An artificial intelligence-based low-altitude remote sensing anomaly identification method

Abstract

The application relates to an artificial intelligence-based low-altitude remote sensing anomaly identification method, and belongs to the technical field of digital image processing and geographic information remote sensing monitoring. The method constructs a dynamic and content-aware feature learning and decision-making framework. First, parallel branches with different expansion rates are used to comprehensively capture diversified receptive field features from local details to global context, then channel compression and excitation and channel dimension pooling operations are respectively performed on the features, a lightweight 'channel-space dual-channel gating mechanism' is realized, appropriate weights are adaptively allocated to feature maps of different branches and different positions, and geographic scene prior knowledge is fused to exert spatial attention at the feature level, further suppress the response of background regions irrelevant to anomalies, and make the model focus on the real suspicious areas. Finally, the detection capability for multi-scale anomalies, especially small-scale and weak-contrast anomalies, can be significantly improved, so that the low-altitude remote sensing anomaly identification accuracy is improved.

Description

Technical Field

[0001] This invention relates to the fields of digital image processing and geographic information remote sensing monitoring technology, and in particular to an artificial intelligence-based method for identifying low-altitude remote sensing anomalies. Background Technology

[0002] Low-altitude remote sensing technology mainly refers to the technology of using platforms such as drones and light aircraft to conduct Earth observation at low altitudes of hundreds to thousands of meters above the ground, acquiring high spatial resolution and timely remote sensing image data. This technology plays an irreplaceable role in fields such as land surveying, geographic information updating, and mineral resource exploration. Through intelligent anomaly identification of low-altitude remote sensing images, it is possible to efficiently detect changes, anomalies, or potential risk targets on the Earth's surface. Examples include illegal mining sites, potential landslide hazards, water pollution spread, and forest fire risk areas. This technology transforms traditional, labor-intensive, time-consuming, and limited-coverage visual interpretation methods into efficient, objective, and large-scale automated monitoring methods, greatly improving the management efficiency and emergency response speed of related industries.

[0003] Because low-altitude remote sensing imagery covers a wide variety of geographical scenes, ranging from vast plains and rolling hills to steep mountains, the textures, spectra, and spatial structures of surface cover (such as vegetation, water bodies, bare rock, and man-made structures) are complex and varied. Therefore, the targets to be identified vary greatly in scale. For example, in the supervision of mineral resource development, it is necessary to identify both large illegal mining faces that may cover tens of thousands of square meters and small, newly built sheds or small illegal mining sites that are only a few dozen square meters in size. Therefore, this "multi-scale" characteristic is a core feature of low-altitude remote sensing anomaly identification.

[0004] Existing general-purpose object detection models typically employ a fixed backbone network (such as ResNet or VGG) in conjunction with a Feature Pyramid Network (FPN) to extract multi-scale features. However, the performance of such methods deteriorates significantly when faced with extreme scale variations (especially small-scale anomalies) and complex background interference. The main reasons are: First, fixed-structure feature extraction networks have limited representation capabilities when dealing with small-scale anomaly features that differ significantly from the training data distribution, particularly those highly similar to the background or with extremely low signal-to-noise ratios. Downsampling operations in deep networks easily lead to the loss of crucial details in small targets. Second, complex natural scene backgrounds (such as shadows from dense vegetation, textures of exposed rocks, and cloud cover) generate numerous interfering features that are similar to real anomalies in shallow texture and color. The model easily misinterprets this background noise as foreground anomalies, or conversely, obscures real anomalies embedded in complex backgrounds.

[0005] This tendency to miss or falsely detect multi-scale anomalies, especially small-scale, low-contrast anomalies in complex backgrounds, severely limits the practical value and reliability of low-altitude remote sensing intelligent monitoring systems in actual deployment. Therefore, there is an urgent need for a low-altitude remote sensing anomaly identification method that can adapt to complex geographical scenes, accurately focus on multi-scale anomaly features, and effectively suppress background interference. Summary of the Invention

[0006] In view of this, the present invention provides an artificial intelligence-based method for identifying low-altitude remote sensing anomalies, in order to solve the technical problems of high false negative rate and weak resistance to background interference in current low-altitude remote sensing anomaly identification methods in complex geographical scenarios.

[0007] The present invention provides a low-altitude remote sensing anomaly identification method based on artificial intelligence, comprising:

[0008] Construct an input feature map from any low-altitude remote sensing image, and calculate a preset number of spectral index maps for the input feature map;

[0009] Construct a preset number of dilated convolutional branches with the same kernel size but different dilation rates. Input the input feature map into each dilated convolutional branch to obtain a preset number of output feature maps. Concatenate all output feature maps according to the channel dimension to obtain multi-receptive field features.

