Multi-temporal multi-spectral image feature recognition method based on pixel image fingerprint

Abstract

This invention relates to the field of intelligent remote sensing image recognition technology, and discloses a method for identifying ground features in multi-temporal multispectral images based on pixel image fingerprints. The method includes: calculating several feature index bands for the original multi-temporal multispectral images and integrating them into a structured data file; selecting sample points of various target ground features based on high-resolution reference images and visual interpretation to construct a training sample set; inputting the data into a deep autoencoder network for training, using the trained autoencoder for feature learning, generating a standard pixel image fingerprint database, and setting a high-confidence recognition threshold; traversing the structured data file to be identified, using the trained autoencoder for matching and recognition, and generating target ground feature recognition results; and outputting the target ground feature recognition product through reclassification mapping rules and configuration of visualization colors. This invention effectively solves the problem of identifying "different objects with the same spectrum" caused by the limited number of bands and insufficient spectral separability in multispectral remote sensing images.

Description

Technical Field

[0001] This invention relates to the field of intelligent remote sensing image recognition technology, and more specifically, to a method for recognizing ground features in multi-temporal and multispectral images based on pixel image fingerprints. Background Technology

[0002] Multispectral remote sensing technology acquires the radiometric information of ground features in several to more than ten discrete spectral bands, enabling the simultaneous acquisition of the spatial distribution and spectral characteristics of ground features, forming a remote sensing data source with integrated map and spectrum attributes. Compared with hyperspectral remote sensing, multispectral imagery has advantages such as large swath width, short revisit period, and low data acquisition cost, and has been widely used in operational fields with high requirements for time series observation, such as agricultural planting structure monitoring, land use / cover change detection, disaster emergency assessment, and urban expansion analysis. In particular, with the networking and operation of multispectral satellite constellations such as Sentinel-2, Landsat-8 / 9, and the Gaofen series, the ability to acquire multi-temporal and multispectral remote sensing imagery has been greatly improved, providing a rich data foundation for realizing dynamic identification of ground features and long-term time series change monitoring.

[0003] However, while multi-temporal multispectral remote sensing imagery offers high-frequency observation capabilities, it also introduces a series of technical challenges distinct from single-temporal hyperspectral imagery. First, multi-temporal images are affected by factors such as solar altitude angle, atmospheric conditions, soil moisture, and phenological stages. Even when observing the same ground feature using the same sensor, the spectral response may exhibit significant differences, a phenomenon known as "temporal spectral variation." Second, multispectral images have a limited number of bands and lower spectral resolution, resulting in insufficient spectral separability between ground features. The problem of "different objects sharing the same spectrum" is more pronounced than with hyperspectral images, making it difficult to achieve high-precision ground feature identification solely based on single-temporal spectral information. Furthermore, effective observation windows are scarce in cloudy and rainy areas, and image quality issues such as cloud and snow cover and band noise often exist between images. Combined with geometric registration errors between multi-temporal images, these factors pose a severe challenge to the stable extraction of temporal spectral features and the generalization ability of identification models. In addition, how to balance computational efficiency and automation while ensuring identification accuracy in the face of long-term image sequences accumulated year by year is also a key bottleneck restricting the operational application of multispectral remote sensing.

[0004] Currently, methods for land cover identification based on multispectral remote sensing imagery can be broadly categorized into two types: classification methods based on single-temporal spectral features and identification methods based on multi-temporal time series features. Traditional single-temporal methods, such as maximum likelihood estimation, support vector machines (SVM), and random forests, classify land cover by extracting spectral features and a limited number of texture features from a single period of imagery. These methods are mature and computationally efficient, but they struggle to handle temporal spectral variations and have stringent requirements for the representativeness of training samples. In recent years, time series analysis methods based on multi-temporal remote sensing imagery have received widespread attention. Deep learning methods such as Dynamic Time Warping (DTW), decision trees based on phenological parameter fitting, Long Short-Term Memory (LSTM) networks, and Temporal Convolutional Networks (TempCNN) can effectively characterize the seasonal variation patterns of land cover and have made significant progress in tasks such as crop identification and forest disturbance detection. However, deep time series models generally rely on a large number of high-quality labeled samples for training, resulting in a large number of model parameters, weak interpretability, and limited generalization ability in real-world scenarios such as scarce training samples, incomplete image temporal phases, or cloud contamination.

[0005] Furthermore, existing identification processes largely focus on optimizing single-phase image classifiers or improving temporal feature extraction algorithms, while neglecting a systematic technical solution that is interpretable and has low sample dependence across the entire chain from multi-temporal raw data to the final land cover identification product. In particular, existing methods often limit the use of feature indices with clear physical meaning (such as Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Normalized Difference Building Index (NDBI)) to auxiliary analysis, failing to deeply integrate them with the original spectral bands to form a highly discriminative pixel-level feature representation. Simultaneously, there is a lack of a systematic solution that can transform multi-temporal image data into a refined "fingerprint"-like representation and adaptively construct a standard fingerprint database and confidence thresholds using deep learning.

[0006] Therefore, exploring a multispectral remote sensing image land cover identification method that can deeply integrate original bands and feature indices, construct pixel image fingerprints, automatically generate a standard fingerprint library using deep learning and set a high confidence threshold, and has strong fault tolerance to temporal spectral variations and image noise is of great theoretical and practical significance for promoting the transformation of multispectral remote sensing technology from "single-temporal classification" to "multi-temporal target identification" and improving the automation and intelligence level of operational remote sensing monitoring.

