A low-resolvable target positioning method and system for space-based multi-spectral remote sensing images

Abstract

The application relates to a low-resolvable target positioning method and system for space-based multi-spectral remote sensing images, and relates to the technical field of multi-spectral remote sensing image target positioning.The technical points are as follows: the space-based multi-spectral remote sensing image is preprocessed, and a candidate region is generated through abnormality detection; a cross-spectrum comparable positioning evidence vector is constructed for the candidate region, a multi-dimensional consistency score is calculated, and a total score is obtained through weighting, a gate value is generated through Sigmoid soft gate mapping; the multi-spectral segment collaborative or single-spectral segment dominant positioning is adaptively selected according to the gate value, a target center position is obtained through weighted fusion, the gate value is synchronously output as a reliability index, and an end-to-end training process based on Landsat8 images is designed.The system corresponds to six serial modules, and built-in function units in the modules realize the above method.The application improves the positioning stability and robustness in a complex scene, reduces the calculation complexity, the modules are reusable, engineering implementation is strong, and the application is suitable for precise positioning of space-based multi-spectral remote sensing low-resolvable targets.

Description

Technical Field

[0001] This invention relates to the field of target localization technology for multispectral remote sensing images, specifically to a method and system for localizing low-resolution targets in space-based multispectral remote sensing images. Background Technology

[0002] The field of low-resolution target localization in space-based multispectral remote sensing imagery, to which this invention pertains, relies on the development of space-based optical remote sensing technology to achieve practical applications. Among them, the Landsat 8 satellite, with its 11 spectral band observation capabilities provided by the OLI / TIRS payload, has the advantages of global coverage, long-term continuous observation, and standardized calibration system, becoming an important data foundation for multispectral remote sensing target detection and localization, and providing key data support for the observation and analysis of low-resolution targets on the Earth's surface.

[0003] In space-based remote sensing applications, many targets exhibit low-resolution features due to their spatial scale being close to or smaller than the pixel scale. These features manifest as local weak anomalies or point / small patch responses in images, easily confused with complex surface backgrounds, image noise, thin clouds / smoke interference, etc., posing a fundamental challenge to accurate positioning. On the one hand, the mixed pixel effect causes the target signal to superimpose with the background within the same pixel, and the target center does not necessarily correspond to the image's maximum response pixel. Simply relying on peak pixels or hard thresholds can easily lead to systematic positioning shifts. On the other hand, multi-spectral observations are not equivalent. The signal-to-noise ratio, saturation characteristics, scattering effects, and point spread / resolution differ across different bands. Furthermore, even multi-spectral products from the same satellite may have cross-spectral registration residuals, resulting in the same target's response not being completely consistent across different spectral bands.

[0004] Existing localization methods for low-resolution targets can be mainly divided into three categories: First, single-spectral-band localization methods, which obtain candidate pixels through anomaly detection and then output the target position using empirical rules such as centroid and fitting. This method is simple to implement but is easily affected by background false alarms, threshold sensitivity, and mixed pixel effects, resulting in poor localization robustness. Second, fixed multi-spectral-band fusion localization methods, which use dual-stream / multi-stream networks or feature fusion structures to fuse multi-spectral-band information in an attempt to improve detection and localization capabilities. However, these methods assume that multi-spectral-band information can be directly fused without considering the reliability differences of spectral evidence. Third, fusion-before-localization methods, which first fuse or super-resolution multi-spectral-band images to improve spatial details before performing localization operations. This method has high requirements for spectral fidelity and geometric consistency, and cross-spectral registration errors can easily introduce new localization biases.

[0005] The aforementioned existing methods have limitations in multi-spectral low-resolution target localization scenarios. Most methods lack explicit mechanisms for handling differences in the reliability of spectral evidence and conflicts between cross-spectral evidence. Under conditions of thin clouds, noise, saturation, or registration residuals, the fusion process may introduce unreliable evidence into the localization process, leading to localization offset. The localization results only output coordinate information and lack a mechanism for characterizing localization credibility / uncertainty, making it impossible for downstream systems to screen results based on reliability or conduct manual verification. Cross-spectral collaboration is treated as a fixed process flow, and the steps of "whether to collaborate and how strong the collaboration is" are not designed as interpretable and decision-making steps, making it difficult to balance localization accuracy and robustness under complex observation conditions. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for locating low-resolution targets in space-based multispectral remote sensing images, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for locating low-resolution targets in space-based multispectral remote sensing imagery, comprising the following steps:

[0008] S1. Perform geometric alignment and unified spatial grid resampling on the input space-based multi-spectral remote sensing image to obtain multi-spectral images with pixel-level correspondence between spectral bands. The space-based multi-spectral remote sensing image includes visible light, near-infrared, short-wave infrared and / or thermal infrared bands.

[0009] S2. Perform rapid anomaly response detection on the processed multi-spectral images, using a single-spectral anomaly detection strategy, a multi-spectral cross-spectral co-occurrence anomaly detection strategy, or a combination of both, combined with connected component analysis or a fixed-size sliding window to aggregate anomaly pixels, to generate a candidate region set based on the principle of high recall.

[0010] S3. For each candidate region, construct local background statistics on each spectral band and generate a normalized anomaly response map. Based on the normalized anomaly response map, extract intensity evidence, contrast evidence, morphological evidence and spatial evidence to form a cross-spectral comparable localization evidence vector for each spectral band.

[0011] S4. Based on the localization evidence vectors of each spectral band, calculate the spatial consistency score, morphological consistency score and energy consistency score respectively, obtain the total consistency score of the candidate region by weighted summation, and then convert the total consistency score into a gate value in the [0,1] interval through Sigmoid soft gating mapping;

[0012] S5. Perform adaptive positioning based on the threshold value: when the threshold value approaches 1, the target center position is estimated using a multi-spectral band cooperative positioning method with spectral band reliability weighting; when the threshold value approaches 0, the most reliable spectral band is selected to perform a single-spectral band dominant positioning method to estimate the target center position; the two positioning results are weighted and fused by the threshold value to obtain the final target center position of the candidate region.

[0013] S6. Output the target center position of each candidate region, and simultaneously output the gating value as a positioning reliability index.

