A method for change detection of unmanned aerial vehicle remote sensing image for rare earth mining area exploration

Abstract

The application provides a UAV remote sensing image change detection method for rare earth mining area identification, which is suitable for complex identification scenes of ion-adsorption type rare earth mining area ecological disturbance and illegal mining change area; the method adopts an overall architecture idea of "front and rear phase feature extraction-multiscale change fusion-cross-scale guided fusion", enhances the expression ability of real change information between front and rear phase features through a difference interaction attention fusion module, so that the model can more effectively distinguish mining disturbance from general background changes; through a boundary change fusion refining module, edge information of the change area is strengthened, and the contour recovery ability of small-scale, slender and boundary broken mining patches is improved; through a mixed attention refining module, shallow spatial details and deep semantic information are comprehensively utilized, and the detection accuracy and false change suppression ability of the change area in a complex background are improved.

Description

Technical Field

[0001] This invention relates to the field of rare earth mining change identification technology, specifically to a method for detecting changes in UAV remote sensing images used for rare earth mining identification. Background Technology

[0002] Rare earth mining areas, especially ion-adsorption type rare earth mining areas, are typically distributed in the hilly and mountainous regions of southern China. These areas are characterized by significant topographic relief, high vegetation cover, and complex surface textures. The disturbance patches formed by mining activities often exhibit small scale, irregular shape, fragmented boundaries, and discrete spatial distribution. In UAV remote sensing imagery, these mining disturbance areas are easily mixed with background information such as exposed bare soil, road construction, vegetation degradation, soil erosion, shadow changes, and differences in illumination, making it difficult to distinguish between the disturbed areas and non-target changes.

[0003] Meanwhile, UAV images from different time periods are affected by factors such as flight altitude, imaging angle, lighting conditions, vegetation growth status, and registration errors, which can easily generate a large amount of false change information, making the identification of mining changes face high uncertainty.

[0004] Although traditional change detection methods can extract surface change information through image difference, threshold segmentation, spectral feature analysis, manual visual interpretation, or conventional machine learning methods, these methods usually rely on human experience and shallow features, making it difficult to fully express the deep semantic differences between two-temporal images in complex backgrounds.

[0005] While existing deep learning change detection methods have achieved certain results in general remote sensing scenarios, they still have problems in high-resolution UAV imagery scenarios in rare earth mining areas, such as missed detection of small-scale mining patches, incomplete extraction of slender disturbance areas, insufficient recovery of broken boundaries, numerous false detections of changes in complex backgrounds, and insufficient fusion of multi-scale features. These problems make it difficult to meet the application requirements of refined, automated, and highly reliable rare earth mining identification. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a UAV remote sensing image change detection model for rare earth mining identification. This model takes UAV remote sensing images from before and after mining as input, extracts multi-scale fusion features through a deep backbone network, and introduces a difference interactive attention fusion module, a boundary change fusion refinement module, and a hybrid attention refinement module during the multi-scale feature fusion process to achieve refined identification of mining change areas in rare earth mining areas.

[0007] Specifically, this invention enhances the ability to express real change information between dual-temporal UAV remote sensing images through a difference-interactive attention fusion module, enabling the model to more effectively distinguish between mining disturbances and general background changes; strengthens the edge information of changed areas through a boundary change fusion refinement module, improving the contour recovery ability of small-scale, elongated, and boundary-fragmented mining patches; and improves the detection accuracy and pseudo-change suppression ability of changed areas in complex backgrounds by comprehensively utilizing shallow spatial details and deep semantic information through a hybrid attention refinement module.

[0008] To achieve the above objectives, this invention proposes a method for detecting changes in UAV remote sensing images for rare earth mining area identification, comprising the following steps: S1: The UAV remote sensing images of the target area before and after mining are registered and cropped into fixed-size image blocks. Then, they are input into the dual-branch backbone network with shared parameters to extract multi-scale pre-temporal features and multi-scale post-temporal features respectively. S2: The multi-scale pre-temporal features and multi-scale post-temporal features are simultaneously input into the change fusion branch at the corresponding scale of the dual-branch backbone network for interactive modeling and change representation. The change fusion branch includes a difference interactive attention fusion module and a boundary change fusion refinement module. The difference interactive attention fusion module performs difference perception and interactive fusion on the pre-temporal features and post-temporal features to obtain a change attention map, which highlights potential change areas and enhances change response. The boundary change fusion refinement module introduces a boundary-aware refinement mechanism, and further combines the pre-temporal features, post-temporal features and change attention map to refine the contour, local details and edge structure of the change area to obtain multi-scale change features. S3: Input the multi-scale change features into the hybrid attention refinement module. The hybrid attention refinement module is mainly used to guide and fuse the multi-scale change features across levels. Through top-down information transmission, the high-level semantics in the deep features are gradually injected into the shallow features to obtain the multi-scale change features after cross-scale guidance and fusion. S4: The multi-scale change features after cross-scale guidance and fusion are recovered through change prediction, and then convolutional mapping is used to generate the final detection result map.