[0007] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0008] The technical problem to be solved by this invention is to overcome the limitations of the limited number of bands and spectral separability of existing multispectral remote sensing images, solve the problem of "different objects with the same spectrum" identification, and propose a multi-temporal multispectral image land cover identification method based on pixel image fingerprint.

[0009] To address the aforementioned technical problems, this invention proposes a multi-temporal, multispectral image-based method for identifying ground features based on pixel image fingerprints.

[0010] A multi-temporal, multispectral image-based land cover identification method based on pixel image fingerprints includes the following steps:

[0011] The original multi-temporal multispectral images are preprocessed to obtain surface reflectance data; several characteristic index bands are calculated based on the surface reflectance data, and the original bands of the surface reflectance data are integrated with all the characteristic index bands to obtain a pixel image; the pixel image is discretized to obtain a spectral coding sequence, and the spectral coding sequence is organized into a structured data file;

[0012] Based on high-resolution reference images and visual interpretation, sample points of various target features are selected, and the corresponding pixel image encoding sequences are extracted from structured data files using the spatial coordinates of the sample points to construct a training sample set.

[0013] The training sample set is input into the deep autoencoder network for training. The trained autoencoder is used to learn features from all training samples to generate deep fingerprint features. The mean of all deep fingerprint features is processed to generate a standard pixel image fingerprint database. The distance distribution of each type of target object between the deep fingerprint features and the standard pixel image fingerprint is analyzed. A high confidence recognition threshold is set based on the distance distribution of each type of target object.

[0014] The algorithm iterates through the image encoding sequence of each pixel in the structured data file to be identified, extracts the deep fingerprint features to be identified using the trained autoencoder, calculates the similarity between the deep fingerprint features to be identified and the standard pixel image fingerprint, matches the similarity with a high confidence recognition threshold, and generates the target ground object identification result.

[0015] The target feature identification results are reconstructed into a two-dimensional classification matrix, and the georeferenced information of the original multi-temporal multispectral images is reused to generate a classification result file;

[0016] By using reclassification mapping rules, the classification result files are reclassified and merged, and visualization colors are configured to output target feature recognition products, thereby enabling the recognition of target features in multi-temporal and multispectral images.

[0017] Furthermore, the original multi-temporal multispectral images are preprocessed to obtain surface reflectance data; several characteristic index bands are calculated based on the surface reflectance data, and the original bands of the surface reflectance data are integrated with all the characteristic index bands to obtain a pixel image; the pixel image is discretized to obtain a spectral coding sequence, and the spectral coding sequence is organized into a structured data file including:

[0018] The metadata file in the original multi-temporal multispectral image is read, the radiance multi-gain, multi-offset, and reflectance scaling coefficients corresponding to each band in the metadata file are parsed, and the radiance multi-gain coefficient is multiplied with the original quantization value of the image. The product result is summed with the multi-offset coefficient to perform radiometric calibration and obtain the spectral radiance of the top atmosphere.

[0019] The planetary reflectance of the top atmosphere is calculated by combining the spectral radiance of the top atmosphere with the Earth-Sun distance, solar zenith angle, and solar irradiance parameters.

[0020] Based on the latitude and longitude and season of the image, the standard atmospheric model and aerosol model are automatically matched. The water vapor content of the atmospheric column is inverted using the water vapor absorption band. Then, the neighborhood effect correction is performed by the atmospheric correction algorithm based on the radiative transfer model, and the planetary reflectance of the top atmosphere is converted into the surface reflectance.

[0021] A pre-defined image is selected as the reference. The same feature points between the image to be registered and the reference image are automatically extracted using the scale-invariant feature transformation algorithm. After removing mismatched points, a multinomial transformation model is constructed. The bicubic convolution interpolation method is used to resample all original multi-temporal multispectral images to obtain the geometrically registered spectral image.

[0022] The geometrically registered spectral images were processed for cloud, cloud shadow and snow masking based on quality assessment bands and spectral characteristics to obtain surface reflectance data.

[0023] Several characteristic indices include: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDRI), and Normalized Difference Building Index (NDBA).

[0024] The normalized vegetation index, normalized water index, and normalized building index bands are calculated based on the surface reflectance data. The original bands of the surface reflectance dataset are then stitched together with all the characteristic index bands in sequence to form a pixel image.

[0025] The characteristic band raster data of all period pixel images are read band by band, converted into a pixel value matrix, and organized into temporary files according to row and column positions.

[0026] The continuous pixel values ​​of each band in the temporary file are discretized using the quantile discretization method. The continuous values ​​are then converted into integer values ​​according to the mapping rules to obtain the spectral coding sequence of each pixel.

[0027] The spectral coding sequence of each pixel is horizontally spliced ​​according to the pixel row and column positions to generate a structured data file.

[0028] Furthermore, cloud, cloud shadow, and snow masking were performed on the geometrically registered spectral image based on quality assessment bands and spectral characteristics to obtain surface reflectance data, including:

[0029] The quality assessment bands inherent in the geometrically registered spectral image are decoded, and key marker bits are extracted through bitwise operations. The key marker bits include expanding clouds, cirrus clouds, high-confidence clouds, cloud shadows, and snow or ice, generating the quality assessment band decoding results.

