A method and system for extracting weak thermal anomaly candidates under clouds in space-based thermal infrared remote sensing and suppressing false anomalies

Abstract

The application provides a cloud-based thermal infrared remote sensing method and system for extracting weak thermal anomaly candidates and suppressing false anomalies, belonging to the field of remote sensing image processing and computer vision. To solve the problem of how to ensure high recall rate of weak thermal anomalies and effectively suppress false anomaly candidates under low spatial resolution and complex observation conditions, and output candidate results with credibility description, the application constructs a candidate learning mechanism for cloud influence perception; introduces a gated false anomaly suppression mechanism to establish a continuous reliability modulation function based on candidate saliency; constructs a candidate probability map and a candidate region confidence index, so that the output result not only includes a candidate region set, but also contains corresponding credibility information. The application realizes the automatic extraction of weak thermal anomaly candidate regions and the suppression of false anomalies under the condition of cloud influence or complex ground surface background interference, providing reliable candidate input for subsequent fine positioning, cross-spectral consistency analysis or manual review.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and computer vision technology, and more specifically, to a method and system for extracting candidate weak thermal anomalies under clouds and suppressing false anomalies in space-based thermal infrared remote sensing. Background Technology

[0002] Space-based thermal infrared remote sensing is widely used in areas such as surface temperature monitoring, fire detection, industrial heat source identification, and environmental anomaly analysis. Multispectral satellites, such as Landsat 8, carry thermal infrared sensors (e.g., TIRS Band 10 and Band 11), capable of providing surface brightness temperature observation data at a spatial resolution of 100m. Due to its long-term stable operation and high radiometric accuracy, Landsat 8 data has become an important data source for medium- and low-resolution thermal anomaly research.

[0003] In existing technologies, most typical methods for thermal anomaly detection employ a two-stage structure of "candidate extraction—false alarm suppression." Early algorithms typically relied on fixed thresholds or relative background anomaly determination, comparing the brightness temperature of a pixel with the statistics of its neighborhood; if it exceeded a preset threshold, it was considered an anomaly candidate. Common enhanced context-sensitive fire detection algorithms achieve anomaly detection by comparing the brightness temperature difference between candidate pixels and the surrounding background. Similar ideas have also been applied to fire detection algorithms in medium-resolution products such as MODIS and VIIRS. These methods are simple in structure and computationally efficient, but they rely on the stability of background statistics.

[0004] For thermal anomaly studies using Landsat 8 data, most methods employ context thresholding strategies or construct anomaly response maps based on single-band brightness-temperature differences. However, under cloud conditions, especially in thin clouds or cloud boundary regions, the statistical characteristics of background brightness-temperature change, leading to inaccurate anomaly detection thresholds. Furthermore, factors such as cloud top radiation, cloud-ground mixed pixel effects, and sensor noise can create false anomaly responses in the thermal infrared band, significantly increasing the false alarm rate. To avoid false alarms, existing techniques typically use cloud masks to directly remove cloud pixels or suspected cloud areas; however, this "one-size-fits-all" strategy also removes genuine weak thermal anomalies beneath the clouds, resulting in systematic missed detections.

[0005] On the other hand, the sources of false anomalies are diverse, including cloud boundary mixed pixels, highly reflective surface structures, persistent industrial heat sources, striped noise, and locally saturated pixels. Existing methods mostly use empirical rule superposition (such as area constraints, neighborhood morphology constraints, or auxiliary band thresholds) for false positive rejection, but lack a unified modeling framework for different types of false anomalies. This rule-based suppression method lacks stability under complex background conditions and struggles to effectively control false positives while maintaining the recall rate of weak anomalies.

[0006] In summary, existing technologies for detecting weak thermal anomalies in space-based thermal infrared radiation have the following main shortcomings:

[0007] (1) The candidate extraction stage relies on fixed thresholds or simple context comparisons, which are difficult to adapt to the brightness temperature changes under the influence of clouds, resulting in the failure to detect weak thermal anomalies under clouds; (2) False anomaly suppression relies on empirical rules or hard mask strategies, which have a high false alarm rate or serious failure to detect in complex backgrounds or cloud boundary areas; (3) Existing methods lack a continuous characterization mechanism for candidate reliability, and cannot provide interpretable confidence information for subsequent analysis.

[0008] Therefore, it is necessary to propose a new technical solution that, without relying on strict cloud removal, improves the robustness of weak thermal anomaly candidate extraction under cloud influence conditions through learnable or structured candidate modeling and gated pseudo-anomaly suppression mechanism, and outputs candidate results with credible characterization, so as to overcome the above-mentioned defects of existing technologies. Summary of the Invention

[0009] The technical problem to be solved by this invention is:

[0010] To address the challenges of ensuring high recall of weak thermal anomalies, effectively suppressing spurious anomaly candidates, and outputting candidate results with reliable characterization under conditions of low spatial resolution and complex observations.