[0009] The present invention has the following beneficial effects: 1. This invention uses UAV remote sensing images from two phases before and after mining as input. Through multi-scale feature extraction, modeling of temporal change information before and after mining, fusion of change features, and output of results, it achieves automated and refined identification of mining disturbance areas in rare earth mining areas. This solves the problems of low efficiency, insufficient automation, and weak adaptability to complex scenarios of traditional manual inspection, visual interpretation, and conventional change detection methods.

[0010] 2. This invention enhances the response to real mining disturbance areas through a difference-interactive attention fusion module, reducing interference from pseudo-changes in complex backgrounds; at the same time, it strengthens the edge information of mining change areas through a boundary change fusion refinement module, improving problems such as blurred boundaries, breaks, holes, and incomplete contours, thereby improving the boundary integrity and spatial positioning accuracy of mining change area detection.

[0011] 3. This invention integrates shallow spatial details and deep semantic information through a hybrid attention mechanism, which improves the model's ability to detect small disturbance patches, elongated mining areas and fragmented change areas. It can also be combined with image cropping, sliding window detection, overlapping area fusion and coordinate writing back to achieve whole-map-level mining change recognition of large-format UAV images. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the overall process framework of the UAV remote sensing image change detection method provided by the present invention. Figure 2 This is a schematic diagram of the structure of the difference interaction attention fusion module of the present invention; Figure 3 This is a schematic diagram of the boundary change fusion refining module of the present invention; Figure 4 This is a schematic diagram of the hybrid attention refining module of the present invention; Figure 5 This is a schematic diagram showing a visual comparison of the method provided by this invention and existing change detection methods on disturbance samples from rare earth mining areas. Detailed Implementation

[0013] like Figure 1-4 As shown, this invention proposes a method for detecting changes in UAV remote sensing images for rare earth mining identification, applicable to complex identification scenarios involving ecological disturbances and illegal mining changes in ion-adsorption type rare earth mining areas. The method adopts an overall architecture of "pre- and post-temporal feature extraction—multi-scale change fusion—cross-scale guided fusion," aiming to simultaneously improve the ability to discriminate changes in areas under complex backgrounds, restore boundaries, and express multiple scales.

[0014] The specific steps of this method are as follows: S1: Feature Extraction The UAV remote sensing images of the target area before and after mining are input into a dual-branch backbone network. The dual-branch backbone network is divided into two temporal branches, both of which use ResNet18 as the backbone network. Four levels of feature representations, namely Stage1 to Stage4 features, are extracted from the shallow to deep layers of the backbone network. The four levels correspond to resolution scales of 1 / 4, 1 / 8, 1 / 16 and 1 / 32, respectively. Finally, the pre-temporal features and post-temporal features of the four scales corresponding to the two temporal branches are obtained.

[0015] S2: Constructing a change fusion branch A change fusion branch is constructed at each scale of the dual-branch backbone network to interactively model and represent the changes in the pre- and post-temporal features of the corresponding scale. The change fusion branch includes a difference-interaction attention fusion module (DIAF) and a boundary change fusion refinement module (BCFR). This combination of difference representation and boundary refinement enables the dual-branch backbone network to obtain more discriminative and structurally complete multi-scale change features during the encoding stage.

[0016] 1. DIAF Module Design Illegal mining disturbances in ion-adsorption rare earth deposits are often mixed with background changes such as vegetation fluctuations, bare soil exposure, and light differences, easily generating spurious change responses. To enhance the ability of earlier and later temporal features to distinguish the true change area, this invention designs a DIAF module, the specific structure of which is as follows: Figure 2 As shown.