[0030] Normalized differential snow index is calculated for the geometrically registered spectral image, and snow is distinguished from bright clouds by combining the low reflectance characteristics of the near-infrared band. Low-temperature cloud tops are identified by using the thermal infrared band temperature threshold, and morphological dilation of cloud pixels in the geometrically registered spectral image is performed based on the solar azimuth angle to obtain spectral enhancement discrimination results.

[0031] The quality assessment band decoding results and the spectral enhancement discrimination results are logically ORed to generate a composite mask image.

[0032] After traversing the spectral images after multiple geometric registrations, the pixel positions marked as clouds, cloud shadows, or snow in the composite mask image are assigned invalid values, and the output is surface reflectance data containing only valid surface information.

[0033] Furthermore, the formula for calculating the Normalized Difference Vegetation Index is:

[0034] NDVI = (NIR - Red) / (NIR + Red);

[0035] The formula for calculating the normalized water index is:

[0036] NDWI=(Green-NIR) / (Green+NIR);

[0037] The formula for calculating the Normalized Building Index is:

[0038] NDBI=(SWIR1-NIR) / (SWIR1+NIR);

[0039] In the formula, NDVI represents the normalized vegetation index; NDWI represents the normalized water index; NDBI represents the normalized building index; NIR represents the near-infrared reflectance; Red represents the red band; Green represents the green band; and SWIR1 represents the shortwave infrared band.

[0040] Furthermore, based on high-resolution reference imagery and visual interpretation, sample points for various target features are selected. The corresponding pixel image encoding sequences are extracted from structured data files using the spatial coordinates of these sample points, and a training sample set is constructed, including:

[0041] By visually interpreting and importing vector sample points and high-resolution reference images in batches, the vector sample points and high-resolution reference images are automatically aligned in coordinate system and space, and the pixel row and column number of each sample point on the original multi-temporal multispectral image is extracted.

[0042] Based on the pixel row and column number of each sample point, the complete pixel image encoding sequence of each sample point is located and extracted from the digitized structured data file;

[0043] The extracted pixel image encoding sequences are automatically organized according to land cover categories and output as structured sample files to form a training sample set.

[0044] Furthermore, the vector sample points and high-resolution reference images are imported in batches. The vector sample points and high-resolution reference images are automatically aligned in coordinate system and space, and the pixel row and column numbers of each sample point on the original multi-temporal multispectral image are extracted, including:

[0045] Based on the land feature identification standards and referring to high-resolution reference images, sample points of various typical land features are selected through visual interpretation.

[0046] Record and store the spatial location and category label of the sample points as vector point sample points with geographic coordinates;

[0047] The vector point sample points are uniformly transformed to a spatial reference system consistent with the original multi-temporal multispectral image, and the pixel row and column number corresponding to each sample point is calculated.

[0048] Furthermore, the training sample set is input into a deep autoencoder network for training. The trained autoencoder is then used to learn features from all training samples to generate deep fingerprint features. The mean of all deep fingerprint features is applied to generate a standard pixel image fingerprint database, and the distance distribution of each category of target features between the deep fingerprint features and the standard pixel image fingerprints is analyzed. Based on the distance distribution of each type of target feature, high-confidence recognition thresholds are set, including:

[0049] A deep autoencoder network is constructed, with the pixel image encoding sequence of the training sample set as input. Low-dimensional depth features are extracted through the deep learning autoencoder, and the input is reconstructed through the deep learning autoencoder to train the deep autoencoder network and minimize the reconstruction error.

[0050] After training is completed, the trained autoencoder is used as the feature extractor. All training samples in the training sample set are input into the feature extractor for feature extraction to obtain the deep fingerprint features of each training sample.

[0051] The mean value of the depth fingerprint features of all samples of each type of target land cover is calculated, and the mean value is used as the standard pixel image fingerprint of each type of target land cover to generate a standard pixel image fingerprint library for each type of target land cover.

[0052] The distance between the depth fingerprint features of each training sample and the standard pixel image fingerprint of the corresponding class of target land cover is calculated to obtain the intra-class distance distribution. The distance between the training sample and the standard pixel image fingerprint of the target land cover of the opposite class is calculated to obtain the inter-class distance distribution. The high confidence recognition threshold of the class land cover is automatically set based on the statistics of the intra-class distance distribution.

[0053] Furthermore, the image encoding sequence of each pixel in the structured data file to be identified is traversed, and the trained autoencoder is used to extract the depth fingerprint features to be identified. The similarity between the depth fingerprint features to be identified and the standard pixel image fingerprint is calculated. The similarity is matched with a high-confidence recognition threshold to generate the target land cover identification results, including:

[0054] Read the standard pixel image fingerprint features of one or more target land features specified by the user from the standard pixel image fingerprint database;

[0055] Traverse the structured data file to be identified, input the spectral encoding sequence of each pixel image into the trained autoencoder to obtain the deep fingerprint features to be identified;

[0056] Calculate the Hamming distance between the depth fingerprint feature to be identified and the standard pixel image fingerprint features of various target features, and use the Hamming distance as the similarity between the depth fingerprint feature to be identified and the standard pixel image fingerprint.