[0014] Furthermore, the single-spectral-band anomaly detection strategy in S2 is as follows: in the spectral band sensitive to the target... An anomaly indicator function is constructed above, which is used when the difference between a pixel response and the mean value of the local window background centered on that pixel exceeds a preset threshold. When a pixel is identified as an anomalous pixel, the anomalous indicator function outputs 1; otherwise, it outputs 0. The multi-spectral cross-spectral co-occurrence anomaly detection strategy is as follows: binary anomaly indicator maps are generated for multiple spectral bands, and a joint anomaly response map is obtained by weighted summation using logical OR. When the joint response value is greater than or equal to 1, the pixel is identified as an anomalous pixel. The method for aggregating anomalous pixels is to perform connected component analysis on the anomalous pixels, or to use a fixed-size sliding window to select scattered anomalous pixels, forming a candidate region as a local window. .

[0015] Furthermore, in S3, the local background statistics are selected as the mean or median, standard deviation or median absolute deviation (MAD), and the calculation formula for the normalized anomaly response map is:

[0016]

[0017] in To avoid small constants that cause numerical instability; the image of the k-th spectral band of the n-th sample is... For each candidate window First, a local background reference is constructed on spectral band k to measure the "degree of anomaly"; and These represent the background statistics within the window. The specific extraction methods for the four types of evidence are as follows: intensity evidence is the maximum value of the normalized anomalous response within the window; contrast evidence is the difference between the intensity evidence and the median of the normalized anomalous response within the window; morphological evidence is the principal axis ratio calculated based on the area, dispersion, or principal directionality of the support domain after thresholding the normalized anomalous response; and spatial evidence is the peak position of the normalized anomalous response within the window or the weighted centroid on the anomalous support domain.

[0018] Furthermore, the spatial consistency score in S4 is calculated as follows: first, the spatial dispersion of the spatial evidence for each spectral band is calculated:

[0019]

[0020] in The mean of the spatial evidence for each spectral band is then mapped through an exponential method. Map the dispersion to a spatial consistency score in the range [0,1]. Scale parameters consistent with the order of magnitude of cross-spectral registration error; morphological consistency score The energy consistency score is obtained by performing exponential mapping after calculating the variance of the dispersion of morphological evidence for each spectral band; the energy consistency score is constructed by analyzing the order consistency or quantile stability of the contrast evidence for each spectral band. The above three types of consistency are combined into a total consistency score.

[0021]

[0022] Weighting coefficient of total consistency score , , satisfy .

[0023] Furthermore, the calculation formula for the Sigmoid soft-gated mapping in S4 is as follows:

[0024]

[0025] in For the Sigmoid function, The threshold value is used for gate control. This is the gate steepness coefficient, the gate value. The closer the value is to 1, the higher the consistency of multi-spectral evidence; the closer it is to 0, the lower the consistency. The selection criterion for the most reliable spectral segment in S5 is to select the spectral segment with the greatest contrast evidence or the smallest degree of morphological abnormality.

[0026] Furthermore, the specific implementation of the multi-spectral band collaborative localization in S5 is as follows: Based on the contrast evidence of each spectral band, an adaptive weight $w_k$ is set, and the joint response within the candidate window is calculated:

[0027]

[0028] The collaborative localization results are then calculated using soft-maximization weighted centroid calculation:

[0029]

[0030] in To control the coefficient of soft maximization concentration, the formula for calculating the final target center position is:

[0031]

[0032] The results are based on the dominant localization of a single spectral band.

[0033] Furthermore, the method also includes a model training step, specifically: using Landsat8 multi-band imagery as the data source, constructing training samples with candidate windows as basic units; labeling the target center pixel coordinates for positive sample windows containing low-resolution targets, and not labeling the center position for negative sample windows composed of background or interference; constructing a joint loss function composed of localization loss and gated supervision loss, whereby the localization loss constrains the distance between the predicted center and the labeled center of the positive sample window, and the gated supervision loss constrains the gate value of the positive sample window to be too high and the gate value of the negative sample window to be too low; performing end-to-end updates on the learnable modules in evidence mapping, consistency evaluation, and gate mapping through standard backpropagation, and ensuring that the localization output is continuous to guarantee the differentiability of the training process.

[0034] Furthermore, the target center position in S6 includes pixel coordinates or latitude and longitude coordinates. The gate value serves as a positioning credibility index, used by the downstream system to screen high-credibility positioning results or to manually verify low-credibility positioning results. The entire positioning process only performs fine analysis of cross-spectral evidence construction, consistency measurement, and adaptive positioning within the candidate area, and only performs lightweight anomaly detection and candidate area generation within the entire map area, thus achieving controllable computational complexity.

[0035] A low-resolution target localization system for space-based multispectral remote sensing imagery includes an input preprocessing module, a candidate region generation module, a cross-spectral localization evidence construction module, a cross-spectral consistency measurement and gating decision module, an adaptive cooperative localization module, and an output module connected in sequence.

[0036] The input preprocessing module is used to perform geometric alignment and unified spatial grid resampling on the input space-based multi-spectral remote sensing images to obtain multi-spectral images with pixel-level correspondence between spectral bands;

[0037] The candidate region generation module is used to perform abnormal response detection on the preprocessed multi-spectral image, aggregate abnormal pixels through a preset strategy to generate a high-recall candidate region set, and transmit the candidate region data to the cross-spectral localization evidence construction module.

[0038] The cross-spectral localization evidence construction module is used to generate a normalized anomaly response map in each spectral band for each candidate region, extract four types of localization evidence (intensity, contrast, morphology, and space) and form an evidence vector, and transmit the evidence vector data to the cross-spectral consistency measurement and gating decision module.

[0039] The cross-spectral consistency measurement and gating decision module is used to calculate the consistency scores of three categories—space, morphology, and energy—based on the localization evidence vector and weight them to obtain the total consistency score. It generates a gating value through a Sigmoid soft gating mapping and transmits the total consistency score and the gating value to the adaptive cooperative localization module.

[0040] The adaptive cooperative positioning module is used to adaptively select multi-band cooperative positioning or single-band dominant positioning mode according to the gate value, obtain the final target center position through weighted fusion of the gate value, and transmit the target center position data and the gate value to the output module.

[0041] The output module is used to output the target center position and simultaneously output the gating value as a positioning reliability index. The target center position includes pixel coordinates or latitude and longitude coordinates.