[0017] Given the first Pre-temporal features at various scales and post-phase characteristics First, we construct explicit differential features, as shown in equation (1): (1) in, Indicates the first Explicit difference features at each scale are used to characterize the intensity of change of features in the preceding and following time phases at corresponding positions. At the same time, in order to characterize the correlation and consistency between features in the preceding and following time phases, implicit interaction features are further constructed, as shown in Equation (2). (2) in, Indicates the first Implicit differential features at each scale This indicates element-wise multiplication. (and) different, It focuses more on describing the co-occurrence relationship and consistency response between features from the previous and subsequent time periods. Both describe cross-temporal change cues from the perspectives of "difference" and "correlation". Based on this, explicit difference features and implicit interaction features are concatenated in the channel dimension, and after convolutional mapping and Sigmoid activation, they are input into the spatial attention module to generate a change attention map, as shown in Equation (3): (3) in, Indicates channel splicing. This represents the convolution operation. This represents the Sigmoid activation function. Indicates spatial attention, For the first The change attention map generated at each scale can further highlight potential change regions and suppress background redundant responses by fusing explicit change information with implicit relevant information.

[0018] After obtaining the change attention map, the DIAF module uses it to weight and enhance the input features of the preceding and following time phases, and then fuses the enhanced features to form a change representation at the current scale. Subsequently, it preserves the explicit difference feature cues at the lower level through convolutional projection and residual compensation, thereby improving the accuracy and robustness of change region detection.

[0019] 2. BCFR Module Design Illegal mining disturbance areas are typically small-scale, irregularly shaped, and have fragmented boundaries, making it difficult to accurately reconstruct their spatial contours based solely on initial change characterizations. To further enhance boundary information and improve the structural integrity of the region, this invention designs a BCFR module, the structure of which is as follows: Figure 3 As shown.

[0020] The BCFR module first performs the following: and Boundary enhancement is performed to obtain the boundary enhancement features of the preceding and following time phases. and As shown in equations (4) and (5): (4) (5) in, , and These represent 3×3, 5×5, and 7×7 convolutional branches, respectively, used to extract local boundary details and contextual information under different receptive fields; This represents an attention operation used to adaptively reweight input features to highlight change boundaries, structurally discontinuous regions, and discriminative semantic responses. and Both represent convolution mapping operations; and They represent the first Boundary enhancement features of the preceding and following phases at various scales.

[0021] Subsequently, the enhanced features of the previous phase boundary were applied. Post-temporal boundary enhancement features With attention to changes Concatenation is performed along the channel dimension, and then mapped using convolution. The variation characteristics after boundary refinement are shown in equation (6): (6) in, Indicates the first The variation characteristics after boundary refinement at each scale are more discriminative and structurally complete multi-scale variation characteristics.

[0022] Through the above process, the BCFR module can further enhance the boundary contours, local details, and structural continuity of the change region while preserving the original change cues and change attention response of the two temporal phases. It can also suppress problems such as boundary breaks, internal voids, and local adhesions, thereby providing more complete and clear change features for subsequent multi-scale feature fusion and change prediction.

[0023] The BCFR module enhances the fused features on two levels: firstly, the features extracted by the multi-scale convolution branches are spatially attention-weighted to highlight the boundary positions and structural discontinuities; secondly, the features from previous and subsequent phases are recalibrated through attention operations to enhance the response to discriminative semantic channels; and the two are uniformly mapped to form refined features, thus simultaneously taking into account both boundary detail restoration and semantic expression stability.

[0024] The BCFR module, combined with a residual refinement strategy, generates the module output. While preserving the original change clues, it further suppresses problems such as boundary breaks, internal voids, and local adhesions, thereby providing more complete and clearer change features for subsequent multi-scale decoding.

[0025] S3: Cross-scale guided fusion The illegally mined rare earth minerals exhibit significant scale variations, making it difficult for single-level features to simultaneously capture both semantic discrimination and detailed representation. To enhance the collaborative representation capability among features of different scale variations, a Hybrid Attention Refinement Module (HARM) was designed, with the following structure: Figure 4 As shown.

[0026] The HARM module is mainly used for cross-level guidance and fusion of multi-scale variation features. The HARM module takes the multi-scale variation features obtained in the encoding stage as input. Different levels of features correspond to spatial detail information and high-level semantic information from shallow to deep. The high-level semantics in the deep features are gradually injected into the shallow features. With the deep features as semantic guidance, adjacent scale features are fused layer by layer. High-level semantic information is gradually injected into the middle and low-level features, thereby enhancing the model's ability to express small-scale variation regions, fine-grained structures and boundary details.