[0057] If the minimum similarity is less than or equal to the high confidence threshold of the corresponding target feature category, the pixel image is determined to be the corresponding target feature category and assigned a category label; otherwise, it is marked as unrecognized, and the target feature recognition result is obtained.

[0058] Furthermore, the target feature identification results are reconstructed into a two-dimensional classification matrix, and the georeferenced information of the original multi-temporal multispectral imagery is reused to generate classification result files, including:

[0059] Unidentified cells in the target feature identification results are uniformly assigned the value of zero to generate a pure two-dimensional classification matrix;

[0060] The classification result files include: the standard ENVI classification file and its header file;

[0061] By combining the two-dimensional classification matrix with the geographic transformation parameters and coordinate system information extracted from the original multi-temporal multispectral imagery, a standard ENVI classification file with geographic reference and its header file are generated.

[0062] Furthermore, by using reclassification mapping rules, the classification result files are reclassified and merged, and visualization colors are configured to output target feature recognition products, thereby achieving the recognition of target features in multi-temporal and multispectral imagery, including:

[0063] Based on user-defined reclassification mapping rules, the category labels in the standard ENVI classification files and their header files are uniformly merged into a single category identifier, and the identifier values ​​of unrecognized areas are standardized to generate standardized classification data.

[0064] Configure a specified display color for each category in the normalized classification data, and write the display color along with georeferenced information into a new file to generate a target feature identification product for thematic mapping;

[0065] By comparing and analyzing the target feature recognition product with the verification sample and calculating the overall accuracy index, the recognition effect is quantitatively evaluated based on the overall accuracy index.

[0066] The beneficial effects of this invention are as follows:

[0067] 1. This invention effectively solves the problem of identifying "different objects with the same spectrum" in multispectral remote sensing images due to the limited number of bands and insufficient spectral separability, by constructing pixel image fingerprints and combining them with deep learning autoencoder feature learning technology. A quantile discretization method is used to deeply fuse multi-temporal, multi-band continuous spectral values ​​with feature indices (NDVI, NDWI, NDBI) to form a highly discriminative pixel image coding sequence, significantly compressing data dimensionality while preserving ground object spectral discrimination information. A deep learning autoencoder is used to automatically learn the deep fingerprint features of pixel images, generating a standard fingerprint library with greater robustness and generalization ability. This method provides a highly interpretable and low-sample-dependency discrimination basis for multispectral remote sensing target ground object identification, significantly reducing the reliance on traditional manual feature design and a large number of labeled samples.

[0068] 2. This invention proposes an automatic threshold setting mechanism based on deep learning, addressing the common problem of "temporal spectral variation" in multi-temporal remote sensing images and achieving a precise balance between high confidence and high coverage. By analyzing the intra-class and inter-class distance distribution of training samples, the high-confidence recognition threshold for various land features is adaptively determined, avoiding the subjectivity of manual parameter tuning. In the recognition stage, only the distance between the depth fingerprint features of the pixel to be identified and the standard fingerprint needs to be measured, and the judgment is made according to the automatic threshold, thus achieving scientific coverage from high-confidence core pixels to edge variation pixels. This mechanism significantly improves the completeness and robustness of target land feature recognition while ensuring recognition accuracy, providing an objective and stable adaptive matching strategy for multispectral remote sensing interpretation in scenarios with complex temporal combinations and uneven image quality.

[0069] 3. This invention achieves seamless conversion from pixel image encoding to geospatial recognition products, bridging the key links between algorithm processing and business applications. Through row and column position reconstruction and reuse of original image georeferenced information, pixel-level recognition results are rapidly converted into standard ENVI classification files and their header files, fully preserving spatial reference and resolution information. The reclassification module supports user-defined label merging, unidentified area labeling, and visualization color configuration, directly generating target feature recognition products that conform to thematic mapping specifications. This output method greatly improves the engineering efficiency and cross-platform usability of multispectral remote sensing recognition results.

[0070] 4. This invention constructs a fully automated identification system encompassing image preprocessing, index calculation, pixel image generation, deep learning fingerprint database construction, automatic threshold matching, and output, forming a lightweight, low-sample-dependency intelligent interpretation solution for multi-temporal and multispectral remote sensing images. This method does not rely on massive amounts of labeled samples and complex deep model training; users only need to provide a small number of typical sample points, and the system can automatically complete the construction of a standard fingerprint database and threshold setting, and select the target land cover type as needed. It is particularly suitable for operational monitoring scenarios with scarce training samples, flexible temporal combinations, and high response speed requirements, such as large-scale crop area estimation, land use dynamic monitoring, and disaster emergency assessment. The proposal of this invention provides efficient, reliable, and easily deployable technical support for promoting the transformation and upgrading of multispectral remote sensing technology from "single-temporal classification" to "multi-temporal target recognition." Attached Figure Description

[0071] The invention will now be further described with reference to the accompanying drawings.

[0072] Figure 1 This is a flowchart illustrating the overall process of the multi-temporal, multispectral image-based ground cover identification method based on pixel image fingerprints according to an embodiment of the present invention.

[0073] Figure 2 This is a flowchart illustrating the multi-temporal, multispectral image-based ground cover identification method based on pixel image fingerprints according to an embodiment of the present invention.

[0074] Figure 3 This is a schematic diagram illustrating the application of the fingerprint recognition principle to image-based ground feature recognition in an embodiment of the present invention.

