effectively filtering out minor changes caused by noise, seasonal differences, or temporary disturbances, and improving the accuracy and robustness of identifying areas of significant change.

[0118] Specifically, determining the change correlation coefficient according to all the surface cover change rates and all the bare land area change rates within the preset temporary determination time period includes:

[0119] Calculating the standard deviation of all the surface cover change rates between the initial moment and each moment within the preset temporarily determined time period to obtain a plurality of cover change fluctuation values;

[0120] Calculating the standard deviation of all the bare land area change rates between the initial moment and each moment within the preset temporarily determined time period to obtain a plurality of bare land change fluctuation values;

[0121] Draw a curve showing the coverage change fluctuation value over time within a preset temporary period, and perform vector processing to obtain a coverage change vector;

[0122] Draw a curve showing the fluctuation value of bare land change over time within a preset temporary period, and perform vector processing to obtain a bare land change vector;

[0123] The cosine similarity between the coverage change vector and the bare ground change vector is calculated to obtain the change correlation coefficient.

[0124] By calculating the standard deviation of the surface cover change rate and the bare land area change rate over a preset, temporarily determined time period, we generate cover change fluctuation values ​​and bare land change fluctuation values, respectively. Subsequently, we plot these two types of fluctuation values ​​over time as curves and perform vectorization (vectorization is a prior art and will not be further described here), generating cover change vectors and bare land change vectors. Finally, we calculate the cosine similarity between these two vectors (calculating cosine similarity is a prior art and will not be further described here), yielding a change correlation coefficient used to measure the consistency of change trends.

[0125] By quantifying the standard deviation fluctuations of the rates of change in land cover and bare land area, we not only capture the synchronous signals of reduced vegetation cover and increased bare soil exposure, but also utilize vectorization and cosine similarity to transform temporal fluctuations into angular similarity metrics in high-dimensional space, accurately reflecting the co-evolutionary characteristics of the two during dynamic changes. The directional analysis based on standard deviation fluctuation sensitivity and cosine similarity not only conforms to the natural coupling of vegetation and bare soil transformations—the loss of vegetation is inevitably accompanied by an increase in bare soil exposure due to human activities such as construction or excavation—but also mathematically reveals the consistency and strength of the changing trends, avoiding the temporal bias associated with single threshold determination and effectively improving the reliability and robustness of identifying areas of significant change.

[0126] Please continue reading Figure 3 As shown, it is a decision logic diagram for determining the area of ​​interest in this embodiment;

[0127] The process of selecting a plurality of new construction areas according to the displacement trajectory length and the displacement trajectory curvature of each temporary area includes:

[0128] When the length of the displacement trajectory is greater than a preset length threshold, determining the temporary area as a focus area to screen out a number of focus areas;

[0129] A number of new construction areas are selected based on the curvature of the displacement trajectory within all the areas of interest.

[0130] The preset length threshold is a standard for screening the length of displacement trajectories. It depends on the operating range and regional scale of the construction machinery and is usually set between tens of meters and hundreds of meters. In this embodiment, it is set to 100 meters, which can effectively distinguish short-distance occasional movements from real construction trajectories.

[0131] For the displacement trajectories of construction machinery extracted in each temporary area, if the total trajectory length exceeds a pre-set length threshold, the area is marked as a "focus area"; then, a statistical analysis is performed on the trajectory curvature of all focus areas to screen out trajectory segments with stable curvature changes and that meet the construction path characteristics, ultimately determining these areas as true new construction areas.

[0132] Displacement length threshold filtering eliminates sporadic, short-distance machine movements, focusing on high-intensity, large-scale construction activities. This aligns with the fundamental physical principle in engineering that "operation distance is positively correlated with operation scale." Trajectory curvature analysis, incorporating the kinematic principle that "straight or regular curved motion is a typical characteristic of machinery executing a predetermined trajectory," effectively distinguishes the orderly movement of construction machinery along roads or building boundaries from random jitter or detours. The synergistic effect of these two methods reflects the inherent coupling relationship between "working condition intensity and path shape" in spatial motion, significantly improving the accuracy and robustness of identifying new construction areas.

