Intelligent acquisition method for specific environment target sample based on artificial intelligence and remote sensing technology

Abstract

The present application relates to the technical field of environmental monitoring target identification, and in particular to a specific environmental target sample intelligent collection method based on artificial intelligence and remote sensing technology. The method comprises the following steps: remote sensing data acquisition and environmental adaptation preprocessing; specific environmental sample library initial construction; AI model training and target sample initial collection; sample environment adaptation correction and standardization; sample library iterative optimization and model updating. Through the technical means of target partition-based degradation consistency inversion correction, partition differentiation error correction, and sample credibility shunting and standardization, the present application can correct the distortion of ground object spectral features caused by environmental degradation factors in remote sensing sample data, and simultaneously correct the deviation of target labeling boundaries and real ground object contours. Based on the corrected high-quality standard samples, model training can reduce the misjudgment and omission probability of the model for real scene ground object features, and improve the reliability of the model identification results.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring target identification technology, and more specifically, to a method for intelligent collection of specific environmental target samples based on artificial intelligence and remote sensing technology. Background Technology

[0002] The deep integration of artificial intelligence and remote sensing technology has provided key technical support for the automated interpretation of large-scale, high-frequency remote sensing images in the fields of ecological environment monitoring and specific target identification. Currently, general-purpose publicly available remote sensing datasets, AI algorithm modules of mainstream remote sensing software platforms, and third-party open-source deep learning models can already achieve basic functions of ecological type change detection and environmental target identification, and have been widely used in remote sensing image interpretation tasks in general scenarios.

[0003] However, existing technologies still have the following technical shortcomings. Specifically, current remote sensing sample data used for ecological change detection and specific environmental target identification are easily affected by complex atmospheric and lighting degradation factors, leading to distortion of the spectral features of ground objects and deviations between the target annotation boundaries and the actual outlines of ground objects. Even models trained on general datasets may misjudge or miss ground object features in real scenes due to uncorrected environmental degradation errors in the samples, making it difficult to meet the actual needs of environmental monitoring operations in terms of identification accuracy. In view of this, we propose an intelligent acquisition method for specific environmental target samples based on artificial intelligence and remote sensing technology. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent acquisition method for specific environmental target samples based on artificial intelligence and remote sensing technology, in order to solve the problem mentioned in the background art that existing remote sensing samples are easily affected by environmental degradation factors such as atmosphere and light, resulting in distortion of the spectral features of ground objects and deviation between the target annotation boundary and the real ground object contour, which in turn leads to the model misjudging and missing the features of ground objects in the real scene, and the recognition accuracy is difficult to meet the actual needs of environmental monitoring operations.

[0005] To address the aforementioned technical problems, this invention provides an intelligent acquisition method for specific environmental target samples based on artificial intelligence and remote sensing technology, comprising the following steps: S1. Remote sensing data acquisition and environmental adaptation preprocessing: Acquire multi-source remote sensing images under specific conditions, suppress or correct environmental interference factors in the multi-source remote sensing images, and obtain preprocessed images. S2. Initial construction of a sample library for specific environments: The target information in the preprocessed images is labeled using artificial intelligence annotation methods to build an initial sample library for model training. S3. AI Model Training and Initial Target Sample Collection: Based on the initial sample library, a target recognition model is trained, and the target recognition model is used to perform target detection on newly added remote sensing images to obtain a coarse sample set. S4. Sample Environment Adaptation Correction and Standardization: The coarse sample set is processed using a degradation consistency inversion method. Based on the target annotation information, target partitioning is performed and environmental degradation parameters are estimated. Candidate correction and degradation observation space remapping are completed. Partitioning errors are compared and parameters and annotation boundaries are iteratively corrected. Samples are split based on confidence level and standardized to generate a high-quality standard sample set. The degradation consistency inversion refers to generating candidate correction samples based on sample environmental degradation parameters and mapping the candidate correction samples back to the degradation observation space to verify their consistency with the original coarse sample. The environmental degradation parameters, correction intensity, or annotation boundaries are corrected based on the verification results. S5. Sample Library Iterative Optimization and Model Update: Select qualified samples from the high-quality standard sample set to expand the initial sample library, and iteratively optimize the target recognition model based on the updated initial sample library to form a closed-loop process linking sample collection and model optimization.

[0006] As a further improvement to this technical solution, step S1, remote sensing data acquisition and environmental adaptation preprocessing, includes the following steps: S1.1 Acquire multi-source remote sensing images of the target monitoring area, integrate image data from different sensors and different imaging time periods, and construct an initial remote sensing image dataset; S1.2 Identify interference factors in the initial remote sensing image and extract environmental interference features such as clouds, water vapor, aerosols, shadows and illumination deviations contained in the image. S1.3. For interference from clouds, fog, water vapor, and aerosols, corresponding atmospheric correction algorithms are used for suppression. S1.4. For interference such as shadows and lighting deviations, corresponding lighting correction algorithms are used for correction. S1.5 Perform consistency verification on the processed image to obtain a preprocessed image with uniform image quality and controllable interference.

[0007] As a further improvement to this technical solution, the initial construction of the specific environment sample library in step S2 includes the following steps: S2.1 Filter the target region of the preprocessed image, define the target object to be labeled and its neighborhood range under a specific environment, and generate a set of images to be labeled; S2.2. Use an artificial intelligence-assisted annotation model to perform preliminary annotation of target information in the image set to be annotated, and generate initial annotation data containing target category, location and boundary; S2.3 Perform consistency verification and noise filtering on the initial annotation data, correct annotation deviations, remove invalid annotations, and obtain standard annotation data; S2.4. Associate and match the standard labeled data with the corresponding preprocessed images, and convert them into a sample format suitable for training the target recognition model; S2.5. Classify and store the associated sample data according to target category and environmental scenario, establish a sample index and version management mechanism, and build an initial sample library.

[0008] As a further improvement to this technical solution, step S3, AI model training and initial target sample acquisition, includes the following steps: S3.1 Divide the initial sample library into a training set, a validation set and a test set, and allocate samples of each category according to a preset ratio to ensure a balanced sample distribution; S3.2. Train the preset deep learning target recognition model based on the training set, adjust the model parameters iteratively through the validation set, and evaluate the model recognition accuracy using the test set to obtain the trained target recognition model. S3.3 Acquire newly added remote sensing images under specific environments, perform environmental interference suppression processing on the newly added remote sensing images, and obtain the images to be detected; S3.4. Use the target recognition model to perform target detection on the image to be detected, and output the detection results including the target location, category and confidence level; S3.5. Filter valid detection results according to the preset reliability threshold, associate and store the corresponding target area image, annotation information and metadata, and construct a coarse sample set.