[0010] Channel compression and activation are performed on the multi-receptive field features to obtain a channel attention weight vector with the same number of dimensions as the multi-receptive field channels. The channel attention weight vector is then multiplied with the multi-receptive field features channel by channel to obtain the channel weighted features.

[0011] Max pooling and average pooling are performed on the channel-weighted features according to the channel dimension. The max pooling result and the average pooling result are concatenated and then passed through a standard convolutional layer of preset size to generate a spatial attention map. The spatial attention map is multiplied element-wise with the channel-weighted features to obtain the fused weighted features.

[0012] Based on the current anomaly detection task objective, relevant spectral index maps are selected from the preset categories of spectral index maps of the input feature map. A prior attention mask is constructed based on the relevant spectral index maps and converted into a prior attention weight map. The fused weighted features are enhanced based on the prior attention weight map to obtain an anomaly sensitive feature map.

[0013] Based on the anomaly-sensitive feature map, anomaly identification is performed for the current anomaly identification task objective.

[0014] Furthermore, the construction of the input feature map includes:

[0015] The low-altitude remote sensing image is segmented to obtain multiple image blocks with a first preset number of pixels. There is an overlapping area with a second preset number of pixels between adjacent image blocks. Any image block is used as the input feature map.

[0016] Furthermore, the preset spectral index maps include a normalized difference vegetation index map, a normalized difference water index map, a normalized difference building index map, and a bare surface index map.

[0017] Furthermore, the process of channel compression and activation of the multi-receptive field features to obtain a channel attention weight vector with the same number of dimensions as the multi-receptive field channels includes:

[0018] The average pooling is performed sequentially on each channel of the multi-receptive field feature to obtain the corresponding average pooling result. The average pooling results of all channels are then arranged in the order of each channel to form a channel compression result vector with the same number of dimensions as the multi-receptive field channels.

[0019] Calculate the channel attention weight vector based on the channel compression result vector:

[0020] ,

[0021] in, This represents the channel attention weight vector. This represents the compression result vector of the channel. Let the first learnable parameter matrix be of size . ,in This represents the total number of dilated convolution branches. This represents the number of channels in the output feature map of each dilated convolution branch. Indicates the reduction ratio. This represents a filtering function, whose purpose is to determine... Check if each element in the calculation result is less than 0, and replace the corresponding element with 0 if it is. This represents the second learnable parameter matrix, whose size is... , This represents the normalization function.

[0022] Furthermore, the step of concatenating the max pooling result and the average pooling result and then generating a spatial attention map through a standard convolutional layer of a preset size includes:

[0023] The max pooling result and the average pooling result are concatenated and then input into the standard convolutional layer of the preset size. The output of the standard convolutional layer is normalized to obtain the spatial attention map.

[0024] Furthermore, the length and width of the convolutional kernels in the preset-size standard convolutional layer are equal and odd numbers.

[0025] Furthermore, the construction of the prior attention mask based on the relevant spectral index map includes:

[0026] The spectral index map in the relevant spectral index map that is negatively correlated with the current anomaly identification task target is denoted as the negative correlation index map, and the spectral index map in the relevant spectral index map that is positively correlated with the current anomaly identification task target is denoted as the positive correlation index map;

[0027] The difference between the constant 1 and any negative correlation index map is denoted as the inverse mapping index map. The prior attention mask is obtained by multiplying each inverse mapping index map element by element with each positive correlation index map.

[0028] Furthermore, the conversion to a priori attention weight map includes:

[0029] The prior attention mask is input into a convolutional layer with a kernel length and width of 1 to obtain the convolutional output result. The convolutional output result is then normalized to obtain the prior attention weight map.

[0030] Furthermore, the abnormality-sensitive feature map is as follows:

[0031] ,

[0032] in, This represents the abnormality sensitive feature map. This represents a learnable prior attention intensity control parameter. This represents the prior attention weight map. This represents element-wise multiplication. This represents the fusion weighted feature.