[0133] Please continue reading Figure 4 As shown, it is a decision logic diagram for determining a new construction area in this embodiment;

[0134] The process of selecting a number of new construction areas based on the curvature of the displacement trajectory in all the areas of interest includes:

[0135] Determining the change consistency according to the curvature of the displacement trajectory of any two of the regions of interest;

[0136] When the change consistency is greater than a preset consistency threshold, it is determined that both of the two focus areas are the new construction areas, so as to screen out several new construction areas.

[0137] The preset consistency threshold is the standard for determining the synchronization of trajectory curvature. It depends on the stability of the mechanical motion and the data noise level. It is usually set between 0.7 and 0.9. In this embodiment, it is set to 0.8, which can effectively distinguish between highly consistent construction paths and occasional movements.

[0138] The consistency of the curvature sequence of the construction machinery displacement trajectories in each pair of focus areas is calculated, and then the obtained consistency is compared with the preset consistency threshold. When the consistency exceeds the threshold, it is determined that the two focus areas are part of the same construction activity. Finally, all areas with highly consistent curvature changes are screened out as new construction areas.

[0139] The morphological similarity of the motion paths of construction machinery in different areas is quantified using trajectory curvature consistency. This is consistent with the principle in kinematics that "similar curvatures indicate similar motion patterns" and also conforms to the law that the round-trip operation paths of machinery in urban construction maintain continuity and repeatability. By setting the consistency threshold, occasional jitter and path deviation can be effectively filtered out, thereby enhancing the accuracy and stability of identifying new construction areas.

[0140] Specifically, the process of determining the change consistency according to the curvature of the displacement trajectory of any two of the regions of interest includes:

[0141] Obtaining the displacement trajectory curvature of each of the focus areas within a preset screening time period to obtain a trajectory curvature sequence;

[0142] The Pearson correlation coefficient of the trajectory curvature sequences of any two of the regions of interest is calculated to determine the change consistency.

[0143] The preset screening time is the time window used to extract the curvature sequence of the displacement trajectory of the area of ​​interest. It depends on the frequency of remote sensing image acquisition and the cycle of urban construction activities. It is usually set between 7 and 30 days. In this embodiment, it is set to 14 days, which can take into account the continuity of the change trend and the response speed of the detection.

[0144] Within the preset screening time, the time-varying curvature sequence of the displacement trajectory of construction machinery in each area of ​​interest is first extracted. Then, the Pearson correlation coefficient of the curvature sequences of any two areas is calculated and used as an indicator of the consistency of their changes. Finally, the areas with highly synchronized curvature changes are screened out as new construction areas based on the preset consistency threshold.

[0145] By quantifying the synchronous fluctuations of the curvature sequence through the Pearson correlation coefficient, it is possible to capture the consistency of the path morphology of construction machinery in different areas, which is in line with the basic physical principle of "similar curvature means similar motion pattern" in kinematics, and eliminate occasional noise and short-term jitter from a statistical perspective; at the same time, curvature, as a measure of the degree of path bending, together with dynamic parameters such as trajectory length and motion speed, reflects the spatial distribution law of mechanical operations, embodying the natural science logic of "the underlying coupling of path geometry and construction behavior", thereby significantly improving the accuracy, robustness and interpretability of the identification of new construction areas.

[0146] Specifically, the process of determining a number of significant change areas according to the image registration residual mean and the coverage change rate of each newly constructed area includes:

[0147] Calculating the standard deviation of the mean of all image registration residuals from the initial moment to each moment within a preset significant determination time length to obtain a plurality of residual fluctuation values;

[0148] Calculating the time derivatives of all the coverage change rates within the preset significant determination time length to obtain a coverage change coefficient;

[0149] Determine a joint change index based on the residual fluctuation value and the coverage change coefficient, wherein the calculation process of the joint change index is: Si=α×Ri+β×max|dCi / dt|, wherein Si is the joint change index, α is the preset residual fluctuation coefficient, Ri is the residual fluctuation value, β is the preset coverage coefficient, and max|dCi / dt| is the maximum value of the time derivative of the coverage change rate sequence Ci;

[0150] Based on the comparison result of the combined change index and the preset standard index, the newly constructed area is determined to be the significant change area, so as to determine several significant change areas.