[0009] As a further improvement to this technical solution, step S4, sample environment adaptation correction and standardization, includes the following steps: S4.1. Use the degradation consistency inversion method to process the coarse sample set and retrieve the annotation information, regional image information, neighborhood background information and remote sensing image metadata corresponding to the target sample. S4.2. Based on the target annotation information, perform target partitioning and estimate the environmental degradation parameters corresponding to the target samples; S4.3. Generate candidate correction samples based on environmental degradation parameters and target partitioning results, and map the candidate correction samples to the degradation observation space to obtain re-degraded samples; S4.4. Compare the partition consistency error between the degraded sample and the original coarse sample, and iteratively correct the environmental degradation parameters, local correction intensity or annotation boundary based on the partition consistency error; S4.5 Calculate sample credibility based on partition consistency error, complete sample diversion based on credibility, and standardize the diverted and retained samples to generate a high-quality standard sample set.

[0010] As a further improvement to this technical solution, the target partitioning and environmental degradation parameter estimation in S4.2 includes the following steps: S4.21. Based on the target's labeled boundary coordinates, divide the target sample into three continuous regions: the target interior region, the target boundary region, and the target neighborhood background region. S4.22 Extract the spectral, radiometric and textural features of the target neighborhood background area, and simultaneously acquire image imaging parameters, sensor parameters and atmospheric related metadata; S4.23. Based on the principle of radiative transfer, using pure pixels, near-pure pixels, or statistically representative pixels in the background area of ​​the target neighborhood, a set of environmental degradation parameters, including atmospheric transmittance and atmospheric comprehensive radiation, is estimated.

[0011] As a further improvement to this technical solution, the candidate correction and degraded observation space remapping in S4.3 includes the following steps: S4.31. Based on the target partitioning results, assign a corresponding local correction intensity to each pixel. , The value range is [0,1]; S4.32, Based on environmental degradation parameters and local correction intensity The original coarse sample is radiometrically corrected to generate candidate corrected samples; a lower limit for atmospheric transmittance is set to avoid calculation anomalies. S4.33. Using the same environmental degradation parameters as in step S4.32, the candidate correction samples are positively mapped back to the original degradation observation space to obtain the re-degraded samples. S4.34. Unify the radiation value range between the re-degraded sample and the original coarse sample to complete the data benchmark alignment process.

[0012] As a further improvement to this technical solution, the partitioning error comparison and parameter and annotation boundary correction in S4.4 includes the following steps: S4.41 Calculate the root mean square error of pixels between the re-degraded sample and the original coarse sample in the target interior region, target boundary region and target neighborhood background region respectively, and use it as the partition consistency error of the corresponding region. S4.42. When the partition consistency error of the target's internal region exceeds the first preset threshold, adjust the environmental degradation parameter value or the local correction intensity. ; S4.43. When the partition consistency error of the target boundary area exceeds the second preset threshold, the coordinates of the target annotation boundary are finely adjusted along the error gradient direction. S4.44 When the partition consistency error of the target neighborhood background area exceeds the third preset threshold, mark the sample as having background mixing anomaly.

[0013] As a further improvement to this technical solution, the iteration termination determination in S4.4 includes the following steps: S4.45. Assign weight coefficients to the target interior region, target boundary region, and target neighborhood background region respectively. ,satisfy The weighting coefficients can be preset or adaptively determined based on the target category, image quality, or application scenario. S4.46. Based on the weight coefficients of each partition, calculate the total partition consistency error corresponding to a single iteration. S4.47. When the total partition consistency error is lower than the preset convergence threshold, or the number of iterations reaches the preset upper limit, stop the iteration operation; S4.48 Output the final iteratively generated samples as environment adaptation and correction samples, and retain the corresponding environment degradation parameters, local correction intensity and annotation boundary information.

[0014] As a further improvement to this technical solution, the sample credibility assessment and standardization output of S4.5 includes the following steps: S4.51 Calculate sample reliability based on the final total partition consistency error. The smaller the overall partition consistency error, the higher the sample reliability. The higher, The value range is (0,1]; S4.52, will satisfy The samples were marked as qualified and retained; S4.53, will satisfy The samples were marked as samples to be reviewed and transferred to the manual review process; S4.54, will satisfy The samples were removed; S4.55. Standardize the format, size, coordinate reference and annotation content of the retained qualified samples to generate a high-quality standard sample set.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention uses techniques such as degradation consistency inversion correction based on target partitions, partition difference error correction, and sample credibility diversion and standardization to correct the distortion of spectral features of ground objects in remote sensing sample data caused by environmental degradation factors such as atmosphere and illumination. At the same time, it corrects the deviation between the target annotation boundary and the real ground object outline. Based on the corrected high-quality standard samples, the model training can reduce the probability of misjudging and missing ground object features in real scene, improve the reliability of model recognition results, and adapt to the actual application needs of environmental monitoring business. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the overall steps of the intelligent acquisition method for target samples in a specific environment in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, this embodiment provides a method for intelligent acquisition of specific environmental target samples based on artificial intelligence and remote sensing technology, including the following steps: S1. Remote sensing data acquisition and environmental adaptation preprocessing: Acquire multi-source remote sensing images under specific conditions, suppress or correct environmental interference factors in the multi-source remote sensing images, and obtain preprocessed images. In this step, S1, the remote sensing data acquisition and environmental adaptation preprocessing includes the following steps: S1.1 Acquire multi-source remote sensing images of the target monitoring area, integrate image data from different sensors and different imaging periods, and construct an initial remote sensing image dataset; specifically, acquire Gaofen-4 visible light images and Sentinel-2 multispectral images of the target monitoring area, covering the period from January to June 2024, including typical environmental conditions such as sunny days, light fog, cloudy days, and dawn and dusk.

[0019] Furthermore, the georegistration information, imaging timestamps, and radiometric calibration parameters of each image are obtained. All images are converted to the WGS-84 coordinate system, resampled to a spatial resolution of 10 meters per pixel through bilinear interpolation, and cataloged according to imaging timestamps to construct an initial remote sensing image dataset.

[0020] S1.2 Identify interference factors in the initial remote sensing images and extract environmental interference features such as clouds, fog, water vapor, aerosols, shadows, and illumination deviations. Specifically, identify the following interference factors for each initial image: Interference from clouds and water vapor: Calculate the dark channel map of the image, identify continuous areas with dark channel pixel brightness values ​​higher than 180 (8-bit image) as cloud or water vapor obstruction areas, and generate a binary mask; Shadow interference: Convert the image to HSV space, extract the luminance channel, and use Otsu's method for automatic thresholding. Continuous areas with luminance values ​​below the threshold are selected as candidate shadow areas. Calculate the gradient intensity at the edge of the candidate area. If the average gradient of the area is lower than the average gradient of the entire image, it is confirmed as a shadow area; otherwise, it is excluded. Localized uneven illumination: The image is divided into a 16×16 grid, and the average brightness of each block is calculated. If the average value of a block deviates from the average value of the whole image by more than ±30%, it is marked as an area with uneven illumination.

[0021] Finally, the interference types and spatial distribution of each image are recorded to form an interference feature description file.