[0033] The advantages of this invention compared to the prior art are:

[0034] This invention abandons single-scale or fixed-weighted multi-scale feature processing modes and constructs a dynamic, content-aware feature learning and decision-making framework. First, it comprehensively captures diverse receptive field features, ranging from local details to global context, through parallel convolutional branches with different dilation rates. Then, by performing channel compression and activation, as well as channel-dimensional pooling operations on these diverse receptive field features, it implements a lightweight "channel-spatial dual-gating mechanism." This mechanism adaptively assigns appropriate weights to feature maps at different branches and locations. Furthermore, it integrates prior knowledge of the geographical scene, applying spatial attention at the feature level to further suppress background region responses unrelated to anomalies, allowing the model's attention to focus on truly suspicious areas. Ultimately, this significantly improves the detection capability for multi-scale anomalies, especially small-scale, low-contrast anomalies, thereby enhancing the accuracy of low-altitude remote sensing anomaly identification. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart illustrating a low-altitude remote sensing anomaly identification method based on artificial intelligence, provided in Embodiment 1 of the present invention. Detailed Implementation

[0037] The overall concept of this invention is as follows:

[0038] This invention argues that the core reason for the shortcomings of existing technologies in low-altitude remote sensing anomaly identification scenarios lies in the fact that the feature extraction and fusion mechanisms of existing models are static or semi-static. These mechanisms fail to adequately adapt to the characteristics of low-altitude remote sensing images, such as the large dynamic range of anomaly scales and the variable contrast between targets and backgrounds. Furthermore, they do not effectively utilize prior geographical scene knowledge. Simple feature pyramid fusion struggles to enhance the features of small targets without introducing excessive background noise or compromising the positioning accuracy of large targets.

[0039] To address this, this invention proposes a low-altitude remote sensing anomaly identification method based on "multi-branch adaptive receptive field and spatial context gating fusion." The core improvement of this method is to abandon single-scale or fixed-weighted multi-scale feature processing and instead design a dynamic, content-aware feature learning and decision-making framework. First, by constructing parallel convolutional branches with different dilation rates, diverse receptive field features, ranging from local details to global context, are explicitly captured, ensuring that anomaly feature primitives at different scales can be effectively extracted. Next, a lightweight "channel-spatial dual-path gating mechanism" is introduced, which adaptively assigns appropriate weights to feature maps of different branches and locations based on the characteristics of the current input image patch. Finally, deep fusion of geographical scene priors is applied at the feature level, further suppressing background region responses unrelated to anomalies, allowing the model's attention to focus on truly suspicious regions.

[0040] Through this series of progressive and adaptive adjustments, the present invention can significantly improve the recall rate and recognition accuracy of multi-scale anomalies in complex backgrounds, especially improving the detection capability of small, dark, and weak-contrast anomalies, thereby effectively solving the key shortcomings in the prior art and improving the accuracy of low-altitude remote sensing anomaly recognition.

[0041] To further illustrate the technical solution of the present invention, specific embodiments are described below.

[0042] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a particular feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. Furthermore, a particular feature, structure, or characteristic in one or more embodiments may be combined in any suitable form, and the terms "comprising," "including," "having," and variations thereof mean "including, but not limited to," unless otherwise specifically emphasized.

[0043] It should be understood that the sequence number of each step in the following embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0044] Method Implementation Examples:

[0045] See Figure 1 This is a flowchart illustrating a low-altitude remote sensing anomaly identification method based on artificial intelligence provided in Embodiment 1 of the present invention. Figure 1 As shown, the method may include the following steps:

[0046] S101, construct an input feature map from any low-altitude remote sensing image, and calculate a preset type of spectral index map of the input feature map.

[0047] The purpose of this step is to standardize the raw low-altitude remote sensing imagery and convert it into a format suitable for input to deep learning models, providing high-quality, uniformly sized input data for subsequent core recognition steps. The specific steps are as follows:

[0048] 1. Data Acquisition and Registration: Continuously acquire low-altitude remote sensing images of the target area from UAVs or other low-altitude remote sensing platforms. These images may include raw multispectral or hyperspectral images (preferably including red, green, blue, and near-infrared bands). If multi-temporal analysis is involved, specialized software (such as ENVI) should be used to perform geometrical fine correction and registration on multiple images to ensure strict pixel alignment and eliminate geometric distortions caused by flight attitude and terrain undulations.

[0049] 2. Radiometric Calibration and Atmospheric Correction: The raw digital quantization (DN) values ​​of the image are converted into surface reflectance to eliminate the effects of atmospheric scattering and absorption, and to make data acquired at different times and from different sensors comparable. This process can be performed using models (such as 6S, FLAASH) or based on field measurement data.

[0050] 3. Image Enhancement and Denoising: To highlight the differences between potentially anomalous areas and the background, selective image enhancement can be performed. For example, histogram equalization or contrast stretching can be applied to the panchromatic or true-color composite bands. To suppress sensor noise or thin cloud interference, adaptive median filtering or wavelet thresholding denoising methods can be used.

[0051] 4. Multispectral Feature Calculation: Combining prior knowledge of geographical and mineral scenarios, a spectral index map is calculated to aid in anomaly identification, serving as auxiliary input or attention guidance for subsequent networks. The spectral index map may include:

[0052] i. Normalized Difference Vegetation Index (NDVI) map: used to distinguish between vegetated and non-vegetated areas. In mineral development monitoring, illegal mining areas usually show a sharp drop in NDVI.