[0075] Figure 4 This is a schematic diagram illustrating how the present invention quantizes a specified number of integer categories using the quantile discretization method;

[0076] Figure 5 This is a schematic diagram illustrating image preprocessing and pixel-level digitization conversion according to the present invention.

[0077] Figure 6A schematic diagram illustrating the construction of a fingerprint template library and the provision of standard feature references according to an embodiment of the present invention;

[0078] Figure 7 A schematic diagram illustrating pixel identification using multi-level progressive Hamming distance matching according to the present invention;

[0079] Figure 8 A schematic diagram illustrating the reconstruction of classification result labels into geospatial products according to the present invention;

[0080] Figure 9 This is a schematic diagram illustrating the standardization and visualization optimization of the reclassification process implemented according to the present invention.

[0081] Figure 10 Sample image fingerprints of two land cover types, water bodies and cultivated land, prepared according to the present invention;

[0082] Figure 11 This is a map showing the results of water feature identification according to the present invention. Detailed Implementation

[0083] The present invention will now be described in detail with reference to the accompanying drawings, which will make the technical approach and operation steps of the present invention clearer.

[0084] According to embodiments of the present invention, a method for identifying ground features in multi-temporal multispectral images based on pixel image fingerprints is provided.

[0085] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figures 1-11 As shown, according to an embodiment of the present invention, a method for identifying ground features based on pixel image fingerprints in multi-temporal and multispectral images is provided. The method includes the following steps:

[0086] The original multi-temporal multispectral images are preprocessed to obtain surface reflectance data; several characteristic index bands are calculated based on the surface reflectance data, and the original bands of the surface reflectance data are integrated with all the characteristic index bands to obtain a pixel image; the pixel image is discretized to obtain a spectral coding sequence, and the spectral coding sequence is organized into a structured data file.

[0087] Specifically, the original multi-temporal multispectral images are preprocessed to obtain surface reflectance data; several characteristic index bands are calculated based on the surface reflectance data, and the original bands of the surface reflectance data are integrated with all the characteristic index bands to obtain a pixel image; the pixel image is discretized to obtain a spectral coding sequence, and the spectral coding sequence is organized into a structured data file including:

[0088] The metadata file in the original multi-temporal multispectral image is read, the radiance multi-gain, multi-offset, and reflectance scaling coefficients corresponding to each band in the metadata file are parsed, and the radiance multi-gain coefficient is multiplied with the original quantization value of the image. The product result is summed with the multi-offset coefficient to perform radiometric calibration and obtain the spectral radiance of the top atmosphere.

[0089] The planetary reflectance of the top atmosphere is calculated by combining the spectral radiance of the top atmosphere with the Earth-Sun distance, solar zenith angle, and solar irradiance parameters.

[0090] Based on the latitude and longitude and season of the image, the standard atmospheric model and aerosol model are automatically matched. The water vapor content of the atmospheric column is inverted using the water vapor absorption band. Then, the neighborhood effect correction is performed by the atmospheric correction algorithm based on the radiative transfer model, and the planetary reflectance of the top atmosphere is converted into the surface reflectance.

[0091] A pre-defined image is selected as the reference. The same feature points between the image to be registered and the reference image are automatically extracted using the scale-invariant feature transformation algorithm. After removing mismatched points, a multinomial transformation model is constructed. The bicubic convolution interpolation method is used to resample all original multi-temporal multispectral images to obtain the geometrically registered spectral image.

[0092] The geometrically registered spectral images were processed for cloud, cloud shadow, and snow masking based on quality assessment bands and spectral characteristics to obtain surface reflectance data.

[0093] Specifically, cloud, cloud shadow, and snow masking processing based on quality assessment bands and spectral characteristics is performed on the geometrically registered spectral image to obtain surface reflectance data including:

[0094] The quality assessment bands inherent in the geometrically registered spectral image are decoded, and key marker bits are extracted through bitwise operations. The key marker bits include expanding clouds, cirrus clouds, high-confidence clouds, cloud shadows, and snow or ice, generating the quality assessment band decoding results.

[0095] Normalized differential snow index is calculated for the geometrically registered spectral image, and snow is distinguished from bright clouds by combining the low reflectance characteristics of the near-infrared band. Low-temperature cloud tops are identified by using the thermal infrared band temperature threshold, and morphological dilation of cloud pixels in the geometrically registered spectral image is performed based on the solar azimuth angle to obtain spectral enhancement discrimination results.

[0096] The quality assessment band decoding results and the spectral enhancement discrimination results are logically ORed to generate a composite mask image.

[0097] After traversing the spectral images after multiple geometric registrations, the pixel positions marked as clouds, cloud shadows, or snow in the composite mask image are assigned invalid values, and the output is surface reflectance data containing only valid surface information.

[0098] Several characteristic indices include: Normalized Difference Vegetation Index, Normalized Difference Water Index, and Normalized Difference Building Index.

[0099] Specifically, the formula for calculating the Normalized Difference Vegetation Index (NDVI) is:

[0100] NDVI = (NIR - Red) / (NIR + Red);

[0101] The formula for calculating the normalized water index is:

[0102] NDWI=(Green-NIR) / (Green+NIR);

[0103] The formula for calculating the Normalized Building Index is:

[0104] NDBI=(SWIR1-NIR) / (SWIR1+NIR);

[0105] In the formula, NDVI represents the normalized vegetation index; NDWI represents the normalized water index; NDBI represents the normalized building index; NIR represents the near-infrared reflectance; Red represents the red band; Green represents the green band; and SWIR1 represents the shortwave infrared band.