[0011] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0012] This invention provides a method for extracting candidate weak thermal anomalies under clouds and suppressing false anomalies in space-based thermal infrared remote sensing, comprising the following steps:

[0013] S100. During the input phase, the thermal infrared brightness temperature map of Landsat 8 is received as the data source, and quality layer information can be used as an auxiliary input. To ensure spatial consistency, all input data have undergone basic geometric alignment and resampling processing before entering the network.

[0014] S200. During the candidate learning stage, the network extracts features from the input thermal infrared brightness temperature image through a convolutional structure and outputs a candidate saliency map. The candidate saliency map is used to represent the probability that each pixel belongs to a weak thermal anomaly candidate. During training, a weakly labeled mask is used for supervision.

[0015] S300. In the gating suppression stage, the network outputs a gating map in parallel to characterize the reliability of the candidate response at each pixel, that is, to reflect whether the pixel is more likely to belong to the pseudo-anomaly region; by continuously multiplying and combining the candidate saliency map with the gating map, the final candidate probability map is obtained.

[0016] S400. During the training phase, positive and negative sample patches constructed using Landsat 8 are used for supervised learning. Positive samples are windows containing known thermal anomaly centers, while negative samples are background windows without thermal anomalies or regions prone to generating false anomalies. Network parameters are updated through standard backpropagation. Gated branches indirectly learn the ability to suppress false anomalies through joint loss and negative sample sampling strategy.

[0017] S500: During the testing phase, after inputting a complete image, the network outputs a candidate probability map; candidate pixels are extracted by setting a lenient threshold, and a set of candidate regions is generated through connected component analysis; the maximum or average probability of each candidate region is used as its confidence index for sorting or filtering; the final output includes the set of candidate regions and their corresponding confidence scores.

[0018] Further, in step S100, let the thermal infrared brightness temperature map of the nth sample be... ,in Represents pixel coordinates, This represents the image domain, i.e., the pixel coordinate x belongs to the image region Ω; when a Landsat 8 quality layer is available, it is denoted as... This is used as an auxiliary input feature and stitched into the thermal infrared brightness temperature map to form a multi-channel input. In cases where a reliable quality layer cannot be obtained, only the brightness temperature map is used as input.

[0019] Furthermore, in step S100, to reduce the difference in brightness temperature distribution between different scenes, the following steps are taken: Standardize the process.

[0020] Further, in step S200, the following are included:

[0021] Using a lightweight convolutional neural network , will input Mapped to candidate saliency map ,Right now:

[0022] (1)

[0023] in, Used to represent the probability that pixel x belongs to the weak thermal anomaly candidate;

[0024] Weakly supervised labeling masks are used during the training phase. As a supervisory signal, let the candidate labels for the corresponding samples be:

[0025] (2)

[0026] in, This indicates that the pixel is located within a candidate region for thermal anomalies, either manually or using external data. This indicates that the pixel is located in the background area;

[0027] Using pixel-level binary cross-entropy loss function As a basic form of supervision:

[0028] (3)

[0029] The pixel-level binary cross-entropy loss function encourages the network to output a higher response on pixels labeled as candidate regions and a lower response on background pixels.

[0030] After training, during the testing phase, unlabeled Landsat 8 thermal infrared images are input into the network to obtain candidate saliency maps. Subsequently, a lenient threshold was set. Construct the initial candidate set :

[0031] (4)

[0032] Pixel-level candidates are transformed into a set of candidate regions through connected component analysis or fixed-size window aggregation. .

[0033] Further, in step S300, the following are included:

[0034] The network structure employs a shared encoder and dual-branch output: candidate branch outputs are defined. Gated branch output ;in This value is used to represent the unreliability or tendency for false anomalies in candidate responses at that pixel; a larger value indicates that it should be suppressed more. The final candidate probability map is constructed based on these two factors. It is implemented using a multiplicative gating method:

[0035] (5).

[0036] Further, in step S400, the following is included:

[0037] Suppose the training samples are from the thermal infrared brightness temperature map of Landsat 8, denoted as . Optional input of quality layer or QA information. Network input is The output is a candidate saliency map. With gated graph Candidate probability maps are obtained using multiplicative gating. Training supervision uses weakly labeled candidate masks.