[0053] ii. Normalized Difference Water Index (NDWI) plot: used to extract water bodies and help identify water anomalies caused by tailings dam leaks or illegal discharges.

[0054] iii. Bare Soil Index (BSI) map: used to enhance the identification of bare soil and rock areas, and helps to discover new engineering sites.

[0055] Furthermore, other spectral index maps can be calculated based on anomaly identification needs, such as the Normalized Difference Built-up Index (NDBI) map, etc.

[0056] 5. Block Processing: The processed multi-band image (including original bands and derived exponential bands) is cut into image blocks of fixed size (e.g., 512x512 pixels) with some overlap (e.g., 128 pixels). Overlapping cutting prevents anomalous targets from being cut off at block boundaries and reconstructs the entire image result during prediction through sliding window and weighted fusion. For each image block, standardization is performed (e.g., subtracting the mean and dividing by the standard deviation for each channel), and it is organized into batch tensors, a core data structure in deep learning used for efficient parallel processing of multiple samples, to serve as input for subsequent steps.

[0057] Through the aforementioned series of standardized preprocessing operations, geometrically and radiometrically consistent image patches containing rich spectral context information are obtained. Each image patch can serve as an input feature map in subsequent steps, thus providing a stable and reliable input foundation for the subsequent deep learning model. Corresponding to the block-based processing, the spectral index map is the spectral index map corresponding to each image patch. Alternatively, in other embodiments, the input feature map can be constructed directly from any low-altitude remote sensing image without block-based processing; in this case, the spectral index map is the spectral index map corresponding to the selected low-altitude remote sensing image.

[0058] S102, construct a preset number of dilated convolutional branches with the same kernel size but different dilation rates, input the input feature map into each dilated convolutional branch to obtain a preset number of output feature maps, and concatenate all output feature maps according to the channel dimension to obtain multi-receptive field features.

[0059] After acquiring and preprocessing low-altitude remote sensing images in the previous step, in order to solve the problem of multi-scale anomaly feature extraction and fusion in complex backgrounds, this step first generates diversified features covering "details", "local context" and "global context" by constructing parallel branches with different dilated convolutions at the feature extraction level, ensuring that everything from pixel-level subtle changes to regional structural patterns can be captured.

[0060] Therefore, to simultaneously capture the detailed texture of anomalous targets (crucial for small targets) and the surrounding environmental context (crucial for large targets and interference resistance), this step abandons single-scale convolution and instead opts for a three-branch parallel structure. Each branch is a dilated convolution branch using the same-sized kernel but with different dilation rates, thereby obtaining differentiated receptive fields without increasing parameters or sacrificing resolution. Specifically, this includes:

[0061] 1) Branch 1 (Detail Branch): Uses a standard 1x1 convolution or a 3x3 dilated convolution with a dilation rate of 1. This branch has the smallest receptive field and focuses on extracting local details, edge and texture features within the pixel neighborhood, which is crucial for identifying small-scale anomalies (such as a single small illegal facility or the origin of a ground fissure).

[0062] 2) Branch 2 (Local Context Branch): Uses a 3x3 dilated convolution with a dilation rate of 3. This branch has a medium receptive field and can perceive local structural patterns and relationships between adjacent features within a certain range around the target. It helps to distinguish targets that are similar to the background texture but have abnormal structures and initially suppresses isolated noise points.

[0063] 3) Branch 3 (Global Context Branch): Uses a 3x3 dilated convolution with a dilation rate of 6. This branch has the largest receptive field, which can integrate the semantic information and global layout of a larger area of ​​the image patch. It helps to understand the overall shape of large-scale anomalies (such as large mining faces and landslides) and their relationship with the whole scene, and provides richer localization context for small targets.

[0064] In this embodiment, the dilated convolution branch is preferably three. It is easy to understand that in other embodiments, the number of branches can be set according to the scale complexity of the anomaly detection region, such as two or four.

[0065] Based on the dilated convolution branches set above, let the input feature map be... Where H, W, and C are the height, width, and total number of channels of the input feature map, respectively, the operations of the above three branches on the input feature map can be formally expressed as:

[0066] or

[0067]

[0068]

[0069] in, This represents the output feature map obtained after the first branch. This represents the output feature map obtained after the second branch. This represents the output feature map obtained after the third branch. In other words, the height and width of the output feature map obtained from each branch are the same as the height and width of the input feature map, only the number of channels is different. This represents a standard convolution with a kernel size of 1*1. This represents a dilated convolution with a kernel size of 3*3 and an inflation rate of 1. This represents a dilated convolution with a kernel size of 3*3 and an inflation rate of 3. This represents a dilated convolution with a kernel size of 3*3 and an inflation rate of 6. The total number of output channels for the output feature map output by each branch is preferred in this embodiment. Ideally, each branch should output the same number of output channels for its output feature map. ,and Not greater than Subsequently, the output feature maps obtained from the three branches are concatenated along the channel dimension to obtain preliminary multi-receptive field features: .