[0022] S1.3. For interference from clouds, fog, water vapor, and aerosols, corresponding atmospheric correction algorithms are used for suppression. Specifically, for cloud and fog interference, a dark channel prior dehazing algorithm is used for pixel-by-pixel correction. Using the unoccluded area determined in S1.2, the original image pixels corresponding to the top 0.1% of the brightest pixels in the dark channel image are selected, and the average value of the RGB three channels is calculated as the global atmospheric light value. ;remember for The components in the corresponding color channel; Estimating transmittance ;in For pixel coordinates, Original image channel exist Pixel value at; Using the original image grayscale as the guide map, a guide filter is used to smooth the transmittance map. The filter radius is set to 1 / 30 of the long side of the image, and the regularization parameter is... ; according to Restore the fog-free image. In the formula... For the original image in The radiation value at that location is truncated with a denominator threshold of 0.1 to avoid division by zero.

[0023] To mitigate aerosol interference, atmospheric correction was performed using the aerosol optical thickness from image-aided data and a 6S radiative transfer model.

[0024] S1.4. For interference such as shadows and lighting deviations, corresponding lighting correction algorithms are used for correction; specifically, homomorphic filtering is used for global lighting correction. Perform a logarithmic transformation on each band of the image to convert the multiplicative model into an additive model; Perform a Fast Fourier Transform (FFT); Constructing a high-pass Gaussian filter ,in For frequency domain coordinates, For high-frequency gain, For low-frequency gain, Frequency point Euclidean distance to the center of the spectrum The cutoff frequency, The pixel length of the image diagonal; The spectrum is multiplied using this filter and then subjected to an inverse fast Fourier transform (IFFT). Perform an exponential transform to restore pixel values.

[0025] Subsequently, local enhancement was performed using Limit Contrast Adaptive Histogram Equalization (CLAHE), with the grid divided into 8×8 sections and the contrast limiting threshold set to 2.0. Bilinear interpolation was then used to eliminate inter-block boundary effects.

[0026] After processing, details in shadow areas are restored, and the illumination distribution becomes more uniform. Homomorphic filtering and CLAHE can be replaced with other equivalent illumination correction techniques.

[0027] S1.5. Perform consistency verification on the processed images to obtain preprocessed images with uniform image quality and controllable interference. Specifically, images are selected according to the following criteria: Radiometric consistency: Calculates the mean pixel values ​​of each band in the processed image. and standard deviation ,like Exceeding the mean of the original image by ±10%, or Items below 50% of the original standard deviation are excluded. Spatial consistency: Calculate the BRISQUE no-reference image quality score and remove images with a score higher than 40; Visual inspection: 5% of the images from the first two steps are randomly sampled for inspection to confirm the absence of artifacts, color distortion, or edge ringing effects. If unqualified images are found during the inspection, the images in the same batch are returned to S1.3 or S1.4 for parameter adjustment and reprocessing.

[0028] Finally, images that pass all verifications are retained as preprocessed images, resulting in images with uniform quality and controllable interference.

[0029] S2. Initial Construction of a Sample Library for Specific Environments: This involves using AI-based annotation to label target information in preprocessed images, building an initial sample library suitable for model training. Specifically, this includes the following steps: S2.1 The preprocessed image is subjected to target region screening to define the target objects and their neighborhoods under specific environmental conditions, generating a set of images to be labeled. Specifically, the preprocessed image output in S1 is subjected to target region screening. Based on the specific environmental target type of the task (e.g., chimneys, storage tanks, cooling towers in industrial facilities), the target morphology is used for screening and definition. For example, chimneys are long and thin, and may have emission plumes or shadows at the top. Screening ensures the target body is fully visible, and the neighborhood includes a background area 1.5 times the target height. For storage tanks, these targets are cylindrical, often appearing circular or elliptical in overhead views, and are frequently distributed in groups. Screening ensures the tank and its base are fully included, and the neighborhood includes a surrounding area 0.5 times the target diameter. The defined target areas are cropped from the original image into independent slices, the geographic coordinates of each slice are recorded, and the slices are uniformly scaled to 1024×1024 pixels to generate a set of images to be labeled.

[0030] S2.2 An AI-assisted annotation model is used to initially annotate the target information in the image set to be annotated, generating initial annotation data containing target category, location, and boundary. Specifically, a pre-trained segmentation model based on SAM (SegmentAnything Model) is used for initial annotation of the image set to be annotated. For each image slice, the SAM model automatically extracts candidate masks. For slender targets (such as chimneys), the vertical line from the image center is used as a cue input to guide segmentation; for near-circular targets (such as storage tanks), the image center point is used as a cue input. The model outputs the instance segmentation mask and bounding box coordinates of the target, which are then assigned category labels (e.g., chimney, storage tank) by the operator, generating initial annotation data containing target category, location coordinates, and boundary mask.

[0031] S2.3. Perform consistency verification and noise filtering on the initial annotation data, correct annotation deviations, and remove invalid annotations to obtain standard annotation data; specifically, perform the following verification processing on the initial annotation data: Consistency check: Cross-compare the annotation results of the same target in different image slices or different time phases. If the boundary intersection-union ratio (IoU) of multiple annotations of the same target is less than 0.5, it is marked as inconsistent annotation and returned to manual review and correction. Noise filtering: Remove labels that are severely obscured, resulting in a visible area of ​​less than 50% of the typical target area; low-quality labels with a confidence level below 0.7; and labels that are obviously mislabeled as non-target features. Label boundary correction: For the filtered labels, the Conditional Random Field (CRF) algorithm is used to refine the boundary of the segmentation mask by using image edge information as a guide, and the correction range is controlled within 5 pixels.

[0032] Finally, after the above processing, standard labeled data is obtained.

[0033] S2.4. Associate and match the standard labeled data with the corresponding preprocessed images, converting it into a sample format suitable for training the target recognition model. Specifically, associate and match the standard labeled data obtained in S2.3 with the corresponding preprocessed image slices one by one. Perform format conversion according to the training input requirements of the target recognition model (such as EnviroDetectNet): convert the instance segmentation mask into a horizontal bounding box, and use TXT format for the labeling file. Each line records the target category number (e.g., chimney is numbered 0, storage tank is numbered 1, and other categories are sequentially incremented) and the normalized bounding box coordinates. ,in and The normalized x and y coordinates of the center point of the bounding box in the image (range [0,1]). and This provides the normalized width and height of the bounding box. Geographic coordinates and original mask annotations are also retained as auxiliary data.

[0034] S2.5. Classify and store the associated sample data according to target category and environmental scenario, establish a sample index and version management mechanism, and build an initial sample library. Specifically, classify and store the format-converted samples according to two dimensions: target category (such as chimney, storage tank, etc.) and environmental scenario (such as industrial area, town, suburb, etc.).

[0035] A sample index file is created, recording the file path, target category, scene type, number of bounding boxes, geographic coordinates, and timestamp of entry into the database for each sample in JSON format. A version management mechanism is adopted, with the initial sample database version designated as V1.0, and the version number updated incrementally as samples are added or corrected.

[0036] Finally, the sample images, annotation files, index files, and version information are archived together to form an initial sample library for specific environment targets. The format and organization of the sample library conform to the requirements of deep learning dataset specifications.