[0151] The preset residual fluctuation coefficient α depends on the unit size of the image registration residual fluctuation value and its impact on the significance of the change. It is usually set between 0.1 / m and 10 / m. In this embodiment, it is set to 1 / m, which can effectively balance the impact of position changes in the joint change index without amplifying pixel-level errors.

[0152] The preset coverage coefficient β depends on the sensitivity of the coverage change rate to regional changes and the severity of its changes in the time dimension. It is usually set between 0.001 / day² and 1 / day². In this embodiment, it is set to 0.1 / day², which can enhance the contribution of rapid coverage changes to the identification of significant areas while avoiding misjudgment caused by interference factors.

[0153] Within the preset significant determination time, the standard deviation of the image registration residual mean from the initial moment to each moment of each new construction area is first calculated to obtain the residual fluctuation value; at the same time, the time derivative of the coverage change rate within the same time period is calculated to obtain the coverage variation coefficient; then the residual fluctuation value and the coverage variation coefficient are fused to generate a joint change index; finally, the joint change index is compared with the preset standard index threshold. When the joint change index exceeds the threshold, the new construction area can be determined as a significant change area.

[0154] The volatility of the registration error is measured by the standard deviation, reflecting the impact of surface structure changes on image alignment; the time derivative is used to capture the instantaneous rate of coverage change, revealing rapid changes in vegetation or surface caused by construction activities; the two are coupled in a joint change index, which not only conforms to the synchronization enhancement effect, but also follows the natural science logic of the interaction of "error fluctuation-change rate" in physical space, thereby jointly verifying the authenticity of the change signal at the mathematical and physical levels, greatly improving the accuracy and interpretability of identifying areas with significant changes.

[0155] Specifically, the process of adjusting the preset change threshold according to the position coordinates of each significantly changed area within the preset adjustment period to obtain the adjusted change threshold includes:

[0156] Obtaining the center position coordinates of all the regions with significant changes within the preset adjustment period to form a position coordinate set;

[0157] Clustering the position coordinate set, for each candidate cluster area, if the spatial distance between the significant change areas within it is less than or equal to a preset spatial cluster radius, and the number of significant change areas contained therein is greater than or equal to a preset minimum cluster number threshold, then extract it as a spatial change cluster area, so as to obtain a plurality of spatial change cluster areas;

[0158] Based on the number and spatial distribution of significantly changed regions in each of the change cluster regions, the corresponding change spatial density value is calculated to obtain a spatial density sequence. For each change cluster region, its spatial density value is calculated, Di=Ni / Ai, where Di is the spatial density value of the i-th cluster region, Ni is the number of significantly changed regions contained in the i-th change cluster region, and Ai is the spatial distribution area of ​​the i-th cluster region (existing technology);

[0159] The preset change threshold is adjusted based on the spatial density sequence and the change frequency information within a preset historical adjustment period to obtain the adjusted change threshold.

[0160] The preset adjustment period refers to the length of the time interval used to count and analyze changes in position coordinates in areas with significant changes. It depends on the typical duration of construction activities in the monitoring area and the frequency of remote sensing image acquisition. It is usually set between 7 and 90 days. In this embodiment, it is set to 30 days, which can effectively cover the spatial change characteristics within the construction period, ensuring that the adjustment threshold responds to recent construction dynamics and avoids misjudgment caused by short-term abnormal changes, thereby improving the accuracy and stability of threshold adjustment.

[0161] In this embodiment, clustering processing adopts the DBSCAN algorithm based on spatial density. The center coordinates of all areas with significant changes are used as the input point set. DBSCAN is called. Any point can form a cluster when it contains at least min_samples neighbors within the preset spatial clustering radius. The algorithm automatically identifies several spatial change clustering areas, whose boundaries can be extracted through the convex hull and used for subsequent area calculation and density evaluation.

[0162] The preset spatial clustering radius refers to the maximum spatial distance threshold used to determine whether each significantly changed area belongs to the same cluster when performing spatial clustering. It depends on the actual scale of the construction area in the target scene and the equipment observation accuracy. It is usually set between 10m and 100m. In this embodiment, it is set to 50m, which can effectively control the clustering range to match the spatial distribution characteristics of the typical construction impact area.