[0106] The normalized vegetation index, normalized water index, and normalized building index bands are calculated based on the surface reflectance data. The original bands of the surface reflectance dataset are then stitched together with all the characteristic index bands in sequence to form a pixel image.

[0107] The characteristic band raster data of all period pixel images are read band by band, converted into a pixel value matrix, and organized into temporary files according to row and column positions.

[0108] The continuous pixel values ​​of each band in the temporary file are discretized using the quantile discretization method. The continuous values ​​are then converted into integer values ​​according to the mapping rules to obtain the spectral coding sequence of each pixel.

[0109] The spectral coding sequence of each pixel is horizontally spliced ​​according to the pixel row and column positions to generate a structured data file.

[0110] Specifically, such as Figure 2 and Figure 3 As shown in this embodiment, the specific steps for preprocessing downloaded multi-phase Landsat 8 / 9 images (i.e., original multi-temporal multispectral images) to obtain high-quality surface reflectance data are as follows:

[0111] First, radiometric calibration, atmospheric correction, and geometric registration are performed. Specifically, the system reads the metadata file (_MTL.txt) from the image data packet and parses the radiance multi-gain (M) corresponding to each band. L ), multiple offsets (A) L ) and reflectivity scaling factor; using formula L λ =M L ×Q cal +A L The original quantized value (DN value, denoted as Q) is used to quantize the original value (DN value, denoted as Q). cal ) converted to top atmospheric spectral radiance (L λ Furthermore, by combining the Earth-Sun distance, solar zenith angle, and solar irradiance parameters, the planetary reflectivity (ρ) of the upper atmosphere is calculated. TOA Subsequently, atmospheric correction algorithms based on radiative transfer models (such as FLAASH or LEDAPS) are employed. These algorithms automatically match standard atmospheric models (such as mid-latitude summer or winter models) and aerosol models based on the latitude, longitude, and season at the time of image formation. The atmospheric column water vapor content is then retrieved using the water vapor absorption band, and neighborhood effect correction is performed to adjust the ρ... TOA Precisely converted to surface reflectance (ρ) Surface In the geometric registration stage, an image with low cloud cover and distinct features is selected as the reference. The Scale Invariant Feature Transform (SIFT) algorithm is used to automatically extract the corresponding feature points between the image to be registered and the reference image. After removing mismatched points, a multinomial transformation model is constructed. Finally, bicubic convolution interpolation is used to resample all multi-phase images to ensure sub-pixel alignment of multi-temporal data in space, and the registration error is controlled within 0.5 pixels.

[0112] Secondly, cloud, cloud shadow, and snow masking based on Quality Assessment (QA) bands and spectral features is performed to remove contaminated pixels. Specific steps include: First, decoding the QA_PIXEL bands inherent in the image and extracting key marker bits through bitwise operations, including expanding clouds (3rd bit), cirrus clouds (4th bit), high-confidence clouds (5th bit), cloud shadows (7th bit), and snow / ice (9th bit), generating a preliminary binary mask; Second, to overcome the limitations of a single QA band in complex scenes, spectral features are introduced for secondary discrimination: the Normalized Differential Snow Index (NDSI) is calculated as (Green - SWIR1) / (Green + SWIR1), and combined with the near-infrared band... The process involves four steps: first, distinguishing snow from bright clouds using low reflectance features; second, identifying low-temperature cloud tops using thermal infrared temperature thresholds; and third, performing morphological dilation on cloud pixels based on solar azimuth to cover missed shadow areas. The QA band decoding results are then logically ORed with the spectral enhancement discrimination results to generate the final composite mask image. Finally, multiple periods of surface reflectance images are traversed, and pixels marked as clouds, cloud shadows, or snow in the composite mask image are assigned invalid values ​​(NoData), resulting in a pure reflectance dataset containing only valid surface information for subsequent analysis.

[0113] For each image period, the NDVI, NDWI, and NDBI indices are calculated pixel by pixel. The calculation formulas are as follows: (1) NDVI = (NIR - Red) / (NIR + Red); (2) NDWI = (Green - NIR) / (Green + NIR); (3) NDBI = (SWIR1 - NIR) / (SWIR1 + NIR); where NIR is the reflectance of the near-infrared band, Red is the red band, Green is the green band, and SWIR1 is the shortwave infrared band 1. The original bands of each image period are compared with the calculated indices. The bands are sequentially stitched together to form a multidimensional feature vector for each pixel, which is the pixel image. The feature band raster data of all periods are read band by band, converted into a pixel value matrix, and organized into a temporary file according to the row and column positions. The continuous pixel values ​​of each band are discretized using the quantile discretization method, and the continuous values ​​are converted into integer values ​​of level 1-5 according to the mapping rules to form the spectral coding sequence of each pixel. The spectral coding sequence of each pixel is horizontally stitched together according to the pixel row and column positions to generate a structured data file, which provides standardized input for subsequent pixel image fingerprint extraction and matching.