[0037] S3. AI Model Training and Initial Target Sample Collection: A target recognition model is trained based on an initial sample library. This model is then used to perform target detection on newly added remote sensing images, obtaining a coarse sample set. Specifically, this includes the following steps: S3.1 Divide the initial sample library into a training set, a validation set, and a test set, and allocate samples of each category according to a preset ratio to ensure a balanced sample distribution; specifically, divide the initial sample library constructed in S2 into a training set, a validation set, and a test set according to a preset ratio.

[0038] This embodiment uses a 7:2:1 partitioning ratio, that is, the training set accounts for 70%, the validation set accounts for 20%, and the test set accounts for 10%.

[0039] A stratified sampling strategy is adopted during the segmentation: the sample is segmented according to two dimensions: target category (such as chimney, storage tank, etc.) and environmental scene (such as industrial area, town, etc.) to ensure that the proportion of each category and scene in each subset is roughly consistent with the overall sample library, and to avoid the model's identification bias for specific categories or scenes due to uneven sample distribution.

[0040] After the partitioning is completed, the sample size, category distribution, and scene distribution of each subset are counted. Once the distribution balance meets the preset requirements, the subset partitioning results are locked.

[0041] S3.2. Train the preset deep learning target recognition model based on the training set, iteratively adjust the model parameters through the validation set, and evaluate the model's recognition accuracy using the test set to obtain the trained target recognition model. Specifically, train the preset deep learning target recognition model based on the training set. This embodiment uses the EnviroDetectNet model as an example. This model uses the YOLOv8 architecture as the baseline network and does not make substantial modifications to the YOLOv8 baseline network structure. Transfer learning is performed based on the pre-trained weights.

[0042] The following configuration was used during training: the optimizer was stochastic gradient descent (SGD), the initial learning rate was set to 0.01, the momentum was set to 0.937, and the weight decay coefficient was set to 5 × 10⁻⁶. 4 The batch size is set to 16 based on the GPU memory capacity; the maximum number of training epochs is set to 300. After each training epoch, the mean precision (mAP@0.5) is calculated on the validation set as an evaluation metric, and an early stopping strategy is adopted: if the mAP@0.5 on the validation set does not improve for 20 consecutive epochs, training is stopped and the system is rolled back to the optimal weights.

[0043] After training, evaluate the model's recognition accuracy on the test set, calculating mAP@0.5 and precision. and recall rate ,in , , For the actual number of cases, The number of false positives This represents the number of false negatives. When the test set mAP@0.5 reaches the preset accuracy requirement (e.g., above 0.85), the model is considered a successfully trained target recognition model; otherwise, the hyperparameters are adjusted and the model is retrained until the requirement is met.

[0044] S3.3 Acquire new remote sensing images under specific conditions, and perform environmental interference suppression processing on the new remote sensing images to obtain the images to be detected; specifically, acquire new remote sensing images under specific conditions. The acquisition range of the new images covers the same or similar monitoring area as the initial sample library, and the imaging time is later than the acquisition time of the images used in the initial sample library, to ensure that the new images are unseen data for the model.

[0045] After acquisition, the newly added remote sensing images undergo the same environmental interference suppression processing as S1: including interference factor identification, atmospheric correction, illumination correction, and consistency verification, to obtain images with uniform image quality and controllable interference. The sensor type, spatial resolution, and coordinate reference system of the newly added images are consistent with those used in the initial sample library.

[0046] S3.4. Apply the target recognition model to detect targets in the image to be detected, and output the detection results including target location, category, and confidence level. Specifically, input the image to be detected obtained in S3.3 into the target recognition model trained in S3.2, and perform image-by-image inference detection. The model outputs all target information detected in each image, including: target bounding box coordinates (in pixels), target category (such as chimney, storage tank, etc.), and corresponding confidence level. Among them, confidence level This represents the model's predicted probability that the target of that class exists within the detection bounding box. Detection results are output in a structured format (such as JSON or XML) on a per-image basis.

[0047] S3.5. Filter valid detection results according to a preset reliability threshold, and associate and store the corresponding target area image, annotation information, and metadata to construct a coarse sample set. Specifically, based on the preset reliability threshold... Valid test results are filtered. This embodiment sets... Confidence level Detection boxes that meet the threshold are marked as valid detections, while detection boxes that are below the threshold are discarded.

[0048] For each valid detection result, the image region within the corresponding bounding box range is cropped from the original image to be detected as a target slice. The category label output by the model and the bounding box coordinates are used as the initial annotation information, and the image's imaging time, sensor parameters, geographic coordinates and other metadata are synchronously associated.

[0049] The target slices, initial annotation information, and metadata are stored one by one to form a coarse sample set. This coarse sample set is the result of automatic collection by the model, and its annotation accuracy is limited by the model's own performance. There may be boundary offsets or misclassifications, which will be further corrected in step S4.

[0050] S4. Sample Environment Adaptation Correction and Standardization: A degradation consistency inversion method is used to process the coarse sample set. Based on the target annotation information, target partitioning is performed and environmental degradation parameters are estimated. Candidate correction and degradation observation space remapping are completed. Partition errors are compared and parameters and annotation boundaries are iteratively corrected. Samples are split based on confidence level, and a high-quality standard sample set is generated after standardization. Specifically, the degradation consistency inversion refers to generating candidate correction samples based on sample environmental degradation parameters, mapping the candidate correction samples back to the degradation observation space to verify their consistency with the original coarse sample, and correcting environmental degradation parameters, correction strength, or annotation boundaries based on the verification results. Specifically, it includes the following steps: S4.1. Use the degradation consistency inversion method to process the coarse sample set and retrieve the annotation information, regional image information, neighborhood background information and remote sensing image metadata corresponding to the target sample. In this step, for each coarse sample in the coarse sample set, its associated data is retrieved, including: target annotation boundary information (boundary box coordinates or mask) output by the model, image tile data of the corresponding region, background area image with a certain range extended outward from the annotation boundary (such as 1.5 times the target scale) as neighborhood background information, and remote sensing image metadata such as the image's imaging time, sensor parameters, solar zenith angle, and aerosol optical thickness.

[0051] Furthermore, the core processing method in this step is degradation consistency inversion, the basic idea of ​​which is as follows: The coarse sample comes from real remote sensing images, which are inevitably affected by environmental degradation factors such as atmospheric scattering and illumination changes during the imaging process. If the environmental degradation parameters of the sample (such as atmospheric transmittance and path radiance) can be accurately estimated, radiometric correction can be performed on the coarse sample to recover a theoretically unaffected "clean" target image, i.e., a candidate correction sample. Conversely, if the environmental degradation parameters are estimated correctly, the candidate correction sample is re-applied with the same degradation parameters to obtain a re-degraded sample, which should be highly consistent with the original coarse sample in radiometric characteristics.

[0052] Based on this principle, degradation consistency inversion achieves fine calibration of samples through the following closed loop: First, the initial degradation parameters are estimated based on background area information and the principle of radiative transfer to generate candidate calibration samples; then, the candidate calibration samples are mapped back to the degradation observation space to obtain re-degraded samples; then, the consistency error between the re-degraded samples in each partition and the original coarse samples is compared; finally, the degradation parameters, local calibration intensity, and even the labeled boundaries are corrected in reverse according to the error, and the iteration continues until the error converges.