[0042] Furthermore, the candidate region generation module incorporates a single-spectral-band anomaly detection unit, a multi-spectral-band cross-spectral co-occurrence anomaly detection unit, and an anomaly pixel aggregation unit to respectively implement the anomaly detection strategy and pixel aggregation method; the cross-spectral localization evidence construction module incorporates a background statistics calculation unit, a normalized response generation unit, and a four-type evidence extraction unit to respectively implement the calculation of local background statistics, generation of normalized anomaly response maps, and extraction of localization evidence; the cross-spectral consistency measurement and gating decision module incorporates a spatial consistency calculation unit, a morphological consistency calculation unit, an energy consistency calculation unit, and a soft-gating mapping unit to respectively implement the consistency score calculation and gating value generation; the adaptive collaborative localization module incorporates a collaborative localization unit and a single-spectral-band localization unit to respectively implement the multi-spectral-band collaborative localization and single-spectral-band dominant localization, and the system supports the end-to-end training of the joint loss function and backpropagation; the output module incorporates a coordinate transformation unit and a confidence labeling unit to realize the conversion between pixel coordinates and latitude and longitude coordinates, and to associate the gating value with the target center position for labeling output.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] This invention, through cross-spectral consistency measurement and gating decision mechanism, no longer assumes that multi-spectral information is naturally consistent. Instead, it first determines whether the evidence from multiple spectral bands points to the same target, and only performs collaborative localization when the consistency is high. When the evidence conflicts, it degenerates into single-spectral-band dominant localization. This fundamentally avoids the superposition of unreliable evidence caused by factors such as thin clouds, noise, and registration residuals, and completely solves the systematic localization offset problem of existing fixed fusion methods, significantly improving the stability of localization results in complex scenarios.

[0045] To address the issues of single-spectral-band localization being susceptible to background interference and forced fusion introducing noise, this invention employs a "cooperative-degenerate" adaptive localization strategy: when multi-spectral-band evidence is reliable, the complementary advantages of different spectral bands are fully utilized to improve localization accuracy; when conflicting evidence exists, a conservative single-spectral-band localization strategy is adopted to ensure the validity of the results. This design enables the algorithm to output stable target localization results even under harsh observation conditions such as complex terrain, thin cloud interference, and mixed pixels, demonstrating robustness far superior to single-spectral-band localization methods and fixed multi-spectral-band fusion methods.

[0046] Unlike existing technologies that only output target coordinates in a "black box" manner, this invention outputs a consistency gating value as a positioning reliability indicator simultaneously with the target location. This indicator directly reflects the reliability of the positioning result. Downstream systems or manual decision-making modules can filter high-reliability results and verify low-reliability results based on the reliability, upgrading the positioning result from a "single coordinate" to a combined output of "coordinates + reliability," significantly improving the actual usability of the positioning result and the controllability of the overall system.

[0047] The cross-spectral positioning evidence representation and cross-spectral consistency measurement mechanism constructed in the positioning process of this invention are independent intermediate modules that do not depend on specific low-resolution target types. They can be directly used as basic support modules for subsequent remote sensing-related research and system development, such as positioning uncertainty modeling, temporal remote sensing image consistency analysis, and multi-source remote sensing data fusion. Compared with existing methods that only optimize single positioning accuracy, this invention has stronger scalability and technology reuse value.

[0048] This invention confines the highly complex cross-spectral evidence construction, consistency assessment, and adaptive localization operations to candidate regions. By reducing the full-image localization problem to a fine-grained analysis problem of local regions through the candidate region generation stage, it avoids performing highly complex calculations on the entire image. With the same data scale, the computational cost of this invention is on the same order of magnitude as existing multi-spectral localization methods. It requires no dedicated hardware or complex physical inversion models and can be directly deployed on general-purpose computer platforms in space-based remote sensing processing systems, making it highly feasible in engineering applications. Attached Figure Description

[0049] Figure 1 This is a flowchart of the low-resolution target localization method for space-based multispectral remote sensing imagery according to the present invention. Detailed Implementation

[0050] Please see Figure 1 This invention provides a technical solution: a method for locating low-resolution targets in space-based multispectral remote sensing imagery, comprising the following steps:

[0051] S1. Perform geometric alignment and unified spatial grid resampling on the input space-based multi-spectral remote sensing image to obtain multi-spectral images with pixel-level correspondence between spectral bands. The space-based multi-spectral remote sensing image includes visible light, near-infrared, short-wave infrared and / or thermal infrared bands.

[0052] S2. Perform rapid anomaly response detection on the processed multi-spectral images, using a single-spectral anomaly detection strategy, a multi-spectral cross-spectral co-occurrence anomaly detection strategy, or a combination of both, combined with connected component analysis or a fixed-size sliding window to aggregate anomaly pixels, to generate a candidate region set based on the principle of high recall.

[0053] The single-band anomaly detection strategy in S2 is as follows: in the spectral bands that are sensitive to the target... An anomaly indicator function is constructed above, which is used when the difference between a pixel response and the mean value of the local window background centered on that pixel exceeds a preset threshold. When a pixel is identified as an anomalous pixel, the anomalous indicator function outputs 1; otherwise, it outputs 0. The multi-spectral cross-spectral co-occurrence anomaly detection strategy is as follows: binary anomaly indicator maps are generated for multiple spectral bands, and a joint anomaly response map is obtained by weighted summation using logical OR. When the joint response value is greater than or equal to 1, the pixel is identified as an anomalous pixel. The method for aggregating anomalous pixels is to perform connected component analysis on the anomalous pixels, or to use a fixed-size sliding window to select scattered anomalous pixels, and the resulting candidate region is the local window. ;

[0054] S3. For each candidate region, construct local background statistics on each spectral band and generate a normalized anomaly response map. Based on the normalized anomaly response map, extract intensity evidence, contrast evidence, morphological evidence and spatial evidence to form a cross-spectral comparable localization evidence vector for each spectral band.