[0163] The preset minimum cluster number threshold refers to the minimum number of areas required to classify multiple areas with significant changes into the same cluster. It depends on the smallest effective change unit that can be observed in actual construction activities. It is usually set between 3 and 10. In this embodiment, it is set to 5. It can eliminate small-scale abnormal changes caused by accidental fluctuations and improve the credibility of the spatial clustering results.

[0164] A set of location coordinates is constructed by obtaining the center coordinates of all significantly changed areas within a preset adjustment period, and cluster analysis is performed on these areas. For each candidate cluster, if the spatial distance between significantly changed areas within it does not exceed the preset spatial cluster radius, and the number of significantly changed areas within it is at least a preset minimum cluster threshold, the cluster is considered a valid spatial change cluster. Based on this, a sequence of spatial density values ​​is calculated based on the number of significantly changed areas within each cluster and the spatial area they occupy. Finally, the original change threshold is dynamically adjusted by combining the spatial density sequence with change frequency data within the historical adjustment period to obtain a more responsive adjusted change threshold.

[0165] By measuring the intensity of localized changes through "spatial density," the concentration of construction activity is reflected by the degree of clustering of spatially significant areas of change. Historical change frequency is also incorporated to capture persistent trends, enabling adaptive adjustment of change thresholds. This adjustment logic is based on the underlying positive correlation between "density, frequency, and intensity." This means that greater spatial change density and frequency correlate with a greater intensity of the underlying disturbance source (new construction). This approach improves sensitivity and enhances the accuracy and robustness of change detection results.

[0166] Specifically, the preset change threshold is adjusted based on the spatial density sequence and the change frequency information within a preset historical adjustment period, and the process of obtaining the adjusted change threshold includes:

[0167] Extracting the number of times the significant change region is determined in each of the change cluster regions to obtain a historical frequency sequence;

[0168] A local adjustment coefficient is generated based on the spatial density sequence, the preset density weight, the historical frequency sequence, and the preset frequency weight to obtain an adjustment coefficient set. The calculation process of the local adjustment coefficient is: Ki=wd×(Di-Dmin) / (Dmax-Dmin)+wf×(Fi-Fmin) / (Fmax-Fmin), Ki is the local adjustment coefficient, wd is the preset density weight, Di is the spatial density value of the i-th change cluster area, Dmin is the minimum value in the spatial density sequence, Dmax is the maximum value in the spatial density sequence, wf is the preset frequency weight, Fi is the historical frequency of the i-th change cluster area, Fmin is the minimum value of the historical frequency sequence, Fmax is the maximum value of the historical frequency sequence, and wf+wd=1;

[0169] For each change concentration area, multiply its corresponding preset change threshold by the calculated local adjustment coefficient to form the regional adjusted change threshold, Ti'=T0×Ki, where Ti' is the adjusted change threshold of the i-th change concentration area and T0 is the preset change threshold;

[0170] Performing spatial interpolation and smoothing processing on the region adjustment change threshold set to obtain the adjustment change threshold.

[0171] The preset density weight refers to the weight parameter used to measure the influence of spatial density on the change threshold adjustment when calculating the local adjustment coefficient. It depends on the importance of the spatial aggregation characteristics of the area with significant changes in the overall change judgment. It is usually set between 0.3 and 0.7. In this embodiment, it is set to 0.5, which can reasonably reflect the guiding role of regional density changes on threshold setting.

[0172] The preset frequency weight refers to the weight parameter used to measure the influence of the historical change frequency on the change threshold adjustment when calculating the local adjustment coefficient. It depends on the reference value of the historical evolution trend for the current change prediction. It is usually set between 0.3 and 0.7. In this embodiment, it is set to 0.5, which can effectively reflect the enhanced guiding role of the continuous change area in change identification.