[0114] Specifically, invalid bands are removed and noise is suppressed from the original multi-temporal multispectral images, while retaining core band images with high information validity. The core band images are then converted pixel-by-pixel into spectral coding sequences composed of integer values, and organized into a structured tabular dataset. This includes: supporting user-defined removal of noisy or redundant invalid band ranges from the original multispectral images, retaining core band images with high information content and strong discriminative power; reading raster data from the retained core band images band by band, converting them into a pixel value matrix, and organizing it into a structured tabular file according to row and column positions; discretizing the continuous pixel values ​​in the structured tabular file of each band using the quantile discretization method to highlight spectral differences between ground features, and forming a spectral feature coding sequence for each pixel according to mapping rules (e.g., level 1-5); and horizontally concatenating the spectral feature coding sequences of each pixel according to pixel row and column positions to generate a structured tabular dataset, providing a standardized input format for subsequent spectral fingerprint extraction and matching.

[0115] Specifically, such as Figure 6 As shown, this corresponds to the fingerprint image acquisition stage in fingerprint recognition. The core task is to transform the raw multi-temporal multispectral remote sensing images into a standard data format that can be used for matching analysis. Through preprocessing of the multi-phase images (including cloud and snow masking, radiometric normalization, and invalid band removal), the information-rich core spectral bands are retained, and the multi-temporal spectral response of each pixel is quantized into discrete integer values. This digitization process essentially generates unique "image fingerprint" raw data for each spatial pixel based on the time dimension, laying the foundation for subsequent template construction and similarity matching.

[0116] This step aims to transform the original image data into a structured pixel image fingerprint file by integrating, preprocessing, calculating feature indices, and discretizing the original multi-temporal multispectral images, thus laying the foundation for subsequent fingerprint construction and matching.

[0117] I. Data Acquisition and Preprocessing:

[0118] The data source is the Landsat 8 / 9 Level-2 surface reflectance products for the study area, downloaded from the Geospatial Data Cloud (http: / / www.gscloud.cn / ). The imagery covers the period from January 2021 to December 2023, with a spatial resolution of 30m. Each imagery contains 7 effective bands (coastal / aerosol band, blue band, green band, red band, near-infrared band, shortwave infrared 1, and shortwave infrared 2).

[0119] Quality screening: Cloud, cloud shadow, and snow masking are processed using the QA band (quality assessment band) built into the image to remove contaminated pixels; geometric registration and radiometric normalization are performed on multiple images to ensure spectral comparability and spatial consistency between different time phases.

[0120] II. Calculation of Characteristic Index:

[0121] To enhance the ability to identify ground features, three types of feature indices with clear physical meaning are calculated pixel by pixel for each image: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Normalized Difference Building Index (NDBI).

[0122] The calculated three exponential bands are sequentially stitched together with the original seven bands to form ten feature bands for each image period. For 15 image periods, each pixel accumulates 150 feature values, which constitute the original multidimensional feature vector of that pixel, i.e., the "pixel image".

[0123] III. Image Data Conversion and Discretization:

[0124] The program to run is the "01 merge.py" program, which is developed based on Python's rasterio and pandas libraries to perform numerical conversion and discretization processing on multi-period image data.

[0125] Conversion process:

[0126] 1. The program reads 10 characteristic band files (TIFF format) of 15 images in sequence, reads raster data band by band and converts it into a two-dimensional numerical matrix.

[0127] 2. To improve subsequent matching efficiency and enhance feature separability, the continuous pixel values ​​of each band are discretized. A quantile discretization method is used, iterating through the numerical distribution of each band and calculating its 20%, 40%, 60%, and 80% quantiles as discretization thresholds, mapping continuous values ​​to integer values ​​of levels 1-5. The thresholds are dynamically determined based on the data distribution. Figure 4 As shown in Table 1, examples of mapping rules are provided.

[0128] Table 1 Mapping Rules Table

[0129]

[0130] Core objective: To achieve data compression through quantile discretization, while highlighting spectral differences between ground features, providing a clearer feature representation for subsequent matching.

[0131] 3. Each band is discretized to generate an independent temporary file, which records the integer pixel value information of the corresponding period-band combination to ensure the integrity and traceability of time-series spectral data.

[0132] IV. Multi-temporal feature sequence fusion and pixel image fingerprint file generation:

[0133] The fusion logic is based on the row and column coordinates of a pixel, and concatenates the discrete values ​​(1-5) of the same pixel across 150 features in temporal order (first period, second period, then band) to form an integer sequence of length 150. Example: If the discrete values ​​of a pixel in each band from period 1 to period 15 are [1,3,2,…,4], then its final pixel image fingerprint sequence is “132…4”.

[0134] The dataset is generated by organizing the feature sequences of all pixels according to their row and column positions, resulting in a structured pixel image fingerprint file (named "image_fingerprint_data.txt"). Each row in the file corresponds to a row of the image, each column corresponds to a column of the image, and each cell stores a 150-dimensional integer sequence for that pixel, achieving spatial, spectral, index, and temporal alignment of multi-temporal remote sensing data.

[0135] Specifically, such as Figure 6 As shown, this step corresponds to the fingerprint image acquisition stage in fingerprint recognition. The core task is to transform the raw multi-temporal, multispectral remote sensing images into a standard data format that can be used for matching analysis. By preprocessing, calculating feature indices, and discretizing the images from multiple periods, we are essentially generating unique "pixel image fingerprint" raw data for each spatial pixel based on the time dimension and physical indices, laying the foundation for subsequent template construction and similarity matching.