[0070] This step explicitly generates diverse features covering different spatial ranges with negligible additional computational cost, providing rich raw materials for subsequent processing and ensuring that the information base of anomalies at different scales is not lost.

[0071] S103, channel compression and activation are performed on the multi-receptive field features to obtain a channel attention weight vector with the same number of dimensions as the multi-receptive field channels. The channel attention weight vector is then multiplied with the multi-receptive field features channel by channel to obtain the channel weighted features.

[0072] While the previous step yielded multi-receptive field features, directly concatenating features from different branches is unreasonable because different branches contribute differently to different input content. For example, for small target regions containing a lot of detail, the detail branch should have a higher weight; for large, uniform regions, the global context branch may be more reliable. Therefore, simply concatenating or adding these heterogeneous features introduces noise and increases computational burden. This embodiment therefore designs a lightweight "channel-spatial dual-gating mechanism." This mechanism can dynamically learn channel and spatial features based on the content of the input feature map and assign appropriate fusion weights to each spatial location of each branch, achieving adaptive optimization fusion of features, suppressing irrelevant responses, and enhancing relevant features.

[0073] Therefore, this step first optimizes the channel weights in the aforementioned "channel-space dual-gating mechanism" to improve the multi-receptive field features and obtain channel-weighted features. Specifically, the multi-receptive field features are first compressed and activated to achieve channel attention construction, thereby obtaining a channel attention weight vector with the same number of dimensions as the multi-receptive field channels, including:

[0074] Channel compression:

[0075] First, regarding the multi-receptor field feature Channel compression is performed based on multi-receptor field characteristics. Each channel in the compression process is performed separately. Essentially, this channel compression process extracts a global feature summary of the currently compressed channel. Therefore, in this embodiment, the preferred implementation is to aggregate spatial information through a global average pooling (GAP) process to obtain multi-receptive-field features. Average pooling result for any channel in the dataset:

[0076]

[0077] in, Indicating multiple receptive fields The average pooling result of the c-th channel, which is also the global feature summary extraction result of the c-th channel. Indicates the height of the output feature map. This indicates the width of the output feature map. Indicating multiple receptive fields The feature value with coordinates [i,j] in the c-th channel.

[0078] Following the average pooling method described above, the multi-receptive-field features are processed sequentially. Each channel is averaged to obtain the corresponding average pooling result. These average pooling results from all channels are then arranged in channel order to form a compressed vector with the same dimension as the multi-receptive field channels, specifically, a length of... The channel compression result vector, where This represents the total number of dilated convolution branches, which is set to 3 in this embodiment. This represents the number of channels in the output feature map produced by each dilated convolution branch.

[0079] Channel incentives:

[0080] The role of channel excitation is to assess the importance of the current channel. Although the channel compression process described above has obtained the aggregated spatial information of each channel (i.e., a global feature summary), assessing the importance of the current channel cannot solely focus on the aggregated spatial information itself; further consideration and trade-offs are required through a neural network. Therefore, this embodiment preferably generates the attention weight vector using two fully connected layers and a nonlinear activation function:

[0081]

[0082] in, This represents the channel attention weight vector. This represents the compression result vector of the channel. Let the first learnable parameter matrix be of size . ,in This represents the total number of dilated convolution branches; in this embodiment, it is set to 3. This represents the number of channels in the output feature map of each dilated convolution branch. This represents the reduction ratio, with a preferred value of 16. (This parameter matrix...) It can be understood as an information filter that filters out longer summaries (from...) (Dimension) is compressed into a more refined intermediate representation ( This dimensionality reduction process can reduce computational costs and help the model capture high-level correlations between channels. This represents a filtering function, whose purpose is to determine... The calculation process checks if each element in the result is less than 0, and replaces the corresponding element with 0 if it is. This can be understood as introducing a kind of selective memory, thus retaining only positive or important information and filtering out noise. This represents the second learnable parameter matrix, whose size is... It will refine the intermediate representation ( (Dimension) Restored back to the original dimension ( (Dimension), but by this point the information is already the result of thought and consideration. This represents the normalization function, which will The output values ​​are mapped to values ​​between 0 and 1. The normalized values ​​of these values ​​are the importance weights of the corresponding channels obtained through the final learning process. The closer the normalized value of any channel is to 1, the more important that channel is in the current task.