[0173] A historical frequency sequence is generated by extracting the number of times significant change areas were identified within each change cluster. The spatial density sequence and the historical frequency sequence are combined, and a set of local adjustment coefficients is generated using a normalized calculation method based on preset density and frequency weights. Subsequently, the preset change threshold for each cluster is multiplied by its corresponding local adjustment coefficient to form a set of regional adjustment change thresholds. This set of regional adjustment change thresholds is spatially interpolated and smoothed to obtain the adjusted change thresholds. Based on the spatial locations of each change cluster and its corresponding regional adjustment change threshold, an inverse distance weighted (IDW) method is used for spatial interpolation to generate a continuously changing threshold field. A local mean filtering algorithm is applied to this threshold field to eliminate local abnormal fluctuations and improve spatial continuity, resulting in an adjusted change threshold with spatial consistency and smooth transition characteristics throughout the detection area.

[0174] By integrating spatial density (reflecting the concentration of change) with historical frequency (reflecting the persistence of change), an adjustment factor model was constructed that aligns with the spatial-temporal coupling characteristics of natural phenomena. This reflects the scientific logic that higher density and more frequent changes should increase change detection sensitivity. Furthermore, a combination of normalization and weighting ensures the universality and comparability of local adjustment coefficients across different scales. Spatial interpolation and smoothing further enhance the continuity and stability of the threshold field across geographic space, effectively reducing the impact of local anomalies on overall detection results and enhancing the model's adaptability to complex feature evolution and detection accuracy.

[0175] Specifically, the process of determining a plurality of target areas according to the ideal change area and the significant change area obtained based on the adjusted change threshold includes:

[0176] The overlapping areas of all the significant change areas and all the target areas are taken as the target areas to determine a plurality of target areas.

[0177] By spatially overlaying the significant change regions obtained by adjusting the change threshold with the ideal change regions extracted by the Siamese residual network model, the overlapping portions are extracted as target regions. This process, based on explicit spatial overlap relationships, accurately locates and screens the true change regions, avoiding false positives or negatives caused by a single feature or model, and ultimately determines several target regions as the core output of the binary detection results.

[0178] By introducing a confidence-enhancing mechanism through spatial overlap, the identification of target areas becomes more robust. Significantly changed areas are derived from surface feature change parameters such as the bare land area change rate and the cover change rate. These parameters are driven by surface physical processes and have a clear spatiotemporal logic. Ideally, ideally changed areas are derived from the high-dimensional representation of image residual signals by the twin network, which is sensitive to subtle changes and structural anomalies. The overlapping areas represent the intersection of "significant physical change" and "high-confidence model identification areas." This aligns with the natural science principle of "multi-source information fusion and discrimination" in remote sensing change detection and embodies the logic of "feature-driven + model-aware." This improves the accuracy and stability of target area identification and has excellent promotional and practical value.

[0179] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A remote sensing image binary detection method that integrates feature differences and twin structures, characterized by: include: Real-time acquisition of dual-temporal remote sensing images of the detection area within the urban development zone; Dividing the dual-temporal remote sensing image into a plurality of unit areas; Extracting the surface coverage change rate, image registration residual mean, displacement track length and displacement track curvature of the construction machinery, and bare land area change rate within each unit area; determining, based on the surface cover change rate, the bare land area change rate, and a preset change threshold, that the surface structure in the unit area has changed, and forming a plurality of temporary areas; Selecting a number of new construction areas based on the displacement trajectory length and the displacement trajectory curvature of each temporary area; Determine a number of significant change areas according to the image registration residual mean and the coverage change rate of each newly constructed area; Adjusting the preset change threshold according to the position coordinates of each of the significantly changed areas within a preset adjustment period to obtain an adjusted change threshold; Inputting the dual-temporal remote sensing image into a preset twin residual network model to obtain a number of ideal change areas; determining a plurality of target areas according to the ideal change area and the significant change area obtained based on the adjusted change threshold; The process of determining a number of areas with significant changes according to the image registration residual mean and the coverage change rate of each newly constructed area includes: Calculating the standard deviation of the mean of all image registration residuals from the initial moment to each moment within a preset significant determination time length to obtain a plurality of residual fluctuation values; Calculating the time derivatives of all the coverage change rates within the preset significant determination time length to obtain a coverage change coefficient; determining a joint variation index according to the residual fluctuation value and the coverage variation coefficient; Determining the newly constructed area as the significant change area based on a comparison result of the combined change index and a preset standard index, so as to determine a number of significant change areas; The process of adjusting the preset change threshold according to the position coordinates of each of the significantly changed areas within the preset adjustment period to obtain the adjusted change threshold includes: Obtaining the center position coordinates of all the regions with significant changes within the preset adjustment period to form a position coordinate set; Performing clustering processing on the position coordinate set, and extracting a number of spatially varying cluster areas based on a preset spatial cluster radius and a preset minimum cluster number threshold; Based on the number and spatial distribution of significant change areas within each of the change clustering areas, the corresponding change spatial density value is calculated to obtain a spatial density sequence; The preset change threshold is adjusted based on the spatial density sequence and the change frequency information within a preset historical adjustment period to obtain the adjusted change threshold.