[0038] The optimization goal is and Consistency is achieved by training candidate outputs using pixel-level binary cross-entropy. for:

[0039] (6)

[0040] Adding lightweight constraints to the training process Apply weak regularization to the overall distribution, or introduce window-level gating supervision to the negative sample patch;

[0041] Final training loss for:

[0042] (7) Among them, This represents a weakly regularized gated or window-level supervised term, where λ is the weighting coefficient.

[0043] Further, in step S500, the following are included:

[0044] During the inference phase, given an unlabeled Landsat 8 thermal infrared brightness temperature map... Forward calculation yields and The candidate cell set is determined by... Obtained by applying a relaxed threshold:

[0045] (8)

[0046] Among them, threshold The goal was to ensure candidate recall; subsequently, [the following was done / implemented]. Perform connected component analysis to aggregate candidate cells into a candidate region set. It outputs candidate confidence scores for each candidate region for sorting and hierarchical output.

[0047] A system for extracting candidate weak thermal anomalies under clouds and suppressing false anomalies in space-based thermal infrared remote sensing is provided. The system has program modules corresponding to the above steps and executes the steps in the above-mentioned method for extracting candidate weak thermal anomalies under clouds and suppressing false anomalies when running.

[0048] A computer-readable storage medium storing a computer program configured to, when invoked by a processor, implement the steps of a method for extracting candidate weak thermal anomalies under clouds and suppressing false anomalies in space-based thermal infrared remote sensing.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] (1) Significantly reduces the false negative rate of weak thermal anomalies under clouds

[0051] Existing methods typically remove cloud pixels directly during the detection process or use fixed thresholds, leading to systematic missed detection of weak thermal anomalies under clouds. This invention utilizes a candidate learning mechanism for cloud impact perception, enabling the triggering of candidate responses even under conditions of reduced brightness-temperature contrast. In simulation experiments of typical cloud boundaries and thin cloud cover scenarios, compared to traditional fixed threshold methods, this invention improves candidate recall by approximately 10%–20% in weak thermal anomaly scenarios while maintaining the same false alarm rate.

[0052] (2) Effectively suppresses pseudo-anomalies caused by cloud boundaries and background structures.

[0053] Traditional methods often misclassify cloud boundary mixed pixels or highly reflective structures as thermal anomalies. This invention uses a gated modulation mechanism to continuously suppress anomalous responses, significantly reducing false anomaly candidates. In datasets containing complex backgrounds and cloud interference, the false alarm rate of this invention can be reduced by approximately 15%–30% compared to direct candidate extraction methods, while maintaining a stable candidate recall rate.

[0054] (3) Improve the stability and interpretability of candidate results

[0055] This invention outputs a candidate probability map and a regional confidence index, providing a reliability measure for each candidate result. Compared to existing technologies that only output binary results, this output format provides richer information support for subsequent multi-source fusion, time-series analysis, or manual verification, significantly enhancing the interpretability and scalability of the system.

[0056] (4) The computational efficiency is comparable to existing methods and is suitable for engineering deployment.

[0057] The candidate learning and gating suppression structure of this invention can be implemented based on lightweight convolutional networks or local computation modules. The computational complexity is mainly concentrated within the candidate region, and the overall running time is on the same order of magnitude as traditional thermal anomaly detection algorithms. It can meet the practical remote sensing data processing needs on conventional computing platforms.

[0058] In summary, this invention effectively solves the problems of missed detection of weak thermal anomaly candidates and false alarms of anomalies under cloud influence conditions in existing technologies while keeping computational complexity under control. It improves the stability and reliability of candidate extraction and provides intermediate results with continuous confidence expression for subsequent research and system expansion, demonstrating significant technological progress and application value. Attached Figure Description

[0059] Figure 1 This is a flowchart of a method for extracting candidate weak thermal anomalies under clouds and suppressing false anomalies in space-based thermal infrared remote sensing, according to an embodiment of the present invention. Detailed Implementation

[0060] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0061] This method addresses the problem of low-resolution thermal anomaly candidate extraction in Landsat 8 thermal infrared imagery. The goal is to generate high-recall thermal anomaly candidate regions with controllable false positives, even under conditions of cloud cover, complex backgrounds, and observational noise. The method consists of three consecutive stages: candidate learning, gating suppression, and candidate output. The entire process maintains a consistent forward structure during training and testing, with the only difference being whether or not annotations are used for parameter updates.

[0062] Specific Implementation Plan 1: Combining Figure 1 As shown, this invention provides a method for extracting candidate weak thermal anomalies under clouds and suppressing false anomalies in space-based thermal infrared remote sensing, comprising the following steps:

[0063] S100. In the input stage, the Landsat 8 thermal infrared brightness temperature map is received as the data source, and quality layer information can be used as an auxiliary input. To ensure spatial consistency, all input data have undergone basic geometric alignment and resampling processing before entering the network. Since low-resolution targets only occupy a small number of pixels, this method does not perform high-complexity analysis on the entire image, but instead constructs local samples on the image using a sliding window or patch segmentation method, thereby reducing computational complexity and enhancing sensitivity to local anomalies.