[0027] In the fusion process at each scale, the HARM module first upsamples the features of adjacent scales and concatenates them with the change features of the current scale. Then, channel alignment and feature compression are completed through 1×1 convolution to obtain the fused features of the current scale. On this basis, attention units are introduced to adaptively reweight the fused features to highlight change-related responses and suppress background redundancy information, thus obtaining the change features of the current level after cross-scale guidance and fusion. This process is carried out step by step at multiple scales, so that deep semantic information and shallow spatial details can be fully coordinated, thereby improving the consistency and complementarity of multi-scale change feature expression and obtaining multi-scale change features after cross-scale guidance and fusion.

[0028] Through the above design, the HARM module can effectively enhance the correlation and discriminative power between features of different scales of change, improve the boundary recovery effect while maintaining the stability of the main structure of the change region, and provide more stable and refined feature support for subsequent change region prediction.

[0029] S4: Output of detection results The multi-scale change features output by the HARM module, after cross-scale guidance and fusion, are used for change prediction and recovery through upsampling and feature fusion. Finally, the detection result map is generated through convolutional mapping.

[0030] The overall architecture of the detection method provided by this invention can more effectively cope with complex backgrounds, strong pseudo-change interference, and blurred target boundaries in the ecological disturbance scenario of ion adsorption rare earth mining areas, thereby improving the extraction accuracy and boundary integrity of illegally mined change areas.

[0031] The mining changes in the same rare earth mining area were detected using the method provided in this invention and existing methods such as HANet, LightCDNet, Changer, BiT, BAN, CGNet, ChangeFormer, and DGMA2net. The detection results are as follows: Figure 5 As shown in the figure, GT is the truth label.

[0032] Depend on Figure 5It is evident that different change detection methods have varying degrees of shortcomings when facing complex rare earth mining disturbance scenarios. HANet and LightCDNet have weak characterization capabilities for slender and fragmented change patches, and the detection results are prone to fragmentation, holes, and local omissions in the change areas. Changer's change extraction results are relatively conservative. Although some non-change areas have less response, it is prone to causing shrinkage of the boundary of the mining change area and local omissions. BiT can identify some change subjects, but there are still omissions in narrow extended structures, edge turning areas, and complex background boundaries, and the boundary roughness or local adhesion may occur. BAN, CGNet, and ChangeFormer have improved in terms of main area recovery, but they are still prone to generating scattered noise in areas with strong vegetation texture, bare soil exposure, and road disturbance. DGMA2Net can respond to the change subjects relatively completely, but it is more sensitive to background disturbances and is prone to scattered or patchy pseudo-responses in non-change areas because it sacrifices some detection accuracy while pursuing higher recall.

[0033] However, compared to other methods, the method of this invention exhibits better regional integrity and boundary fit. For mining disturbance patches that are long and narrow, fragmented, small-scale, and have blurred boundaries, this invention can restore the main structure of the change more completely and reduce fractures, holes, and local omissions. For areas with complex vegetation, bare soil, roads, and mixed shadows, this invention can significantly reduce the response to background pseudo-changes, making the detection results closer to the true annotation.

[0034] Figure 5 The visualization results show that the method of the present invention not only improves the overall detection capability of the changed region, but also enhances the ability to express the boundary details and local structure of the changed region.

[0035] Meanwhile, the performance comparison results of existing change detection methods and the method provided by this invention in the same rare earth mining change detection task are shown in Table 1 below: Table 1. Performance comparison of existing change detection methods and the method of the present invention in change detection tasks.

[0036] As shown in Table 1, the method of the present invention achieves the best results in the main evaluation indicators such as F1, IoU, Precision, and Oa. Although the Recall of the method of the present invention is 90.28%, which is slightly lower than DGMA2-Net's 90.68% by 0.40 percentage points, the method of the present invention significantly improves Precision and Oa. This indicates that the method of the present invention does not simply expand the range of change detection, but effectively reduces false detections while maintaining a high detection rate. It achieves a more reasonable balance between change integrity and false detection control. Therefore, the overall performance of the method of the present invention is better than that of existing change detection methods.