[0083] After that, and By multiplying each channel sequentially, channel weights can be optimized to improve multi-receptive-field features and obtain channel-weighted features. .

[0084] S104. Max pooling and average pooling are performed on the channel-weighted features according to the channel dimension. The max pooling result and the average pooling result are concatenated and then a spatial attention map is generated through a standard convolutional layer of a preset size. The spatial attention map is multiplied element-wise with the channel-weighted features to obtain the fused weighted features.

[0085] Following the previous step, this step continues to optimize the spatial weights in the aforementioned "channel-space dual-path gating mechanism" to further improve the fused weighted features based on the channel weighted features.

[0086] Specifically, the channel weighting features Max pooling and average pooling are performed separately along the channel dimension. The max pooling result and the average pooling result are concatenated and then passed through a standard convolutional layer of a preset size to generate a spatial attention map. The purpose of this process is to evaluate which part of the content in any channel is most critical.

[0087]

[0088]

[0089] in, This represents the spatial attention map. This represents the normalization function, which normalizes a standard convolutional layer of a preset size. The output values ​​are mapped to values ​​between 0 and 1. The term refers to a standard convolutional layer of a preset size. In this embodiment, it is preferred that the length and width of the preset-size standard convolutional layer are both equal to 7. Using a 7x7 convolutional kernel is an effective empirical choice. In other embodiments, other sizes of convolutional kernels can also be used. The design principle of this convolutional kernel is to use a sufficiently large odd-sized convolutional kernel to capture the relationship between each spatial location and its surrounding larger neighborhood, thereby determining whether the importance of the location has spatial continuity (for example, anomalous regions usually appear in patches rather than isolated noise points). Odd-sized kernels facilitate the definition of the center pixel and keep the spatial size of the feature map unchanged after padding. This indicates a max pooling operation. This indicates an average pooling operation;

[0090] as well as, This indicates the fusion weighted features. This indicates element-wise multiplication.

[0091] The principle behind obtaining the spatial attention map described above is roughly as follows:

[0092] First, the channel weighting features We can observe each position from two perspectives: the first perspective is the most prominent value of any position in all channels (corresponding to max pooling operation), and the second perspective is the average performance of any position in all channels (corresponding to average pooling operation). By observing these two perspectives, we can obtain two observation result images.

[0093] Then, by combining these two observation images, we obtain a dual-view map that includes the "most prominent feature" and the "average feature".

[0094] Next, a larger convolutional kernel (such as the preferred 7x7 kernel in this embodiment) is used to scan and analyze the dual-view map. This large convolutional kernel can perceive the relationship between each point and its surrounding larger area, thereby determining whether the importance of a location has spatial consistency (for example, an outlier is usually related to a region around it, rather than isolated noise).

[0095] Finally, also through The function generates a feature map of the same size as the input feature map space. Spatial attention map with values ​​between 0 and 1 at various internal locations. The value at each location in the graph represents the spatial importance weight of that pixel location.

[0096] By learning the attention weights, this step achieves adaptive fusion of multi-branch features, effectively suppressing irrelevant or noisy features, strengthening the feature representation most relevant to the current recognition task, and improving the discriminative power of the features.

[0097] S105, select relevant spectral index maps from the preset categories of spectral index maps of the input feature map according to the current anomaly recognition task objective, construct a prior attention mask based on the relevant spectral index maps and convert it into a prior attention weight map, and enhance the fused weighted features based on the prior attention weight map to obtain an anomaly sensitive feature map.

[0098] In specific applications of low-altitude remote sensing anomaly identification, the occurrence of anomalies often has geospatial relevance. For example, new engineering activities typically occur in bare land (high BSI) rather than dense forests (high NDVI); water anomalies (pollution, eutrophication) are only likely to occur within or at the edge of water bodies (high NDWI). Therefore, the spectral indices calculated in step S101 of this embodiment provide strong prior knowledge for this purpose. This step injects this prior information into the network in the form of attention, forcing the model to focus its limited attention resources on the "highly suspicious" areas indicated by the prior knowledge, thereby performing secondary refinement at the feature level and further improving the model's discrimination ability and robustness in complex geographical contexts.