[0064] In Landsat 8 thermal infrared imagery, low-resolution thermal anomalies typically manifest as a slight increase in local brightness temperature, with a spatial scale close to or smaller than the pixel resolution. In the presence of thin clouds or cloud boundary interference, the thermal anomaly signal may be partially attenuated, while the statistical characteristics of the background brightness temperature change. Therefore, the goal of the candidate extraction stage is not to directly complete the final determination, but to learn a "weak thermal anomaly candidate response" from the thermal infrared imagery that is robust to cloud impacts while ensuring a high recall rate.

[0065] Let the thermal infrared brightness temperature diagram of the nth sample be... ,in Represents pixel coordinates, Let Ω represent the image domain, meaning that pixel coordinate x belongs to the image region Ω, and Ω is a two-dimensional integer grid. A subset of the square; to reduce the impact of differences in brightness temperature distribution between different scenes, firstly... Standardize the input data, for example, by subtracting the global or local mean and dividing by a scale factor, to ensure the data distribution remains within a stable range; if a Landsat 8 quality layer (e.g., cloud, shadow, or saturation marker) is available, denote it as... Then, it is used as an auxiliary input feature and spliced ​​into the thermal infrared brightness temperature map to form a multi-channel input. In cases where a reliable quality layer cannot be obtained, brightness temperature maps can be used as input alone.

[0066] S200. In the candidate learning stage, the network extracts features from the input thermal infrared brightness temperature map through a convolutional structure and outputs a candidate saliency map. This candidate saliency map represents the probability that each pixel belongs to a "weak thermal anomaly candidate". During training, a weakly labeled mask is used for supervision. The label only indicates the approximate range of the candidate region and does not require precise boundaries. The design goal of this stage is to improve the candidate recall rate, that is, to avoid missing real weak thermal anomalies as much as possible, while allowing a certain number of false positives.

[0067] Specifically, including,

[0068] To enable candidate learning for weak thermal anomalies, this method employs a lightweight convolutional neural network. , will input Mapped to candidate saliency map ,Right now:

[0069] (1)

[0070] in, Used to represent the probability that pixel x belongs to the "weak thermal anomaly candidate";

[0071] The network structure can adopt an encoder-decoder convolutional structure or a shallow convolutional network, with the focus on maintaining local responsiveness and computational efficiency, rather than relying on large-scale semantic features;

[0072] Weakly supervised labeling masks are used during the training phase. As a supervisory signal, let the candidate labels for the corresponding samples be:

[0073] (2)

[0074] in, This indicates that the pixel is located within a candidate region for thermal anomalies, either manually or using external data. This indicates that the pixel is located in the background area;

[0075] It is important to emphasize that this mask is not required to precisely characterize the boundaries of real anomalies; its role is to guide the network to learn "anomaly tendencies" rather than to perform precise segmentation. Therefore, a pixel-level binary cross-entropy loss function is employed. As a basic form of supervision:

[0076] (3)

[0077] This loss function encourages the network to output higher responses on pixels labeled as candidate regions and lower responses on background pixels;

[0078] Considering that cloud effects may weaken the brightness-temperature difference of real weak thermal anomalies, this step emphasizes "high recall priority". Therefore, when constructing training data, the labeled area can be moderately expanded to avoid overfitting caused by uncertain precise boundaries. At the same time, the sample composition should include a certain proportion of "weak thermal under clouds" samples so that the network can learn feature patterns that can still identify anomalies under the condition of weakened brightness-temperature contrast.

[0079] After training, during the testing phase, unlabeled Landsat 8 thermal infrared images were directly input into the network to obtain candidate saliency maps. Subsequently, a lenient threshold was set. Construct the initial candidate set :

[0080] (4)

[0081] Then, pixel-level candidates are transformed into a set of candidate regions through connected component analysis or fixed-size window aggregation. This candidate set only represents regions where "weak thermal anomalies may exist," providing input for subsequent gated pseudo-anomaly suppression and fine analysis.

[0082] The core contribution of this stage is that, without relying on hard threshold rules, we obtain weak thermal anomaly candidate responses with a certain adaptability to cloud impact through learnable candidate saliency mapping, providing a reliable initial candidate basis for subsequent consistency-gated suppression mechanisms.