2. The remote sensing image binary detection method of fusing feature differences and twin structures according to claim 1 is characterized in that: The process of determining that the surface structure in the unit area has changed according to the surface cover change rate, the bare land area change rate, and a preset change threshold, and forming a plurality of temporary areas includes: Determining the change correlation based on all of the surface cover change rates and all of the bare land area change rates within a preset temporary determined time period; According to the comparison result of the change correlation and the preset change threshold, it is determined that the surface structure in the unit area has changed, and a plurality of temporary areas are formed.

3. The remote sensing image binary detection method of fusing feature differences and twin structures according to claim 2 is characterized in that: Determining the change correlation coefficient based on all the surface cover change rates and all the bare land area change rates within the preset temporary determined time period includes: Calculating the standard deviation of all the surface cover change rates between the initial moment and each moment within the preset temporarily determined time period to obtain a plurality of cover change fluctuation values; Calculating the standard deviation of all the bare land area change rates between the initial moment and each moment within the preset temporarily determined time period to obtain a plurality of bare land change fluctuation values; A change correlation coefficient is determined based on all of the coverage change fluctuation values ​​and all of the bare land change fluctuation values.

4. The remote sensing image binary detection method of fusing feature differences and twin structures according to claim 3 is characterized in that: The process of selecting a plurality of new construction areas according to the displacement trajectory length and the displacement trajectory curvature of each temporary area includes: Screening out a number of areas of interest based on a comparison result of the displacement trajectory length and a preset length threshold; A number of new construction areas are selected based on the curvature of the displacement trajectory within all the areas of interest.

5. The remote sensing image binary detection method of fusing feature differences and twin structures according to claim 4 is characterized in that: The process of selecting a number of new construction areas based on the curvature of the displacement trajectory in all the areas of interest includes: Determining the change consistency according to the curvature of the displacement trajectory of any two of the regions of interest; According to the comparison result of the change consistency and the preset consistency threshold, it is determined that the two focus areas are both the new construction areas, so as to screen out several new construction areas.

6. The remote sensing image binary detection method of fusing feature differences and twin structures according to claim 5, characterized in that: The process of determining the change consistency according to the curvature of the displacement trajectory of any two of the regions of interest includes: Obtaining the displacement trajectory curvature of each of the focus areas within a preset screening time period to obtain a trajectory curvature sequence; The Pearson correlation coefficient of the trajectory curvature sequences of any two of the regions of interest is calculated to determine the change consistency.

7. The remote sensing image binary detection method of fusing feature differences and twin structures according to claim 6, characterized in that: The process of adjusting the preset change threshold based on the spatial density sequence and the change frequency information within the preset historical adjustment period to obtain the adjusted change threshold includes: Extracting the number of times the significant change region is determined in each of the change cluster regions to obtain a historical frequency sequence; Generating a local adjustment coefficient based on the spatial density sequence, the preset density weight, the historical frequency sequence, and the preset frequency weight to obtain an adjustment coefficient set; Modifying the preset change threshold corresponding to each of the change cluster regions based on the adjustment coefficient set to form a region adjustment change threshold set; Performing spatial interpolation and smoothing processing on the region adjustment change threshold set to obtain the adjustment change threshold.

8. The remote sensing image binary detection method of fusing feature differences and twin structures according to claim 7, characterized in that: The process of determining a plurality of target areas according to the ideal change area and the significant change area obtained based on the adjusted change threshold comprises: The overlapping areas of all the significant change areas and all the target areas are taken as the target areas to determine a plurality of target areas.

Images