[0099] Specifically, the various spectral index maps corresponding to the current input feature map obtained in step S101, namely NDVI, NDWI, BSI, and NDBI index maps, are respectively denoted as... , , as well as Then, spectral index maps related to the current anomaly detection task target are selected from these maps. Based on the correlation between each relevant spectral index map and the current anomaly detection task target, it is labeled as a negative correlation index map or a positive correlation index map. Then, a prior attention mask is constructed based on each negative correlation index map and the positive correlation index map. For example, for detecting anomalies in mineral development, a mask can be constructed to highlight areas of bare land with sparse vegetation: ;for For each value, the closer it is to 1, the higher the prior probability that the location belongs to a highly suspicious area.

[0100] Prior attention mask With fusion weighted features After spatial dimension alignment, the prior attention mask is transformed into a prior attention weight map through a learnable transformation (usually a 1*1 convolutional layer with a sigmoid normalization function). This transformation allows the network to soften and adjust prior information to better adapt to the data.

[0101]

[0102] in, This represents the prior attention weight map. Represents the normalization function. This indicates a convolutional layer with a kernel size of 3*3. This represents the prior attention mask.

[0103] Finally, this prior attention weight map is used. Weighted features of fusion Final modulation is performed to obtain an enhanced anomaly-sensitive feature map. :

[0104]

[0105] in, This represents the abnormality sensitive feature map. This represents a learnable prior attention strength control parameter used to control the strength of prior attention (typically, it can be...). Initializing it to a small positive number (e.g., 0.1, 0.5, or 1.0), this "1+" form ensures that even with low prior attention weights, fundamental feature information is not completely erased. This represents the prior attention weight map. This represents element-wise multiplication. This represents the fusion weighted feature.

[0106] This step integrates spectral index information into the deep learning framework, achieving a combination of data-driven and knowledge-driven approaches. This allows the network to concentrate its computational resources on geographically more likely areas to be anomalies, enhancing its response to weak signals in real suspicious areas and improving the model's discrimination accuracy and robustness in complex scenarios.

[0107] Thus, through the progressive processing of the aforementioned steps, this embodiment transforms the input feature map into an anomaly-sensitive feature map rich in multi-scale information, adaptively optimized and fused, and guided by scene priors for focusing. This prepares the ground for the final anomaly identification.

[0108] S106, Based on the anomaly sensitive feature map, complete the anomaly identification under the current anomaly identification task target.

[0109] After obtaining the abnormally sensitive feature map Then, the abnormal sensitive feature map can be generated. The anomaly is identified by inputting the data into a task-specific detection head.

[0110] For object detection tasks (such as outlining illegal mining sites), a lightweight RPN (Region Proposal Network) can be used in conjunction with classification and regression subnetworks. For pixel-level segmentation tasks (such as delineating ground fissures or pollution areas), a simple decoder (consisting of several transposed convolutional layers) can be used to upsample the feature maps back to the input image size. To ensure recognition accuracy, the following operations may also be included in the recognition process:

[0111] 1. Confidence threshold filtering: Set a confidence threshold (e.g., 0.5) according to the application scenario to filter out predictions with too low confidence.

[0112] 2. Non-maximum suppression (NMS): For target detection results, NMS is applied to eliminate overlapping boxes for the same anomalous target and retain the predicted box with the highest confidence.

[0113] 3. Result Mapping and Aggregation: Since overlapping blockization is used in step S101 of this embodiment, it is necessary to map the prediction result of each image block back to the original full-image coordinate system according to its position, and to perform a weighted average on the prediction results of the overlapping areas (such as based on the distance between the prediction position and the block center) to generate a smooth and consistent full-image prediction result.

[0114] 4. Visualization and Output: The final anomaly identification results (such as bounding boxes with category labels and color segmentation masks) are overlaid with the original low-altitude remote sensing imagery to generate intuitive thematic visualization maps. Simultaneously, structured result data, such as GeoJSON vector files, can be output, containing information such as the location, area, type, and confidence level of anomalies, for further analysis by Geographic Information Systems (GIS) or for use by decision support systems.

[0115] This invention abandons single-scale or fixed-weighted multi-scale feature processing modes and constructs a dynamic, content-aware feature learning and decision-making framework. First, it comprehensively captures diverse receptive field features, ranging from local details to global context, through parallel convolutional branches with different dilation rates. Then, by performing channel compression and activation, as well as channel-dimensional pooling operations on these diverse receptive field features, a lightweight "channel-spatial dual-gating mechanism" is implemented. This mechanism adaptively assigns appropriate weights to feature maps at different branches and locations. Furthermore, it integrates prior knowledge of the geographical scene, applying spatial attention at the feature level to further suppress background region responses unrelated to anomalies, allowing the model's attention to focus on truly suspicious areas. This invention significantly improves the detection capability of multi-scale anomalies, especially small-scale, low-contrast anomalies, thereby enhancing the accuracy of low-altitude remote sensing anomaly identification.