Claims

1. A method for detecting changes in UAV remote sensing images for rare earth mining area identification, characterized in that, Includes the following steps: S1: The UAV remote sensing images of the target area before and after mining are registered and cropped into fixed-size image blocks. Then, they are input into the dual-branch backbone network with shared parameters to extract multi-scale pre-temporal features and multi-scale post-temporal features respectively. S2: The multi-scale pre-temporal features and multi-scale post-temporal features are simultaneously input into the change fusion branch at the corresponding scale of the dual-branch backbone network for interactive modeling and change representation. The change fusion branch includes a difference interactive attention fusion module and a boundary change fusion refinement module. The difference interactive attention fusion module performs difference perception and interactive fusion on the pre-temporal features and post-temporal features to obtain a change attention map, which highlights potential change areas and enhances change response. The boundary change fusion refinement module introduces a boundary-aware refinement mechanism, and further combines the pre-temporal features, post-temporal features and change attention map to refine the contour, local details and edge structure of the change area to obtain multi-scale change features. S3: Input the multi-scale change features into the hybrid attention refinement module. The hybrid attention refinement module is mainly used to guide and fuse the multi-scale change features across levels. Through top-down information transmission, the high-level semantics in the deep features are gradually injected into the shallow features to obtain the multi-scale change features after cross-scale guidance and fusion. S4: The multi-scale change features after cross-scale guidance and fusion are recovered through change prediction, and then convolutional mapping is used to generate the final detection result map.

2. The method for detecting changes in UAV remote sensing images for rare earth mining area identification according to claim 1, characterized in that, The dual-branch backbone network includes two temporal branches, each using ResNet18 as the backbone network, and extracts multi-scale pre-temporal features and multi-scale post-temporal features from the shallow to deep layers of the backbone network.

3. The method for detecting changes in UAV remote sensing images for rare earth mining area identification according to claim 2, characterized in that, The operation flow of the differential interaction attention fusion module is as follows: given the first... The pre-temporal features at each scale and the aforementioned post-phase features First, we construct explicit differential features, as shown in equation (1): （1） in, Indicates the first The explicit difference features at each scale are used to characterize the intensity of change of the preceding and following phase features at corresponding positions. Meanwhile, to characterize the correlation and consistency between the preceding and following phase features, implicit interaction features are further constructed, as shown in equation (2). （2） in, Indicates the first The implicit difference features at each scale. Represents element-wise multiplication, and different, The focus is more on describing the co-occurrence relationship and consistency response between the preceding and following phase features; Based on this, the explicit difference features and the implicit interaction features are concatenated in the channel dimension, and after convolutional mapping and sigmoid activation, they are input into the spatial attention module to generate the change attention map, as shown in equation (3): （3） in, Indicates channel splicing. This represents the convolution operation. This represents the Sigmoid activation function. Indicates spatial attention. For the first The change attention map generated at each scale.

4. The method for detecting changes in UAV remote sensing images for rare earth mining area identification according to claim 3, characterized in that, The operation process of the boundary change fusion and refinement module is as follows: First, respectively... and Boundary reinforcement is performed, as shown in equations (4) and (5): （4） （5） in, , and These represent 3×3, 5×5, and 7×7 convolutional branches, respectively, used to extract local boundary details and contextual information under different receptive fields; This represents an attention operation used to adaptively reweight input features to highlight change boundaries, structurally discontinuous regions, and discriminative semantic responses. and Both represent convolution mapping operations; and They represent the first Boundary enhancement features of the preceding and following phases at various scales; Subsequently, the enhanced features of the previous phase boundary were applied. Post-temporal boundary enhancement features Note the changes mentioned above. Concatenation is performed along the channel dimension, and then mapped using convolution. The variation characteristics after boundary refinement are shown in equation (6): （6） in, This represents the multi-scale variation characteristics.

5. The method for detecting changes in UAV remote sensing images for rare earth mining area identification according to claim 1, characterized in that, The operation process of the hybrid attention refinement module is as follows: In the fusion process at each scale, the features of adjacent scales are first upsampled and then spliced ​​with the change features of the current scale. Subsequently, channel alignment and feature compression are completed through 1×1 convolution to obtain the fused features at the current scale. On this basis, an attention unit is introduced to adaptively reweight the fused features to highlight change-related responses and suppress background redundancy information, thus obtaining the change features of the current level after cross-scale guidance and fusion. This process is carried out step by step at multiple scales, so that deep semantic information and shallow spatial details can be fully coordinated to obtain the multi-scale change features after cross-scale guidance and fusion.

6. The method for detecting changes in UAV remote sensing images for rare earth mining area identification according to claim 1, characterized in that, The change prediction and recovery is performed by upsampling and feature fusion of the multi-scale change features after cross-scale guidance and fusion.

Images