[0053] This method can automatically correct quality problems in coarse samples caused by environmental degradation and labeling bias without manual annotation of the true values.

[0054] S4.2. Based on the target annotation information, perform target partitioning and estimate the environmental degradation parameters corresponding to the target samples; specifically, this includes the following steps: S4.21. Based on the target's labeled boundary coordinates, divide the target sample into three continuous regions: the target interior region, the target boundary region, and the target neighborhood background region. In this step, the sample region is continuously divided into target interior regions based on the annotation boundaries of the current coarse sample. Target boundary area and target neighborhood background area The division method is as follows: using the marked boundary curves. Using this as a baseline, shrink inward by 3 pixels to obtain the internal region. ;Depend on The ring-shaped region between the inward contraction of 3 pixels and the outward expansion of 3 pixels constitutes the boundary area. ;from A ring-shaped area extending outwards from 3 pixels to 15 pixels is designated as the background area. The three regions do not overlap and completely cover the sample area. The inner region corresponds to the target body, the boundary region contains the target edge transition zone, and the background region is a set of pure background pixels.

[0055] S4.22 Extract the spectral, radiometric and textural features of the target neighborhood background area, and simultaneously acquire image imaging parameters, sensor parameters and atmospheric related metadata; In this step, the background area is extracted. Mean value of pixel values ​​in each band and standard deviation ( , The total number of bands is used as a spectral feature. Background pixel values ​​are converted to apparent radiance using image calibration coefficients. The mean and variance of the data are calculated as radiation features. Then, the gray-level co-occurrence matrix of the background area is calculated, and contrast, homogeneity, and second angular moment are extracted as texture features to help determine whether the background features are uniform.

[0056] Simultaneously, the imaging parameters and atmospheric-related metadata of the image are acquired, mainly including: sensor spectral response function. Solar zenith angle Observing the zenith angle Relative azimuth Aerosol optical thickness Water vapor content Ozone concentration Solar irradiance at the top of the atmosphere (dimension: ) and Earth-Sun distance correction factor (Dimensionless, representing the ratio of the Earth-Sun distance to the average Earth-Sun distance on that day). The above data can be obtained from image header files, ESA Sentinel series auxiliary products, or NASA MERRA-2 meteorological reanalysis data.

[0057] S4.23. Based on the principle of radiative transfer, using pure pixels, near-pure pixels, or statistically representative pixels in the background area of ​​the target neighborhood, a set of environmental degradation parameters, including atmospheric transmittance and atmospheric comprehensive radiation, is estimated.

[0058] In this step, atmospheric transmittance is estimated using background region pixels based on the radiative transfer equation. and atmospheric integrated radiation items This constitutes the set of environmental degradation parameters { The apparent radiance received by the sensor. Compared with the true radiance of the Earth's surface The following conditions must be met: ; The true surface radiance (Lambertian surface assumption) converted from surface reflectance is as follows: ; in, The atmospheric transmittance for downward solar radiation is assumed to be equal for both upward and downward atmospheric transmittance in this method. , (where is the sensor's upward atmospheric transmittance). This assumption applies to flat observation areas with uniform atmospheric conditions.

[0059] Furthermore, based on the background area's landforms, one of the following three methods is selected for estimation: Approach 1: Solve simultaneously using dark and bright targets. When both dark targets (e.g., dense vegetation, typically ranging from 0.05 to 0.15) and bright targets (e.g., bare soil, typically ranging from 0.15 to 0.30) with known reflectance exist within the background area, let their surface reflectances be respectively... and The corresponding apparent radiance mean is and Substituting into the aforementioned Lambertian surface radiance formula: ; Substitution Substituting the assumptions into the dark and bright targets respectively, we obtain the following system of equations: ; Solving for the given information yields: ; ; Approach Two: Statistical Representative Pixel Regression. If the background area lacks pure pixels with known reflectance, then the 5%, 25%, 50%, 75%, and 95% quantile pixels, sorted by brightness, are selected as statistical representative pixels. Furthermore, the radiance variations of the same ground features in multiple temporal images of the same area, or reference ground features with known reflectance within the image, are used to construct a statistical representative pixel regression model. and The linear regression relationship is given by the slope of the regression line. The intercept is This method assumes that the surface reflectance of the same land cover remains constant at different time phases, and is suitable for observation areas with stable surface cover and no significant changes.

[0060] Approach 3: Direct simulation using a radiative transfer model. When neither of the above two conditions is met, the data obtained in S4.22 will be used. , , , , , and Input the 6S radiative transfer model and set empirical values ​​for the reflectance of typical surfaces in the background area (e.g., vegetation reflectance is set to 0.05 to 0.15, and bare soil reflectance is set to 0.15 to 0.30). The model will then output the results. and The estimated value.

[0061] Regardless of the approach used, the environmental degradation parameters of the coarse sample are ultimately obtained. } serves as the initial input for subsequent candidate corrections and iterative adjustments.

[0062] S4.3. Generate candidate correction samples based on environmental degradation parameters and target partitioning results, and map the candidate correction samples to the degradation observation space to obtain re-degraded samples; specifically, this includes the following steps: S4.31. Based on the target partitioning results, assign a corresponding local correction intensity to each pixel. , The value range is [0,1]; In this step, based on the target three-part partitioning result of S4.21, each pixel in the sample is... Distribute local correction intensity , The allocation rule is: target internal area. All pixels within ; Target neighborhood background area All pixels within ; Target boundary area Inside, By pixel Boundary and The distance to the boundary is linearly interpolated, with values ​​closer to 1.0 closer to the interior region and closer to 0.0 closer to the background region. This gradual transition ensures a smooth change in the correction intensity of the boundary region, avoiding abrupt radiation differences between the interior and background regions.

[0063] S4.32, Based on environmental degradation parameters and local correction intensity The original coarse sample is radiometrically corrected to generate candidate corrected samples; a lower limit for atmospheric transmittance is set to avoid calculation anomalies. In this step, the set of environmental degradation parameters estimated by S4.23 is used { } and local correction intensity of each pixel For the original coarse sample image Perform radiometric correction pixel by pixel to generate candidate correction samples The correction formula is: ; Indicates the original coarse sample in pixels The radiation value at that location Indicates candidate correction samples at pixels The radiation value at that location Atmospheric transmittance, For the comprehensive atmospheric radiation item, For pixels The local correction intensity at the location.

[0064] To prevent If the value is too small, the denominator will approach zero; therefore, a lower limit for atmospheric transmittance should be set. .when At that time, take After this correction, the atmospheric degradation effect in the interior region is completely removed, the background region values ​​remain unchanged, and the boundary region transitions smoothly between the two.

[0065] S4.33. Using the same environmental degradation parameters as in step S4.32, the candidate correction samples are positively mapped back to the original degradation observation space to obtain the re-degraded samples. In this step, candidate correction samples are... Remapping back to the original degraded observation space yields re-degraded samples. The re-degradation process uses the same environmental degradation parameters as S4.32. } and local correction intensity The formula is: ; Indicates the re-degraded sample at the pixel The radiation value at that location is consistent with S4.32 for the rest of the symbols.