[0055] In S3, the local background statistics are selected as the mean or median, standard deviation or median absolute deviation (MAD). The formula for calculating the normalized anomaly response plot is:

[0056]

[0057] in To avoid small constants that cause numerical instability; the image of the k-th spectral band of the n-th sample is... For each candidate window First, a local background reference is constructed on spectral band k to measure the "degree of anomaly"; and These represent the background statistics within the window. The specific extraction methods for the four types of evidence are as follows: intensity evidence is the maximum value of the normalized abnormal response within the window; contrast evidence is the difference between the intensity evidence and the median of the normalized abnormal response within the window; morphological evidence is the principal axis ratio calculated based on the area, dispersion, or principal directionality of the support domain after thresholding the normalized abnormal response; and spatial evidence is the peak position of the normalized abnormal response within the window or the weighted centroid on the abnormal support domain.

[0058] S4. Based on the localization evidence vectors of each spectral band, calculate the spatial consistency score, morphological consistency score and energy consistency score respectively. Obtain the total consistency score of the candidate region by weighted summation, and then convert the total consistency score into a gate value in the [0,1] interval through Sigmoid soft gating mapping.

[0059] The spatial consistency score in S4 is calculated as follows: First, the spatial dispersion of the spatial evidence in each spectral band is calculated.

[0060]

[0061] in The mean of the spatial evidence for each spectral band is then mapped through an exponential method. Map the dispersion to a spatial consistency score in the range [0,1]. Scale parameters consistent with the order of magnitude of cross-spectral registration error; morphological consistency score The energy consistency score is obtained by performing exponential mapping after calculating the variance of the dispersion of morphological evidence for each spectral band; the energy consistency score is constructed by analyzing the order consistency or quantile stability of the contrast evidence for each spectral band. The above three types of consistency are combined into a total consistency score.

[0062]

[0063] Weighting coefficient of total consistency score , , satisfy .

[0064] The formula for calculating the Sigmoid soft-gated mapping in S4 is:

[0065]

[0066] in For the Sigmoid function, The threshold value is used for gate control. This is the gate steepness coefficient, the gate value. The closer the value is to 1, the higher the consistency of multi-spectral evidence; the closer it is to 0, the lower the consistency. The selection criterion for the most reliable spectral segment in S5 is to select the spectral segment with the greatest contrast evidence or the smallest degree of morphological abnormality.

[0067] S5. Perform adaptive positioning based on the gate value: When the gate value approaches 1, the target center position is estimated using a multi-spectral band cooperative positioning method with spectral band reliability weighting; when the gate value approaches 0, the most reliable spectral band is selected to perform a single-spectral band dominant positioning method to estimate the target center position; the two positioning results are weighted and fused by the gate value to obtain the final target center position of the candidate region.

[0068] The specific implementation of multi-spectral band collaborative localization in S5 is as follows: adaptive weights are set based on the contrast evidence of each spectral band. Calculate the joint response within the candidate window:

[0069]

[0070] The collaborative localization results are then calculated using soft-maximization weighted centroid calculation:

[0071]

[0072] in To control the coefficient of soft maximization concentration, the formula for calculating the final target center position is:

[0073]

[0074] The results are based on the dominant localization of a single spectral band.

[0075] S6. Output the target center position of each candidate region, and simultaneously output the gating value as the positioning reliability index.

[0076] In S6, the target center location includes pixel coordinates or latitude and longitude coordinates. The gate value serves as a positioning credibility index, which is used by downstream systems to screen high-credibility positioning results or to manually verify low-credibility positioning results. The entire positioning process only performs fine analysis of cross-spectral evidence construction, consistency measurement, and adaptive positioning within the candidate region. Only lightweight anomaly detection and candidate region generation are performed across the entire map, thus achieving controllable computational complexity.

[0077] The method also includes a model training step, specifically: using Landsat 8 multi-band imagery as the data source, training samples are constructed with candidate windows as the basic unit. The center pixel coordinates of the target are labeled for positive sample windows containing low-resolution targets, while the center position of negative sample windows composed of background or interference is not labeled. A joint loss function consisting of localization loss and gated supervision loss is constructed. The localization loss constrains the distance between the predicted center and the labeled center of the positive sample window, while the gated supervision loss constrains the gate value of the positive sample window to be too high and the gate value of the negative sample window to be too low. End-to-end updates are performed on the learnable modules in evidence mapping, consistency evaluation, and gate mapping through standard backpropagation, and the localization output is in a continuous form to ensure that the training process is differentiable.

[0078] A low-resolution target localization system for space-based multispectral remote sensing imagery includes an input preprocessing module, a candidate region generation module, a cross-spectral localization evidence construction module, a cross-spectral consistency measurement and gating decision module, an adaptive cooperative localization module, and an output module connected in sequence.

[0079] The input preprocessing module is used to perform geometric alignment and unified spatial grid resampling on the input space-based multi-spectral remote sensing images to obtain multi-spectral images with pixel-level correspondence between spectral bands;

[0080] The candidate region generation module is used to perform anomaly response detection on the preprocessed multi-spectral images, aggregate abnormal pixels through a preset strategy to generate a high-recall candidate region set, and transmit the candidate region data to the cross-spectral localization evidence construction module.

[0081] The cross-spectral localization evidence construction module is used to generate normalized anomaly response maps in each spectral band for each candidate region, extract four types of localization evidence (intensity, contrast, morphology, and space) and form evidence vectors, and transmit the evidence vector data to the cross-spectral consistency measurement and gating decision module.

[0082] The cross-spectral consistency measurement and gating decision module is used to calculate the consistency scores of spatial, morphological and energy categories based on the localization evidence vector and weight them to obtain the total consistency score. The module generates a gating value through a Sigmoid soft gating mapping and transmits the total consistency score and the gating value to the adaptive cooperative localization module.

[0083] The adaptive cooperative positioning module is used to adaptively select multi-band cooperative positioning or single-band dominant positioning mode according to the gate value, obtain the final target center position through weighted fusion of the gate value, and transmit the target center position data and gate value to the output module.

[0084] The output module is used to output the center position of the target, and simultaneously outputs the gating value as a positioning reliability indicator. The center position of the target includes pixel coordinates or latitude and longitude coordinates.