[0083] S300. In the gating suppression stage, the network outputs a gating map in parallel to characterize the reliability of candidate responses at each pixel. The gating map reflects whether the pixel is more likely to belong to a pseudo-anomaly region, such as a location affected by cloud boundaries, background structures, or noise. By continuously multiplying the candidate saliency map with the gating map, the final candidate probability map is obtained. This mechanism enables the algorithm to automatically suppress potential pseudo-anomalies while preserving true weak thermal anomaly responses, without relying on fixed rules or hard thresholds for elimination.

[0084] Candidate saliency map obtained from step S200 While ensuring weak thermal anomaly recall, it is inevitable to introduce pseudo-anomaly responses caused by cloud boundary mixed pixels, background structure edges, stripe noise, and local anomaly calibration factors; if directly addressing... Hard thresholding and outputting candidate regions significantly increases false alarms, and the spatial morphology of false alarms often presents as linear, band-like, or large-area low-frequency rises, which differs significantly from the statistical morphology of low-resolution point / small patch thermal anomalies. Therefore, this invention introduces a gated false anomaly suppression mechanism: while retaining the "loose triggering" condition, a gated map related to observation reliability is learned. The candidate saliency is suppressed pixel by pixel, thereby improving the "candidate extraction" from a single threshold decision to an "interpretable joint decision of suppression and retention".

[0085] Specifically, including,

[0086] The network structure employs a shared encoder and a dual-branch output: candidate branch output. Gated branch output ;in This value is used to represent the unreliability or tendency for false anomalies in candidate responses at that pixel; a larger value indicates that it should be suppressed more. The final candidate probability map is constructed based on these two factors. It is implemented using a multiplicative gating method:

[0087] (5)

[0088] This form has two key properties: firstly, right and Both are continuously differentiable, facilitating end-to-end training; secondly, at a certain position Higher but At the same time, when the value is high (typical pseudo-anomaly). It will be explicitly suppressed, thus avoiding the triggering of candidates based solely on a single thermal anomaly response; when higher and At lower levels (more likely a true anomaly). By maintaining a high level, candidates are retained. Unlike traditional "direct elimination by cloud mask" or "rule stacking rejection", gating suppression is continuous and learnable, and allows the retention of low-confidence candidates when the cloud influence is strong but there is still weak thermal anomaly evidence, providing an entry point for subsequent multi-source verification or manual review.

[0089] In terms of training supervision, weakly labeled candidate masks are still used. As the primary monitoring signal, it is used to constrain the final output. Consistency with candidate regions; candidate weak annotations are defined as:

[0090] (6)

[0091] in, This indicates the candidate region for thermal anomalies (obtained by point label expansion or coarse labeling). Indicates the background area; based on this annotation, for Supervision is performed using pixel-level binary cross-entropy or Dice loss to ensure the model outputs higher values ​​within candidate regions. Low output in background area ;

[0092] It should be emphasized that this oversight does not require gated branches. It has independent truth labels; The learning comes from the backpropagation of the error gradient in the multiplicative relationship: when a certain background region is... When a fault is activated, optimization will tend to improve... To suppress the spurious response; when a real candidate region has a weak response due to cloud influence, the optimization will tend to reduce... To avoid over-suppression, this mechanism enables the gating branch to learn the behavior pattern of "suppressing false anomalies and allowing weak true anomalies" without explicit gating labels, thereby meeting the dual requirements of "high recall and controllable false alarms" for weak hot candidate extraction in the cloud.

[0093] During the inference phase, the final candidate output is determined by... Generate; firstly for Use a lenient threshold Extract candidate cell set Subsequently, connected component analysis was used to aggregate candidate pixels into a candidate region set. For each candidate region Candidate confidence levels can be further defined (e.g.) (or regional mean) is used for candidate ranking and hierarchical output: high-confidence candidates can directly enter the subsequent localization / verification module, while low-confidence candidates can enter the review or further research and analysis stage; through gating suppression, this method retains the ability to sensitively trigger weak thermal anomalies while significantly reducing the false candidate output caused by cloud boundaries and complex backgrounds, making the candidate set more suitable as the upstream input for subsequent cross-spectral consistency verification, spatiotemporal consistency verification or physical modeling research;

[0094] S400. During the training phase, positive and negative sample patches constructed using Landsat 8 are used for supervised learning. Positive samples are windows containing known thermal anomaly centers, while negative samples are background windows without anomalies or regions prone to generating false anomalies. The model uses the difference between the candidate probability map and the weakly labeled mask as the main optimization objective and updates the network parameters through standard backpropagation. The gating branch indirectly learns the ability to suppress false anomalies through joint loss and negative sample sampling strategy.