[0066] The physical meaning of this step is to place the corrected, "clean" image back into its original atmospheric degradation environment. If... and If all estimates are accurate, then Should be with Pixel-by-pixel consistency; if there is a deviation between the two, it indicates an error in the estimation of degradation parameters or the allocation of correction intensity. This deviation will be used in S4.4 for iterative correction of driving parameters and annotation boundaries.

[0067] S4.34. Unify the radiation value range between the re-degraded sample and the original coarse sample to complete the data benchmark alignment process.

[0068] In this step, to avoid the slight offset caused by floating-point arithmetic precision or numerical truncation during the correction and re-degradation process affecting subsequent error comparisons, the re-degradation samples are... Compared with the original coarse sample Align the radiation values ​​across ranges. Calculate separately. minimum value Maximum value and minimum value Maximum value .by Based on the numerical range, Linear mapping to : ; This represents the pixel values ​​of the re-degraded sample after interval alignment. If Then directly set all pixels The above mapping will no longer be executed. Aligned and They are in the same radiation dimension and numerical range, which is used by S4.4 to calculate the partition error.

[0069] S4.4, compare the partition consistency error between the degraded sample and the original coarse sample, and iteratively correct the environmental degradation parameters, local correction intensity, or annotation boundary based on the partition consistency error; where: The following steps are used to compare partitioning errors and correct parameters and annotation boundaries: S4.41 Calculate the root mean square error of pixels between the re-degraded sample and the original coarse sample in the target interior region, target boundary region and target neighborhood background region respectively, and use it as the partition consistency error of the corresponding region. In this step, the internal area of ​​the target is... Target boundary area and target neighborhood background area Calculate the re-degraded samples respectively Compared with the original coarse sample The pixel-level root mean square error between regions is used as the partition consistency error for each region: ; ; ; in, , , These represent the total number of pixels in the three regions, respectively.

[0070] This error reflects the effectiveness of the internal region correction. If the atmospheric degradation parameters or the internal region correction intensity are inaccurate, the error will be too high. It reflects the accuracy of boundary area labeling and the effect of transition area correction. Offset of the labeled boundary will lead to an increase in the difference between the measured value and the re-degraded value in this area. It reflects the purity of the background area. If there are non-background pixels mixed in or the environmental degradation parameter estimation is biased, the error will increase.

[0071] S4.42. When the partition consistency error of the target's internal region exceeds the first preset threshold, adjust the environmental degradation parameter value or the local correction intensity. ; In this step, a first preset threshold is set. In this embodiment, .when This indicates a deviation in the internal region correction, which may be due to atmospheric transmittance. Or comprehensive atmospheric radiation item Inaccurate estimation, or internal zone correction strength The value was inappropriate.

[0072] The correction method is: keep Unchanged, for along Fine-tune the direction of decrease and adjust the step size. Take 1% of the current value; if adjusted back If it does not decrease, then... Make fine adjustments and adjust the step size. Take 1% of the current value; if adjusted separately and None of them effectively reduced Then the internal area Based on the current value, search and adjust within the range [0.8, 1.0] with a step size of 0.05. After each adjustment, re-execute the correction and re-degradation process of S4.32 and S4.33, and recalculate. .

[0073] S4.43. When the partition consistency error of the target boundary area exceeds the second preset threshold, the coordinates of the target annotation boundary are finely adjusted along the error gradient direction. In this step, a second preset threshold is set. In this embodiment, .when This indicates a significant re-degradation deviation in the boundary region, suggesting an offset between the labeled target boundary and the actual target boundary. The correction method is to adjust the labeled boundary curve... Key control points along The movement follows the negative gradient direction. For boundary points... Its update volume is: ; in, Indicates the first The coordinates of the boundary points at the next iteration The moving step size (in this embodiment, it is taken as...) (pixels) for exist The gradient at a given location is approximated using a 3×3 window center difference method. Each movement is no more than 1 pixel. After boundary adjustment, the target partitioning in S4.21, along with subsequent correction and re-degradation steps, must be re-executed.

[0074] S4.44 When the partition consistency error of the target neighborhood background area exceeds the third preset threshold, mark the sample as having background mixing anomaly.

[0075] In this step, a third preset threshold is set. In this embodiment, .when At this point, it is inferred that there may be anomalies in the background area, such as the incorporation of non-background features, abrupt changes in the type of background features, or severe local degradation affecting the area itself. Such cases are not automatically corrected at this step; instead, the sample is marked as "background incorporation anomaly," and this label will participate in the triage decision during the S4.5 credibility assessment stage.

[0076] The iteration termination determination includes the following steps: S4.45. Assign weight coefficients to the target interior region, target boundary region, and target neighborhood background region respectively. ,satisfy The weighting coefficients can be preset or adaptively determined based on the target category, image quality, or application scenario. In this step, weighting coefficients are assigned to the errors of the three partitions. These correspond to the internal area, boundary area, and background area, respectively, satisfying... This embodiment has the default settings. The weighting coefficients can be adjusted based on the target category or image quality: for target types with complex boundary contours (such as irregularly shaped facilities), the weighting coefficients should be appropriately increased. To enhance the requirements for boundary accuracy; for images with complex background features and poor homogeneity, the resolution should be appropriately reduced. To reduce the impact of background fluctuations on the overall judgment.

[0077] S4.46. Based on the weight coefficients of each partition, calculate the total partition consistency error corresponding to a single iteration. In this step, the total partition consistency error for a single iteration is calculated by weighting the weight coefficients of each partition: ; It comprehensively reflects the overall accuracy of the correction parameters, correction intensity, and annotation boundaries in this iteration.

[0078] S4.47. When the total partition consistency error is lower than the preset convergence threshold, or the number of iterations reaches the preset upper limit, stop the iteration operation; In this step, a convergence threshold is set. and maximum number of iterations This embodiment takes After each iteration, a decision is made: like This indicates that the consistency between the re-degraded sample and the original coarse sample has met the accuracy requirements, and the iteration stops. However, the number of iterations has reached [a certain number]. To avoid an infinite loop, the iteration stops, and the non-converged state of the sample is recorded. If neither of the above two conditions is met, the process returns to S4.42 to S4.44 to continue adjusting the parameters and boundaries, and enters the next iteration.

[0079] S4.48 Output the final iteratively generated samples as environment adaptation and correction samples, and retain the corresponding environment degradation parameters, local correction intensity and annotation boundary information.

[0080] In this step, after the iteration terminates, the candidate correction samples generated in the last iteration are... This serves as the output for environmental adaptation and correction. Simultaneously, the optimal parameters determined during this processing are retained, including environmental degradation parameters. and Local correction intensity distribution map The corrected target boundary coordinates are also included. This information is archived along with the sample for subsequent quality traceability and sample reliability assessment.