[0085] The candidate region generation module incorporates a single-spectral-band anomaly detection unit, a multi-spectral-band cross-spectral co-occurrence anomaly detection unit, and an anomaly pixel aggregation unit, respectively implementing anomaly detection strategies and pixel aggregation methods. The cross-spectral localization evidence construction module incorporates a background statistics calculation unit, a normalized response generation unit, and four types of evidence extraction units, respectively implementing local background statistics calculation, normalized anomaly response map generation, and localization evidence extraction. The cross-spectral consistency measurement and gating decision module incorporates a spatial consistency calculation unit, a morphological consistency calculation unit, an energy consistency calculation unit, and a soft-gating mapping unit, respectively implementing consistency score calculation and gating value generation. The adaptive collaborative localization module incorporates a collaborative localization unit and a single-spectral-band localization unit, respectively implementing multi-spectral-band collaborative localization and single-spectral-band dominant localization, and the system supports joint loss functions and end-to-end training via backpropagation. The output module incorporates a coordinate transformation unit and a confidence annotation unit, realizing the conversion between pixel coordinates and latitude and longitude coordinates, and associating the gating value with the target center position for annotation output.

[0086] Example 1, based on the above method

[0087] In space-based remote sensing imagery, low-resolution targets typically occupy only a few pixels, or even have a spatial scale smaller than the effective resolution of a single pixel. Under these conditions, performing complex cross-spectral collaborative analysis and fine-grained localization calculations directly on every pixel location across the entire image is not only computationally expensive but also easily affected by large-area background noise and irrelevant ground features. Therefore, this method introduces a candidate region generation stage before fine-grained localization to quickly filter local areas "potentially containing low-resolution targets" across the entire image, thereby significantly reducing the search space for subsequent calculations. Let the nth Landsat8 multi-band input sample be:

[0088]

[0089] Where k represents different spectral bands, x represents pixel coordinates, and Ω is the image spatial domain. The goal of the candidate region generation stage is not to accurately distinguish targets, but rather to cover as many potential target locations as possible based on the principle of "high recall priority," allowing for a certain number of false detections. This design ensures that this stage only performs a coarse screening function, while false detections will be effectively suppressed in the subsequent cross-spectral consistency gating and localization stages.

[0090] In its implementation, candidate region generation is based on the rapid detection of anomalous responses. An anomalous response refers to a response pattern in which a pixel or small neighborhood exhibits a significant deviation from its local background in one or more spectral bands. Considering that the imaging characteristics of low-resolution targets may differ in different application scenarios, this method does not limit a single spectral band to the sole criterion, but allows the use of one or a combination of the following two lightweight strategies to generate candidate regions. The first strategy is to perform anomaly detection in a single spectral band that is relatively sensitive to the target. For example, in a certain spectral band... Above, a simple anomaly indicator function is constructed by comparing pixel responses with their local background statistics:

[0091]

[0092] in This represents the mean background value within a local window centered at x. The first strategy uses a preset threshold. The core idea of ​​this strategy is to quickly capture response locations that stand out significantly from the background without relying on complex modeling. The second strategy involves simultaneously detecting spatially co-occurring anomalous points or small patches across multiple spectral bands. Specifically, binary anomaly indicator maps can be generated on several spectral bands. The joint anomaly response graph is obtained by using a logical OR weighted summation method.

[0093]

[0094] in This is the set of spectral bands participating in candidate generation. The advantage of this strategy is that it does not rely on the absolute reliability of a specific spectral band, but rather improves the coverage of the real target through "cross-spectral co-occurrence". Based on the above anomaly indication map A(x), scattered anomalous pixels are further aggregated into several candidate regions through connected component analysis or a fixed-size sliding window. Each candidate region can be represented as a local window. This can be equivalently represented as a set of candidate pixels. It's important to emphasize that no final judgment is made at this stage regarding whether the candidate region actually contains the target; therefore, false anomalies caused by noise, complex surface textures, or cloud interference are allowed. In supervised learning scenarios, a strict one-to-one correspondence is not required between the candidate region generation stage and the training labels. The target reference location or region-level annotation provided in the training samples can be represented as:

[0095]

[0096] Candidate regions only need to ensure that The location only needs to have high coverage, while for The location allows for a certain proportion of coverage redundancy. This relaxed coverage requirement ensures the robustness of the candidate generation stage and provides sufficient candidate support for subsequent fine-grained localization based on cross-spectral consistency gating. Through the above candidate region generation process, the localization problem within the original full map area is effectively reduced to a fine-grained analysis problem within several local regions. Subsequent cross-spectral evidence construction, consistency evaluation, and adaptive localization processes are all performed only within these candidate regions, thereby significantly reducing the overall computational complexity while ensuring high recall for low-discriminability targets and laying the foundation for robust localization under complex interference conditions.

[0097] A local window set is obtained during the candidate region generation stage. Subsequently, this method does not directly fuse or jointly decide on multi-spectral images. Instead, it first constructs a set of "location-related evidence" in each spectral band to form a cross-spectral comparable evidence representation. The core motivation for this design is that the response intensity, noise level, and background interference patterns of low-resolution targets often differ across different spectral bands. If fusion is performed directly without a reliability assessment, false anomalies in a certain spectral band (such as strong surface edges, cloud boundary residuals, and stripe noise) can easily be introduced into the collaborative localization process, causing localization shifts. Conversely, by first constructing local evidence within each spectral band and then comparing the consistency of this evidence at the cross-spectral level, "multi-spectral collaboration" can be transformed from a fixed operation into a controllable subsequent decision. Let the k-th spectral band image of the n-th sample be... For each candidate window First, a local background reference is constructed on spectral band k to measure the "degree of anomaly". Specifically, let... and Let the background statistics within the window be represented by mean / median and standard deviation / MAD, respectively (robust scales can be taken). Then, define the normalized outlier response plot within the window:

[0098]

[0099] in To avoid numerical instability, a small constant is used. Based on R(k)(x), this method extracts four types of evidence directly related to localization and with cross-spectral comparability in each spectral band: intensity evidence, contrast evidence, morphological evidence, and spatial evidence. First, intensity evidence is used to answer the question "Does a significant peak or anomaly exist within the spectral window?" A reproducible implementation is to take the maximum value of the anomalous response within the window:

[0100]

[0101] This is used to describe the intensity of the anomaly in that spectral band. Secondly, contrast evidence is used to characterize the prominence and robustness of the anomaly relative to the background, and can be represented by the difference between the peak and background quantiles or the mean difference, for example:

[0102]