[0095] Specifically, including,

[0096] Let the training samples come from the thermal infrared brightness temperature map of Landsat 8 (preferably TIRS Band 10), denoted as Optional input of quality layer or QA information. (This is only for supplementary observation reliability information); network input is... (Used only when there is no QA) The output is a candidate saliency map. With gated graph Candidate probability maps are obtained using multiplicative gating. Training supervision uses weakly labeled candidate masks, which take the following form:

[0097] (7)

[0098] in, Indicates candidate regions for thermal anomalies. This indicates the background region. It's important to emphasize that this annotation is used to characterize the "candidate region" rather than the precise boundary; therefore, it can be obtained through point annotation dilation, coarse polygons, or mapping based on external event coordinates followed by manual verification. To ensure training stability and generalization ability, the training samples include both positive sample regions (…). The window or patch, also including the negative sample region ( The background patch suggests intentionally sampling "spurious anomaly regions" (such as near cloud boundaries, strong background edges, noise stripes, or high brightness temperatures but non-target regions) in negative samples to force the gating branch to learn suppression strategies.

[0099] Optimization target and The primary goal is consistency, and candidate outputs are trained using pixel-level binary cross-entropy or Dice loss, for example, using binary cross-entropy. :

[0100] (8)

[0101] because right and All are continuously differentiable functions. The gradient can be used to update the parameters θ of both the candidate branch and the gated branch simultaneously through standard backpropagation; to avoid the degradation of the gated branch (e.g. Prematurely approaching all 1s or all 0s and enhancing its "pseudo-anomaly suppression" semantics can be achieved by adding lightweight constraints during training: a reproducible approach that does not rely on additional annotations is... The overall distribution is subjected to weak regularization (e.g., constraining its mean to be within a reasonable range to avoid full suppression); another more stable approach is to introduce window-level gating supervision into the negative sample patch, causing negative samples to tend to have higher average gating values ​​and positive samples to tend to have lower average gating values, thereby... It more explicitly assumes the function of "suppressing false anomalies"; the final training loss It can be written as:

[0102] (9) Among them, This indicates a weakly regularized gated or window-level supervised term, where λ is the weighting coefficient.

[0103] S500: During the testing phase, the model no longer uses any label information; after inputting a complete image, the network outputs a candidate probability map; candidate pixels are extracted by setting a loose threshold, and a set of candidate regions is generated through connected component analysis; the maximum or average probability of each candidate region is used as its confidence index for sorting or filtering; the final output includes the set of candidate regions and their corresponding confidence scores, rather than the final event classification result; this design allows this method to serve as an upstream module for subsequent fine-tuning, multi-spectral consistency verification, or manual review;

[0104] Specifically, including,

[0105] During the inference phase, given an unlabeled Landsat 8 thermal infrared brightness temperature map... (and optional) ), obtained directly from forward computation and The candidate cell set is determined by... Obtained by applying a relaxed threshold:

[0106] (10)

[0107] Among them, threshold The threshold is set to ensure candidate recall (typically significantly lower than the final decision threshold). Then, [the following is done / then...] Perform connected component analysis to aggregate candidate cells into a candidate region set. And output the candidate confidence score for each candidate region (e.g. (or regional mean) is used for sorting and stratified output; the output of this stage is "candidate region + confidence level", which is positioned to provide upstream candidate input with high recall and controllable false alarms for subsequent research and analysis (such as further multi-spectral consistency verification, spatiotemporal consistency test, manual review or physical modeling), rather than directly giving the final thermal anomaly category conclusion.

[0108] Specific Implementation Scheme 2: The present invention provides a system for extracting candidate weak thermal anomalies under clouds and suppressing false anomalies in space-based thermal infrared remote sensing. This system has program modules corresponding to the above steps, and executes the steps in the above-mentioned method for extracting candidate weak thermal anomalies under clouds and suppressing false anomalies when running.

[0109] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0110] Specific Implementation Scheme 3: The present invention provides a computer-readable storage medium storing a computer program configured to, when called by a processor, implement the steps of a method for extracting candidate weak thermal anomalies under clouds and suppressing false anomalies in space-based thermal infrared remote sensing.

[0111] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0112] Example

[0113] This embodiment selects thermal infrared imagery acquired by the Landsat 8 satellite as input data, prioritizing TIRSBand 10 brightness temperature maps as the primary observation data, and optionally incorporating corresponding quality layer information to indicate the location of clouds, shadows, or saturated pixels. Before entering the algorithm, the input imagery undergoes radiometric correction and geometric registration, and is uniformly resampled to the same spatial resolution to ensure the accuracy of subsequent pixel-by-pixel processing. The experimental area is selected from scenes containing thin cloud cover, cloud boundary transitions, and complex surface backgrounds, targeting low-resolution thermal anomalies, such as weak heat sources with scales close to or smaller than a single pixel.