[0081] S4.5. Calculate sample reliability based on partition consistency error, perform sample splitting based on reliability, and standardize the remaining split samples to generate a high-quality standard sample set. Specifically, this includes the following steps: S4.51 Calculate sample reliability based on the final total partition consistency error. The smaller the overall partition consistency error, the higher the sample reliability. The higher, The value range is (0,1]; In this step, the total partition consistency error at the end of iteration S4.46 is used as the basis. Calculate the confidence level of the sample. Credibility is negatively correlated with total error, and the calculation formula is as follows: ; in, As the attenuation adjustment factor, this embodiment takes... . The smaller, The closer it is to 1; The larger, The closer it gets to 0. The value range is (0,1).

[0082] S4.52, will satisfy The samples were marked as qualified and retained; In this step, a high confidence threshold is set. In this embodiment, .like This indicates that after degradation consistency inversion correction, the heavily degraded image is highly consistent with the original coarse image, and the degradation parameter estimation and boundary labeling are relatively accurate. Therefore, this sample is marked as a qualified sample and directly retained.

[0083] S4.53, will satisfy The samples were marked as samples to be reviewed and transferred to the manual review process; In this step, a low confidence threshold is set. In this embodiment, .like This indicates that the calibration process for this sample has some deviations, possibly due to the complexity of the target morphology, uneven background, or special local degradation effects, making it difficult for automatic calibration to converge completely. Such samples are marked as samples awaiting review and, along with their calibration parameters, annotation boundaries, and partitioning error information, are transferred to the manual review process. The operator will then determine whether to accept the sample or require further corrections before acceptance.

[0084] S4.54, will satisfy The samples were removed; In this step, if This indicates that the overall partition consistency error of the sample is too large, the degradation parameter estimation deviates significantly from the true value, the labeled boundary has a large offset, or the background is severely contaminated, and automatic correction can no longer effectively restore its quality. Therefore, this sample is removed and will not proceed to the next step.

[0085] S4.55. Standardize the format, size, coordinate reference and annotation content of the retained qualified samples to generate a high-quality standard sample set.

[0086] In this step, the qualified samples retained in S4.52 are standardized and organized in a unified manner, including: Standardized format: Sample images are uniformly stored in lossless compressed TIFF or PNG format, and annotation files are uniformly stored in TXT format consistent with S2.4, with each line recording the target category number and normalized bounding box coordinates. ; Uniform size: Scaling or padding the sample images to 512×512 pixels to keep the target subject centered and fully visible; Unified coordinate reference: The geographic coordinate information of the samples is uniformly converted to the WGS-84 coordinate system to ensure that the spatial reference of samples from different sources is consistent; Uniform annotation content: Only target category and bounding box information are retained in the annotation file, and auxiliary annotations and metadata generated during intermediate processing are removed to ensure a concise and standardized format.

[0087] After the above standardization process, all qualified samples are standardized into a uniform specification, generating a high-quality standard sample set for subsequent sample library expansion and model iteration training.

[0088] S5. Sample Library Iterative Optimization and Model Update: Select qualified samples from the high-quality standard sample set to expand the initial sample library, and iteratively optimize the target recognition model based on the updated initial sample library to form a closed-loop process linking sample collection and model optimization.

[0089] In this step, all qualified samples generated in the high-quality standard sample set in S4.55 are first archived according to the classification and storage structure established in S2.5. Using target category (e.g., chimneys, storage tanks) and environmental scene (e.g., industrial areas, towns) as classification dimensions, the image files and annotation files of new samples are placed into the corresponding directories. The sample index file is updated synchronously, appending entries such as the file path, target category, scene type, number of bounding boxes, geographic coordinates, and timestamp of entry into the database for new samples. The sample database version number is updated progressively from the current version, for example, from V1.0 to V2.0. The expanded sample database has a larger total size than before, with more comprehensive category and scene coverage.

[0090] Subsequently, based on the expanded sample library, the object recognition model was retrained following the procedures in S3.1 and S3.2. The sample library partitioning remained at a 7:2:1 ratio, and the stratified sampling strategy was consistent with S3.1. The training configuration used the optimizer, learning rate, batch size, and other hyperparameters set in S3.2, with the optimal weights obtained in the previous training serving as the initial weights for fine-tuning. After training, mAP@0.5, precision, and recall were evaluated on the test set and compared with the metrics of the previous version model to verify the performance improvement effect of the expanded sample library.

[0091] Finally, the updated target recognition model is redeployed to stages S3.3 and S3.4 to perform target detection on newly added remote sensing images, collecting a new round of coarse sample sets. After degradation consistency inversion correction and standardization processing in stage S4, the new batch of coarse samples is further screened to supplement the sample library.

[0092] The aforementioned closed-loop process of "new image acquisition → model detection to generate coarse samples → degradation consistency inversion correction → qualified sample entry into the database → model retraining → model update and re-acquisition" can be executed multiple times. With each completion of a closed loop, the sample database gradually expands in size and scene coverage becomes richer, and the model recognition accuracy is continuously optimized in the iteration, forming a self-evolutionary mechanism that links sample acquisition and model optimization.

[0093] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0094] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for intelligent acquisition of specific environmental target samples based on artificial intelligence and remote sensing technology, characterized in that, Includes the following steps: S1. Remote sensing data acquisition and environmental adaptation preprocessing: Acquire multi-source remote sensing images under specific conditions, suppress or correct environmental interference factors in the multi-source remote sensing images, and obtain preprocessed images. S2. Initial construction of a sample library for specific environments: The target information in the preprocessed images is labeled using artificial intelligence annotation methods to build an initial sample library for model training. S3. AI Model Training and Initial Target Sample Collection: Based on the initial sample library, a target recognition model is trained, and the target recognition model is used to perform target detection on newly added remote sensing images to obtain a coarse sample set. S4. Sample Environment Adaptation Correction and Standardization: The coarse sample set is processed using a degradation consistency inversion method. Based on the target annotation information, target partitioning is performed and environmental degradation parameters are estimated. Candidate correction and degradation observation space remapping are completed. Partitioning errors are compared and parameters and annotation boundaries are iteratively corrected. Samples are split based on confidence level and standardized to generate a high-quality standard sample set. The degradation consistency inversion refers to generating candidate correction samples based on sample environmental degradation parameters and mapping the candidate correction samples back to the degradation observation space to verify their consistency with the original coarse sample. The environmental degradation parameters, correction intensity, or annotation boundaries are corrected based on the verification results. S5. Sample Library Iterative Optimization and Model Update: Select qualified samples from the high-quality standard sample set to expand the initial sample library, and iteratively optimize the target recognition model based on the updated initial sample library to form a closed-loop process linking sample collection and model optimization.

2. The intelligent acquisition method for specific environmental target samples based on artificial intelligence and remote sensing technology according to claim 1, characterized in that, In step S1, the remote sensing data acquisition and environmental adaptation preprocessing includes the following steps: S1.1 Acquire multi-source remote sensing images of the target monitoring area, integrate image data from different sensors and different imaging time periods, and construct an initial remote sensing image dataset; S1.2 Identify interference factors in the initial remote sensing image and extract environmental interference features such as clouds, water vapor, aerosols, shadows and illumination deviations contained in the image. S1.