[0103] This reduces the incomparability caused by differences in brightness levels among different ground features. Third, morphological evidence is used to answer the question, "Is the anomaly more like a dot / small patch, or more like a texture edge or banded structure?" To ensure simplicity and matching with low-resolution target features, we can first... Thresholding is performed within the window to obtain the anomaly support domain:

[0104]

[0105] in A fixed threshold or the same as can be used. An adaptive threshold related to internal statistics. Subsequently, based on... The area, dispersion, or principal orientation of the structural features are used to describe the morphology, for example, by using the second moment to obtain the principal axis ratio to distinguish between "points / clumps" and "linear edges." Finally, spatial evidence is used to answer "approximately where in the window is the anomalous peak located?", the most direct form of which is the peak location:

[0106]

[0107] Noise immunity can also be further improved by using weighted centroids, for example in Top Weighted centering. The above four types of evidence together constitute spectral segment k within the window. Internal evidence vector:

[0108]

[0109] in This represents the specific quantitative form of morphological evidence. This evidence vector does not serve a category discrimination function; its purpose is solely to provide a "comparable information source" needed for cross-spectral consistency judgment: if different spectral bands provide similar results within the same candidate window... (Spatial consistency), and exhibiting similar dotted / small patch morphology (morphological consistency), and If multiple spectral bands show significant positive values ​​(consistent contrast), then these spectral bands are more likely to point to the same true target; conversely, if only a few spectral bands show significant anomalies, or If there is a significant shift between spectral bands, or if morphological evidence shows linear edges rather than clumps, subsequent co-localization should be suppressed or degraded to avoid introducing unreliable evidence into the multi-spectral band joint localization process.

[0110] It is important to emphasize that this evidence construction process can be run under completely unsupervised conditions or used in conjunction with labels within a supervised learning framework. Training samples can provide region-level or point-level reference annotations; for example, region-level annotations can be represented as:

[0111]

[0112] In this case Its purpose is to train or evaluate whether candidate windows cover the target region, and to construct supervision signals for subsequent gating and localization modules, while the evidence vector... The definition and calculation do not rely on any prior assumptions about the target category, thus ensuring applicability to "generally low-discriminability targets".

[0113] Cross-spectral consistency measurement, gating decision and adaptive co-localization in completing candidate windows Evidence for locating each spectral band After the initial construction, the third part aims to transform "multi-spectral band coordination" from a fixed fusion operation into a controllable decision-making process: first, it determines whether multi-spectral band evidence points to the same target in space and morphology; then, based on this determination, it selects strong co-location or degenerate location, thereby avoiding erroneous fusion caused by factors such as thin clouds, saturation, noise, and cross-spectral registration residuals. To this end, this method first defines a cross-spectral consistency metric to quantify the degree of consistency of multi-spectral band evidence within the candidate window. Let the k-th spectral band be... Internal spatial evidence is (Can be obtained from the weighted centroid), morphological evidence is (For example, compactness / anisotropy obtained from weighted second moments), contrast evidence is Consistency can then be decomposed into three complementary components: spatial consistency, morphological consistency, and energy consistency, which are then combined in a weighted manner to form the total consistency score. Spatial consistency is used to describe whether anomalous response centers in different spectral bands cluster near the same location. Considering that low-resolution targets should exhibit approximately co-located responses across spectral bands, spatial dispersion can be defined as follows:

[0114]

[0115] Map it to a spatial consistency score of [0,1] ,in A scale parameter consistent with the magnitude of cross-spectral registration error. Morphological consistency is used to suppress erroneous coordination caused by "spurious anomalies at the edges of single spectral bands," that is, when some spectral bands exhibit linear or edge structures, coordination should be reduced. This can be achieved by describing the morphology of each spectral band. The discreteness is mapped to the consistency score. For example, variance can be calculated and exponentially mapped for compactness or anisotropy indices. Energy consistency is used to characterize the reliable co-occurrence of anomalies relative to the background; that is, real targets typically exhibit non-random contrast enhancements across multiple spectral bands, while noise or saturation often leads to imbalances in cross-spectral energy relationships. This can be addressed by... Construction of sorting consistency or quantile stability Finally, the three types of consistency are combined into a total consistency score:

[0116]

[0117] Weighting coefficient of total consistency score , , satisfy The consistency score does not serve as the final determination of the target's existence, but is used for subsequent gating decisions: when C(Ωc) is high, it indicates that the multi-spectral evidence supports each other in terms of spatial location, morphological structure, and contrast relationship, making it suitable for strong co-localization; when... A low score indicates significant conflicts or unreliable factors between different spectral bands. Strong fusion may introduce bias, and the system should degenerate into a more robust single-band dominant localization. (Based on consistency score) Define the gate value To control the intensity of cooperative localization, and to avoid policy jitter caused by hard thresholding, this method prefers to use soft-gated mapping.

[0118]

[0119] in For the Sigmoid function, The threshold value is used for gate control. is the gating steepness coefficient. When When the value approaches 1, strong cooperative positioning is executed; when... When the response approaches 0, degenerate localization is performed. One possible form of strong cooperative localization is to calculate the center position by weighting the multi-band responses within a candidate window based on their reliability; for example, by setting each band at... The normalized response within is Then, the joint response is defined as follows:

[0120]

[0121] in Desirable and The relevant adaptive weights or preset weights are used, and the weighted centroid is applied to obtain the cooperative localization result:

[0122]

[0123] β controls the degree of concentration of "soft maximization". Degeneracy localization selects a single most reliable spectral band. (e.g., having the largest) Or the spectral band with the smallest morphological anomaly) is used as the dominant factor, and the response estimation center is used only for that spectral band. The final positioning output can be written in a gated weighted adaptive form:

[0124] ,

[0125] And output simultaneously This serves as a reliable measure of the localization result, providing a "localizable and interpretable" output for subsequent systems under complex interference conditions. In supervised training scenarios, region-level or point-level annotations can be used for joint learning of gating and localization. Region-level annotations can be represented as:

[0126]

[0127] in This indicates the region where the target exists. Positive candidate windows (covering...) can be constructed during training. (region) and negative candidate window (sampled from) (the area), and the gating value A binary supervision mechanism is applied, favoring higher gating for positive windows and lower gating for negative windows; simultaneously, a central error constraint is applied to the positioning result ˆx within the positive window. Because... , and soft-core form All of them are continuously differentiable. The above joint objectives can be optimized through standard backpropagation, enabling the gating mechanism to obtain a more stable cross-spectral collaborative decision-making capability under the learning meaning.