[0114] For sample construction, a sliding window approach was used to crop fixed-size image patches from the original image as training samples. Each sample includes corresponding thermal infrared brightness temperature data and weak annotation information for candidate regions. The annotation method uses a coarse candidate mask formed by center point expansion, that is, expanding a certain pixel radius around the target center position provided by manual annotation or external data to form regions that may have thermal anomalies. This annotation is only used to guide the model to learn candidate tendencies and does not require precise boundaries. To improve the suppression ability of gating branches, typical pseudo-anomaly regions such as cloud boundaries, high-reflectivity background areas, and stripe noise were deliberately included in the construction of negative samples.

[0115] In the specific implementation process, the input thermal infrared image is first fed into the candidate saliency learning module. This module extracts local spatial features through a convolutional neural network and outputs a candidate saliency map, which represents the probability that each pixel belongs to a weak thermal anomaly candidate. The design goal of this stage is to maximize the candidate recall rate so that weak anomalies under cloud conditions can still be triggered.

[0116] Subsequently, the model generates a gated map in parallel to represent the reliability of candidate responses. The gated map reflects whether the location is more likely to belong to a false anomaly region, such as a cloud boundary mixed pixel or a background structure interference region. By continuously modulating the candidate saliency map with the gated map, the final candidate probability map is obtained. This modulation process is numerically a continuous function, allowing the suppression behavior to be learned automatically through training, rather than relying on fixed rules.

[0117] During the training phase, a weakly supervised candidate mask is used to optimize the final candidate probability map, resulting in higher responses from real candidate regions and lower responses from background regions. Simultaneously, a negative sample strategy is employed to enable the gating branch to learn to increase suppression strength in pseudo-anomaly regions. Model parameters are updated using the standard backpropagation algorithm until stable convergence is achieved on the validation data.

[0118] During the testing phase, complete Landsat 8 thermal infrared images were input into the trained model to obtain a candidate probability map. Candidate pixels were extracted by setting a relaxed threshold, and connected component analysis was used to aggregate the candidate pixels into a set of candidate regions. The confidence score of each candidate region was also output for subsequent ranking or filtering. This output retains high sensitivity to weak thermal anomalies under clouds while effectively suppressing false anomalies caused by cloud boundaries and background structures through a gating mechanism.

[0119] In simulation experiments, the method of this invention was compared with traditional fixed-threshold candidate extraction methods. Experimental results show that in scenarios with thin cloud cover, this method can significantly improve the recall rate of weak thermal anomaly candidates, while effectively reducing the number of false positives under complex background conditions. Compared with algorithms that rely solely on threshold rules, this invention achieves improved stability and enhanced interpretability of candidate extraction while maintaining essentially no increase in computational complexity.

[0120] Furthermore, in other embodiments, the network structure of the candidate saliency model, the form of the gating modulation function, and the selection of the input data channel can all be replaced or extended according to the specific application scenario. As long as they contain the basic idea of ​​candidate saliency learning and gating suppression, they all fall within the protection scope of this invention.

[0121] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for candidate extraction and pseudo-anomaly suppression of weak thermal anomalies under clouds in space-based thermal infrared remote sensing, characterized in that, Includes the following steps: S100. During the input phase, the thermal infrared brightness temperature map of Landsat 8 is received as the data source, and quality layer information can be used as an auxiliary input. To ensure spatial consistency, all input data have undergone basic geometric alignment and resampling processing before entering the network. S200. During the candidate learning stage, the network extracts features from the input thermal infrared brightness temperature image through a convolutional structure and outputs a candidate saliency map; the candidate saliency map is used to represent the probability that each pixel belongs to a weak thermal anomaly candidate. Weakly labeled masks are used for supervision during training; S300. In the gating suppression stage, the network outputs a gating map in parallel to characterize the reliability of the candidate response at each pixel, that is, to reflect whether the pixel is more likely to belong to the pseudo-anomaly region; by continuously multiplying and combining the candidate saliency map with the gating map, the final candidate probability map is obtained. S400. During the training phase, positive and negative sample patches constructed using Landsat 8 are used for supervised learning. Positive samples are windows containing known thermal anomaly centers, while negative samples are background windows without thermal anomalies or regions prone to false anomalies. Network parameters are updated through standard backpropagation. Gated branches indirectly learn the ability to suppress false anomalies through joint loss and negative sample sampling strategies. S500: During the testing phase, after inputting a complete image, the network outputs a candidate probability map; candidate pixels are extracted by setting a lenient threshold, and a set of candidate regions is generated through connected component analysis; the maximum or average probability of each candidate region is used as its confidence index for sorting or filtering; the final output includes the set of candidate regions and their corresponding confidence scores.