3. For interference from clouds, fog, water vapor, and aerosols, corresponding atmospheric correction algorithms are used for suppression. S1.

4. For interference such as shadows and lighting deviations, corresponding lighting correction algorithms are used for correction. S1.5 Perform consistency verification on the processed image to obtain a preprocessed image with uniform image quality and controllable interference.

3. The intelligent acquisition method for specific environmental target samples based on artificial intelligence and remote sensing technology according to claim 1, characterized in that, In step S2, the initial construction of the specific environment sample library includes the following steps: S2.1 Filter the target region of the preprocessed image, define the target object to be labeled and its neighborhood range under a specific environment, and generate a set of images to be labeled; S2.

2. Use an artificial intelligence-assisted annotation model to perform preliminary annotation of target information in the image set to be annotated, and generate initial annotation data containing target category, location and boundary; S2.3 Perform consistency verification and noise filtering on the initial annotation data, correct annotation deviations, remove invalid annotations, and obtain standard annotation data; S2.

4. Associate and match the standard labeled data with the corresponding preprocessed images, and convert them into a sample format suitable for training the target recognition model; S2.

5. Classify and store the associated sample data according to target category and environmental scenario, establish a sample index and version management mechanism, and build an initial sample library.

4. The intelligent acquisition method for specific environmental target samples based on artificial intelligence and remote sensing technology according to claim 1, characterized in that, In step S3, the AI ​​model training and initial collection of target samples include the following steps: S3.1 Divide the initial sample library into a training set, a validation set and a test set, and allocate samples of each category according to a preset ratio to ensure a balanced sample distribution; S3.

2. Train the preset deep learning target recognition model based on the training set, adjust the model parameters iteratively through the validation set, and evaluate the model recognition accuracy using the test set to obtain the trained target recognition model. S3.3 Acquire newly added remote sensing images under specific environments, perform environmental interference suppression processing on the newly added remote sensing images, and obtain the images to be detected; S3.

4. Use the target recognition model to perform target detection on the image to be detected, and output the detection results including the target location, category and confidence level; S3.

5. Filter valid detection results according to the preset reliability threshold, associate and store the corresponding target area image, annotation information and metadata, and construct a coarse sample set.

5. The intelligent acquisition method for specific environmental target samples based on artificial intelligence and remote sensing technology according to claim 1, characterized in that, In step S4, the sample environment adaptation correction and standardization includes the following steps: S4.

1. Use the degradation consistency inversion method to process the coarse sample set and retrieve the annotation information, regional image information, neighborhood background information and remote sensing image metadata corresponding to the target sample. S4.

2. Based on the target annotation information, perform target partitioning and estimate the environmental degradation parameters corresponding to the target samples; S4.

3. Generate candidate correction samples based on environmental degradation parameters and target partitioning results, and map the candidate correction samples to the degradation observation space to obtain re-degraded samples; S4.

4. Compare the partition consistency error between the degraded sample and the original coarse sample, and iteratively correct the environmental degradation parameters, local correction intensity or annotation boundary based on the partition consistency error; S4.5 Calculate sample credibility based on partition consistency error, complete sample diversion based on credibility, and standardize the diverted and retained samples to generate a high-quality standard sample set.

6. The intelligent acquisition method for specific environmental target samples based on artificial intelligence and remote sensing technology according to claim 5, characterized in that, The target zoning and environmental degradation parameter estimation in S4.2 includes the following steps: S4.

21. Based on the target's labeled boundary coordinates, divide the target sample into three continuous regions: the target interior region, the target boundary region, and the target neighborhood background region. S4.22 Extract the spectral, radiometric and textural features of the target neighborhood background area, and simultaneously acquire image imaging parameters, sensor parameters and atmospheric related metadata; S4.

23. Based on the principle of radiative transfer, using pure pixels, near-pure pixels, or statistically representative pixels in the background area of ​​the target neighborhood, a set of environmental degradation parameters, including atmospheric transmittance and atmospheric comprehensive radiation, is estimated.

7. The intelligent acquisition method for specific environmental target samples based on artificial intelligence and remote sensing technology according to claim 6, characterized in that, The candidate correction and degraded observation space remapping in S4.3 includes the following steps: S4.

31. Based on the target partitioning results, assign a corresponding local correction intensity to each pixel. , The value range is [0,1]; S4.32, Based on environmental degradation parameters and local correction intensity The original coarse sample is subjected to radiometric correction to generate candidate corrected samples; S4.

33. Using the same environmental degradation parameters as in step S4.32, the candidate correction samples are positively mapped back to the original degradation observation space to obtain the re-degraded samples. S4.

34. Unify the radiation value range between the re-degraded sample and the original coarse sample to complete the data benchmark alignment process.

8. The intelligent acquisition method for specific environmental target samples based on artificial intelligence and remote sensing technology according to claim 7, characterized in that, The partition error comparison and parameter and annotation boundary correction in S4.4 include the following steps: S4.41 Calculate the root mean square error of pixels between the re-degraded sample and the original coarse sample in the target interior region, target boundary region and target neighborhood background region respectively, and use it as the partition consistency error of the corresponding region. S4.

42. When the partition consistency error of the target's internal region exceeds the first preset threshold, adjust the environmental degradation parameter value or the local correction intensity. ; S4.

43. When the partition consistency error of the target boundary area exceeds the second preset threshold, the coordinates of the target annotation boundary are finely adjusted along the error gradient direction. S4.44 When the partition consistency error of the target neighborhood background area exceeds the third preset threshold, mark the sample as having background mixing anomaly.

9. The intelligent acquisition method for specific environmental target samples based on artificial intelligence and remote sensing technology according to claim 8, characterized in that, The iteration termination determination in S4.4 includes the following steps: S4.

45. Assign weight coefficients to the target interior region, target boundary region, and target neighborhood background region respectively. ,satisfy The weighting coefficients can be preset or adaptively determined based on the target category, image quality, or application scenario. S4.

46. Based on the weight coefficients of each partition, calculate the total partition consistency error corresponding to a single iteration. S4.

47. When the total partition consistency error is lower than the preset convergence threshold, or the number of iterations reaches the preset upper limit, stop the iteration operation; S4.48 Output the final iteratively generated samples as environment adaptation and correction samples, and retain the corresponding environment degradation parameters, local correction intensity and annotation boundary information.

10. The intelligent acquisition method for specific environmental target samples based on artificial intelligence and remote sensing technology according to claim 9, characterized in that, The sample reliability assessment and standardization output of S4.5 includes the following steps: S4.51 Calculate sample reliability based on the final total partition consistency error. The smaller the overall partition consistency error, the higher the sample reliability. The higher, The value range is (0,1]; S4.52, will satisfy The samples were marked as qualified and retained; S4.53, will satisfy The samples were marked as samples to be reviewed and transferred to the manual review process; S4.54, will satisfy The samples were removed; S4.

55. Standardize the format, size, coordinate reference and annotation content of the retained qualified samples to generate a high-quality standard sample set.

Images