[0128] This method uses Landsat 8 multispectral remote sensing imagery as the dataset source. Each sample consists of multispectral images of the same geographic location, and the selected spectral bands can include visible light, near-infrared, shortwave infrared, and thermal infrared bands, all uniformly resampled to the same spatial grid. The input to the algorithm is not the entire image, but several local windows obtained during the candidate generation stage; therefore, both training and testing use "candidate windows" as the basic processing unit.

[0129] Training data and annotation method: Training samples are labeled using center points instead of detection boxes or segmentation masks. For samples containing targets, a reference target center position is provided within the candidate window. (Pixel coordinates); For samples that do not contain the target, the candidate window is labeled as a negative sample, and no center position is provided. This labeling method conforms to the imaging characteristics of low-resolution targets that are "smaller than a pixel or approximately a point target," and also avoids artificially constructing inaccurate target boundaries under low-resolution conditions. The training set contains both positive sample windows (covering the target) and negative sample windows (background or interference) to simultaneously learn localization and consistency gating.

[0130] During the training phase, the multispectral imagery of each candidate window first enters the evidence construction module, where local evidence related to localization (abnormal intensity, relative background contrast, anomalous morphology, and peak position) is extracted from each spectral band. Subsequently, this cross-spectral evidence is fed into the consistency evaluation module, which calculates the cross-spectral consistency score of the candidate window and generates a gating value. Based on the gating result, the algorithm adaptively selects either "multi-spectral co-localization" or "single-spectral dominant localization" during forward propagation and outputs the predicted target center location. .

[0131] The training loss consists of two parts: one is the localization loss, which is used to constrain the prediction center in the positive sample window. With Reference Center The training process employs two loss functions: first, the distance between samples; and second, a gated supervision loss, used to constrain the gate value to be too high in positive samples and too low in negative samples. All learnable modules (such as parameters in evidence mapping, consistency assessment, and gate mapping) are updated end-to-end through standard backpropagation, driven by this loss function. Because the localization output is continuous, the entire training process remains differentiable.

[0132] In the testing phase, the algorithm no longer uses any annotation information. Given a Landsat 8 multispectral image, the algorithm first uses a candidate generation module to select several candidate windows that may contain low-resolution targets. Then, for each candidate window, the same forward process as in the training phase is performed to calculate cross-spectral evidence, consistency gating, and output the target localization result. The final output of the testing phase includes: the target center location (pixel coordinates or latitude and longitude coordinates) and the corresponding consistency gating value. The latter serves as the localization confidence level, used to filter high-confidence results or indicate situations requiring manual verification.

Claims

1. A method for locating low-resolution targets in space-based multispectral remote sensing imagery, characterized in that, Includes the following steps: S1. Perform geometric alignment and unified spatial grid resampling on the input space-based multi-spectral remote sensing image to obtain multi-spectral images with pixel-level correspondence between spectral bands. The space-based multi-spectral remote sensing image includes visible light, near-infrared, short-wave infrared and / or thermal infrared bands. S2. Perform rapid anomaly response detection on the processed multi-spectral images, using a single-spectral anomaly detection strategy, a multi-spectral cross-spectral co-occurrence anomaly detection strategy, or a combination of both, combined with connected component analysis or a fixed-size sliding window to aggregate anomaly pixels, to generate a candidate region set based on the principle of high recall. S3. For each candidate region, construct local background statistics on each spectral band and generate a normalized anomaly response map. Based on the normalized anomaly response map, extract intensity evidence, contrast evidence, morphological evidence and spatial evidence to form a cross-spectral comparable localization evidence vector for each spectral band. S4. Based on the localization evidence vectors of each spectral band, calculate the spatial consistency score, morphological consistency score and energy consistency score respectively, obtain the total consistency score of the candidate region by weighted summation, and then convert the total consistency score into a gate value in the [0,1] interval through Sigmoid soft gating mapping; S5. Perform adaptive positioning based on the threshold value: when the threshold value approaches 1, the target center position is estimated using a multi-spectral band cooperative positioning method with spectral band reliability weighting; when the threshold value approaches 0, the most reliable spectral band is selected to perform a single-spectral band dominant positioning method to estimate the target center position; the two positioning results are weighted and fused by the threshold value to obtain the final target center position of the candidate region. S6. Output the target center position of each candidate region, and simultaneously output the gating value as a positioning reliability index.

2. The method for locating low-resolution targets in space-based multispectral remote sensing imagery according to claim 1, characterized in that, The single-band anomaly detection strategy in S2 is as follows: in the spectral band sensitive to the target... An anomaly indicator function is constructed above, which is used when the difference between a pixel response and the mean value of the local window background centered on that pixel exceeds a preset threshold. When a pixel is identified as an anomalous pixel, the anomalous indicator function outputs 1; otherwise, it outputs 0. The multi-spectral cross-spectral co-occurrence anomaly detection strategy is as follows: binary anomaly indicator maps are generated for multiple spectral bands, and a joint anomaly response map is obtained by weighted summation using logical OR. When the joint response value is greater than or equal to 1, the pixel is identified as an anomalous pixel. The method for aggregating anomalous pixels is to perform connected component analysis on the anomalous pixels, or to use a fixed-size sliding window to select scattered anomalous pixels, forming a candidate region as a local window. .

3. The method for locating low-resolution targets in space-based multispectral remote sensing imagery according to claim 2, characterized in that, The local background statistics in S3 are selected as mean or median, standard deviation or median absolute deviation (MAD), and the calculation formula for the normalized anomaly response map is: in To avoid small constants that cause numerical instability; the image of the k-th spectral band of the n-th sample is... For each candidate window First, a local background reference is constructed on spectral band k to measure the "degree of anomaly"; and These represent the background statistics within the window. The specific extraction methods for the four types of evidence are as follows: intensity evidence is the maximum value of the normalized anomalous response within the window; contrast evidence is the difference between the intensity evidence and the median of the normalized anomalous response within the window; morphological evidence is the principal axis ratio calculated based on the area, dispersion, or principal directionality of the support domain after thresholding the normalized anomalous response; and spatial evidence is the peak position of the normalized anomalous response within the window or the weighted centroid on the anomalous support domain.