2. The method for candidate extraction and pseudo-anomaly suppression of weak thermal anomalies under clouds in space-based thermal infrared remote sensing according to claim 1, characterized in that: In step S100, let the thermal infrared brightness temperature diagram of the nth sample be... ,in Represents pixel coordinates, This represents the image domain, i.e., the pixel coordinate x belongs to the image region Ω; when a Landsat 8 quality layer is available, it is denoted as... This is used as an auxiliary input feature and stitched into the thermal infrared brightness temperature map to form a multi-channel input. In cases where a reliable quality layer cannot be obtained, only the brightness temperature map is used as input.

3. The method for candidate extraction and false anomaly suppression of weak thermal anomalies under clouds in space-based thermal infrared remote sensing according to claim 2, characterized in that: In step S100, to reduce the difference in brightness temperature distribution between different scenes, the following steps are taken: Standardize the process.

4. The method for candidate extraction and false anomaly suppression of weak thermal anomalies under clouds in space-based thermal infrared remote sensing according to claim 3, characterized in that: Step S200 includes, Using a lightweight convolutional neural network , will input Mapping to candidate saliency map ,Right now: （1） in, Used to represent the probability that pixel x belongs to the weak thermal anomaly candidate; Weakly supervised labeling masks are used during the training phase. As a supervisory signal, let the candidate labels for the corresponding samples be: （2） in, This indicates that the pixel is located within a candidate region for thermal anomalies, either manually or using external data. This indicates that the pixel is located in the background area; Using pixel-level binary cross-entropy loss function As a basic form of supervision: （3） The pixel-level binary cross-entropy loss function encourages the network to output a higher response on pixels labeled as candidate regions and a lower response on background pixels. After training, during the testing phase, unlabeled Landsat 8 thermal infrared images are input into the network to obtain candidate saliency maps. Subsequently, a lenient threshold was set. Construct the initial candidate set : （4） Pixel-level candidates are transformed into a set of candidate regions through connected component analysis or fixed-size window aggregation. .

5. The method for candidate extraction and pseudo-anomaly suppression of weak thermal anomalies under clouds in space-based thermal infrared remote sensing according to claim 4, characterized in that: Step S300 includes, The network structure employs a shared encoder and dual-branch output: candidate branch outputs are defined. Gated branch output ;in This value is used to represent the unreliability or tendency for false anomalies in candidate responses at that pixel; a larger value indicates that it should be suppressed more. The final candidate probability map is constructed based on these two factors. It is implemented using a multiplicative gating method: （5）。 6. The method for candidate extraction and false anomaly suppression of weak thermal anomalies under clouds in space-based thermal infrared remote sensing according to claim 5, characterized in that: In step S400, the following are included: Suppose the training samples are from the thermal infrared brightness temperature map of Landsat 8, denoted as . Optional input of quality layer or QA information. Network input is The output is a candidate saliency map. With gated graph Candidate probability maps are obtained using multiplicative gating. ; Training supervision uses weakly labeled candidate masks; The optimization goal is and Consistency is achieved by training candidate outputs using pixel-level binary cross-entropy. for: （6） Adding lightweight constraints to the training process Apply weak regularization to the overall distribution, or introduce window-level gating supervision to the negative sample patch; Final training loss for: (7) Among them, This represents a weakly regularized gated or window-level supervised term, where λ is the weighting coefficient.

7. The method for candidate extraction and false anomaly suppression of weak thermal anomalies under clouds in space-based thermal infrared remote sensing according to claim 6, characterized in that: Step S500 includes, During the inference phase, given an unlabeled Landsat 8 thermal infrared brightness temperature map... Forward calculation yields and ; Candidate cell set through the Obtained by applying a relaxed threshold: （8） Among them, threshold The goal was to ensure candidate recall; subsequently, [the following was done / implemented]. Perform connected component analysis to aggregate candidate cells into a candidate region set. It outputs candidate confidence scores for each candidate region for sorting and hierarchical output.

8. A system for extracting candidate weak thermal anomalies under clouds and suppressing false anomalies in space-based thermal infrared imaging, characterized in that: The system has a program module corresponding to the steps of any one of the claims 1-7 above, and executes the steps in the above-described method for extracting candidate weak thermal anomalies under clouds and suppressing false anomalies when running.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the method for candidate extraction and pseudo-anomaly suppression of weak thermal anomalies under clouds in space-based thermal infrared remote sensing according to any one of claims 1-7.

Images