[0116] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An artificial intelligence-based low-altitude remote sensing anomaly identification method, characterized in that, The method includes: Construct an input feature map from any low-altitude remote sensing image, and calculate a preset number of spectral index maps for the input feature map; Construct a preset number of dilated convolutional branches with the same kernel size but different dilation rates. Input the input feature map into each dilated convolutional branch to obtain a preset number of output feature maps. Concatenate all output feature maps according to the channel dimension to obtain multi-receptive field features. Channel compression and activation are performed on the multi-receptive field features to obtain a channel attention weight vector with the same number of dimensions as the multi-receptive field channels. The channel attention weight vector is then multiplied with the multi-receptive field features channel by channel to obtain the channel weighted features. Max pooling and average pooling are performed on the channel-weighted features according to the channel dimension. The max pooling result and the average pooling result are concatenated and then passed through a standard convolutional layer of preset size to generate a spatial attention map. The spatial attention map is multiplied element-wise with the channel-weighted features to obtain the fused weighted features. Based on the current anomaly detection task objective, relevant spectral index maps are selected from the preset categories of spectral index maps of the input feature map. A prior attention mask is constructed based on the relevant spectral index maps and converted into a prior attention weight map. The fused weighted features are enhanced based on the prior attention weight map to obtain an anomaly sensitive feature map. Based on the anomaly sensitivity feature map, complete the anomaly identification under the current anomaly identification task objective; The preset spectral index maps include a normalized difference vegetation index map, a normalized difference water index map, a normalized difference building index map, and a bare surface index map. The construction of the prior attention mask based on the relevant spectral index map includes: The spectral index map in the relevant spectral index map that is negatively correlated with the current anomaly identification task target is denoted as the negative correlation index map, and the spectral index map in the relevant spectral index map that is positively correlated with the current anomaly identification task target is denoted as the positive correlation index map; The difference between the constant 1 and any negative correlation index map is denoted as the inverse mapping index map. The prior attention mask is obtained by multiplying each inverse mapping index map element by element with each positive correlation index map. 2.The low-altitude remote sensing anomaly identification method based on artificial intelligence according to claim 1, characterized in that, The construction of the input feature map includes: The low-altitude remote sensing image is segmented to obtain multiple image blocks with a first preset number of pixels. There is an overlapping area with a second preset number of pixels between adjacent image blocks. Any image block is used as the input feature map. 3.The low-altitude remote sensing anomaly identification method based on artificial intelligence according to claim 1, characterized in that, The process of channel compression and activation of multi-receptive field features to obtain a channel attention weight vector with the same number of dimensions as the multi-receptive field channels includes: The average pooling is performed sequentially on each channel of the multi-receptive field feature to obtain the corresponding average pooling result. The average pooling results of all channels are then arranged in the order of each channel to form a channel compression result vector with the same number of dimensions as the multi-receptive field channels. Calculate the channel attention weight vector based on the channel compression result vector: ， in, This represents the channel attention weight vector. This represents the compression result vector of the channel. Let the first learnable parameter matrix be of size . ,in This represents the total number of dilated convolution branches. This represents the number of channels in the output feature map of each dilated convolution branch. Indicates the reduction ratio. This represents a filtering function, whose purpose is to determine... Check if each element in the calculation result is less than 0, and replace the corresponding element with 0 if it is. This represents the second learnable parameter matrix, whose size is... , This represents the normalization function.

4. The low-altitude remote sensing anomaly identification method based on artificial intelligence according to claim 1, characterized in that, The step of concatenating the max pooling result and the average pooling result and then generating a spatial attention map through a standard convolutional layer of a preset size includes: The max pooling result and the average pooling result are concatenated and then input into the standard convolutional layer of the preset size. The output of the standard convolutional layer is normalized to obtain the spatial attention map.

5. The method for identifying low-altitude remote sensing anomalies based on artificial intelligence according to claim 1 or 4, characterized in that, The kernels of the standard convolutional layer of the preset size have equal length and width, and are odd numbers.

6. The method for identifying low-altitude remote sensing anomalies based on artificial intelligence according to claim 1, characterized in that, The conversion to a priori attention weight map includes: The prior attention mask is input into a convolutional layer with a kernel length and width of 1 to obtain the convolutional output result. The convolutional output result is then normalized to obtain the prior attention weight map.

7. The low-altitude remote sensing anomaly identification method based on artificial intelligence according to claim 1, characterized in that, The abnormality-sensitive feature map is as follows: ， in, This represents the abnormality sensitive feature map. This represents a learnable prior attention intensity control parameter. This represents the prior attention weight map. This represents element-wise multiplication. This represents the fusion weighted feature.

Images