4. The method for locating low-resolution targets in space-based multispectral remote sensing imagery according to claim 3, characterized in that, The spatial consistency score mentioned in S4 is calculated as follows: First, the spatial dispersion of the spatial evidence for each spectral band is calculated. in The mean of the spatial evidence for each spectral band is then mapped through an exponential method. Map the dispersion to a spatial consistency score in the range [0,1]. Scale parameters consistent with the order of magnitude of cross-spectral registration error; morphological consistency score The variance of the morphological evidence for each spectral band is calculated and then an exponential mapping is performed. Energy consistency scores are constructed through analysis of the ordering consistency or quantile stability of contrast evidence across spectral bands. The above three types of consistency are combined into a total consistency score. Weighting coefficient of total consistency score , , satisfy .

5. The method for locating low-resolution targets in space-based multispectral remote sensing imagery according to claim 4, characterized in that, The calculation formula for the Sigmoid soft-gated mapping in S4 is as follows: in For the Sigmoid function, The threshold value is used for gate control. This is the gate steepness coefficient, the gate value. The closer the value is to 1, the higher the consistency of multi-spectral evidence; the closer it is to 0, the lower the consistency. The selection criterion for the most reliable spectral segment in S5 is to select the spectral segment with the greatest contrast evidence or the smallest degree of morphological abnormality.

6. The method for locating low-resolution targets in space-based multispectral remote sensing imagery according to claim 5, characterized in that, The specific implementation of the multi-spectral band collaborative localization in S5 is as follows: setting adaptive weights based on the contrast evidence of each spectral band. Calculate the joint response within the candidate window: The collaborative localization results are then calculated using soft-maximization weighted centroid calculation: in To control the coefficient of soft maximization concentration, the formula for calculating the final target center position is: The results are based on the dominant localization of a single spectral band.

7. The method for locating low-resolution targets in space-based multispectral remote sensing imagery according to claim 6, characterized in that, The method also includes a model training step, specifically: using Landsat 8 multi-band imagery as the data source, constructing training samples with candidate windows as the basic unit; labeling the center pixel coordinates of the target in positive sample windows containing low-resolution targets, and not labeling the center position of negative sample windows composed of background or interference; constructing a joint loss function composed of localization loss and gated supervision loss, whereby the localization loss constrains the distance between the predicted center and the labeled center of the positive sample window, and the gated supervision loss constrains the gate value of the positive sample window to be too high and the gate value of the negative sample window to be too low; and performing end-to-end updates on the learnable modules in evidence mapping, consistency evaluation, and gate mapping through standard backpropagation, with the localization output being in a continuous form to ensure that the training process is differentiable.

8. The method for locating low-resolution targets in space-based multispectral remote sensing imagery according to claim 7, characterized in that, The target center position in S6 includes pixel coordinates or latitude and longitude coordinates. The gate value serves as a positioning reliability index, used by the downstream system to screen high-reliability positioning results or to manually verify low-reliability positioning results. The entire localization process only performs fine analysis of cross-spectral evidence construction, consistency measurement, and adaptive localization within the candidate region, while only lightweight anomaly detection and candidate region generation are performed across the entire map, thus achieving controllable computational complexity.

9. A low-resolution target localization system for space-based multispectral remote sensing imagery, employing the method described in claim 8, characterized in that... It includes an input preprocessing module, a candidate region generation module, a cross-spectral localization evidence construction module, a cross-spectral consistency measurement and gating decision module, an adaptive cooperative localization module, and an output module, which are connected in sequence. The input preprocessing module is used to perform geometric alignment and unified spatial grid resampling on the input space-based multi-spectral remote sensing images to obtain multi-spectral images with pixel-level correspondence between spectral bands; The candidate region generation module is used to perform abnormal response detection on the preprocessed multi-spectral image, aggregate abnormal pixels through a preset strategy to generate a high-recall candidate region set, and transmit the candidate region data to the cross-spectral localization evidence construction module. The cross-spectral localization evidence construction module is used to generate a normalized anomaly response map in each spectral band for each candidate region, extract four types of localization evidence (intensity, contrast, morphology, and space) and form an evidence vector, and transmit the evidence vector data to the cross-spectral consistency measurement and gating decision module. The cross-spectral consistency measurement and gating decision module is used to calculate the consistency scores of three categories—space, morphology, and energy—based on the localization evidence vector and weight them to obtain the total consistency score. It generates a gating value through a Sigmoid soft gating mapping and transmits the total consistency score and the gating value to the adaptive cooperative localization module. The adaptive cooperative positioning module is used to adaptively select multi-band cooperative positioning or single-band dominant positioning mode according to the gate value, obtain the final target center position through weighted fusion of the gate value, and transmit the target center position data and the gate value to the output module. The output module is used to output the target center position and simultaneously output the gating value as a positioning reliability index. The target center position includes pixel coordinates or latitude and longitude coordinates.

10. The low-resolution target localization system for space-based multispectral remote sensing imagery according to claim 9, characterized in that, The candidate region generation module incorporates a single-spectral-band anomaly detection unit, a multi-spectral-band cross-spectral co-occurrence anomaly detection unit, and an anomaly pixel aggregation unit to implement the aforementioned anomaly detection strategy and pixel aggregation method, respectively. The cross-spectral localization evidence construction module incorporates a background statistics calculation unit, a normalized response generation unit, and a four-type evidence extraction unit to implement the aforementioned local background statistics calculation, normalized anomaly response map generation, and localization evidence extraction, respectively. The cross-spectral consistency measurement and gating decision module incorporates a spatial consistency calculation unit, a morphological consistency calculation unit, an energy consistency calculation unit, and a soft-gating mapping unit to implement the aforementioned consistency score calculation and gating value generation, respectively. The adaptive collaborative localization module incorporates a collaborative localization unit and a single-spectral-band localization unit to implement the aforementioned multi-spectral-band collaborative localization and single-spectral-band dominant localization, respectively, and the system supports the aforementioned joint loss function and end-to-end training of backpropagation. The output module incorporates a coordinate transformation unit and a confidence labeling unit to implement the conversion between pixel coordinates and latitude and longitude coordinates, and to associate the gating value with the target center position for labeling and output.

Images