[0136] Based on high-resolution reference images and visual interpretation, sample points of various target features are selected, and the corresponding pixel image encoding sequences are extracted from the structured data files using the spatial coordinates of the sample points to construct a training sample set.

[0137] Specifically, based on high-resolution reference imagery and visual interpretation, sample points for various target features are selected. The spatial coordinates of these sample points are used to extract corresponding pixel image encoding sequences from structured data files, constructing a training sample set including:

[0138] By visually interpreting and importing vector sample points and high-resolution reference images in batches, the vector sample points and high-resolution reference images are automatically aligned in coordinate system and space, and the pixel row and column number of each sample point on the original multi-temporal multispectral image is extracted.

[0139] Specifically, the process involves batch importing vector sample points and high-resolution reference images, automatically aligning the vector sample points with the high-resolution reference images in terms of coordinate system and space, and extracting the pixel row and column numbers of each sample point on the original multi-temporal multispectral image, including:

[0140] Based on the land feature identification standards and referring to high-resolution reference images, sample points of various typical land features are selected through visual interpretation.

[0141] Record and store the spatial location and category label of the sample points as vector point sample points with geographic coordinates;

[0142] The vector point sample points are uniformly transformed to a spatial reference system consistent with the original multi-temporal multispectral image, and the pixel row and column number corresponding to each sample point is calculated.

[0143] Based on the pixel row and column number of each sample point, the complete pixel image encoding sequence of each sample point is located and extracted from the digitized structured data file;

[0144] The extracted pixel image encoding sequences are automatically organized according to land cover categories and output as structured sample files to form a training sample set.

[0145] Specifically, it supports batch import of vector sample points (SHP format) and high-resolution reference images (TIFF format), automatically aligns the vector sample points with the high-resolution reference images in coordinate system and space, and extracts the pixel row and column numbers of each sample point on the original multispectral image. Using the pixel row and column numbers of the sample points, it locates and extracts the complete spectral feature vector of each sample point from the digitized tabular dataset. It automatically organizes the spectral feature vectors of all sample points according to the land cover category and outputs them as a structured tabular file, forming a standardized spectral fingerprint template library.

[0146] Specifically, such as Figure 7 As shown, this step corresponds to feature point marking and registration in fingerprint recognition. In the preprocessed image, training sample points of various typical land features are selected based on the principles of land feature purity and representativeness. These sample points are like registered fingerprints in a fingerprint database, requiring assurance of the purity and typicality of their spectral features. Through the auxiliary interpretation of high-resolution reference images, the precise spatial location of the sample points is obtained, and then their corresponding multi-temporal spectral feature vectors are extracted from the digital data, forming the raw materials for constructing a "standard spectral fingerprint template."

[0147] like Figure 11 The image shown illustrates a pixel-based image fingerprint of a water feature sample. This fingerprint is a 15-row, 10-column matrix structure, where the rows represent image data from 15 different time periods, and the columns represent the combination of 7 original bands and 3 feature indices (NDVI, NDWI, NDBI) for each image period. Each cell contains a discretized integer value ranging from 1 to 5. This example visually demonstrates the spatiotemporal, spectral, and exponential multidimensional feature representation of a pixel-based image fingerprint.

[0148] This stage corresponds to the feature template storage in fingerprint recognition. The feature information of various land cover samples obtained in the previous step is systematically organized and stored to construct a standardized "spectral fingerprint template library." By aggregating and analyzing the features of multiple samples of the same type of land cover (e.g., calculating the mean vector), standard templates that can represent the spectral characteristics of this type of land cover are generated. The quality of the template library directly determines the accuracy of subsequent identification; therefore, it is necessary to ensure that the templates have good representativeness and discriminative power. This step, through sample selection, feature extraction, and template generation, constructs a highly representative "spectral fingerprint" template library for land covers, providing a core reference for subsequent matching and identification of target land covers specified by users.

[0149] I. Sample point collection and coordinate extraction:

[0150] The land cover identification standard is based on the first-level classification standard of the "Classification of Current Land Use" (GB / T 21010-2017). Combined with the actual land cover status of the study area, land cover is divided into four core categories: cultivated land (including paddy fields and dry land), forest land (including arbor forests and shrub forests), water bodies (including rivers, lakes and reservoirs), and artificial surfaces (including urban built-up areas, rural settlements and transportation land).

[0151] The sample selection principles include: purity principle, which is based on visual interpretation of high-resolution Google Earth imagery to ensure that the sample area corresponds to a single homogeneous land cover unit, avoiding mixed pixels, land cover boundaries, shadows, and cloud-contaminated areas; representativeness and balance principle, which adopts a stratified random sampling strategy, ensuring that the samples are spatially evenly distributed within the study area, reflecting the spectral response differences of different growth stages of land cover and under different environmental conditions, and that the number of samples in each category is on the same order of magnitude (no less than 40 per category); independence and sufficiency principle, which ensures that the training samples and accuracy verification samples are completely separated in space, and that the total number of samples meets the requirements for model learning and reliability testing; and coordinate saving, which saves the selected sample points as vector point files (SHP format), recording the geographic coordinates (latitude and longitude) and category label of each sample point.

[0152] II. Extraction of the image encoding sequence of sample points and pixels:

[0153] The program to run is the "02_sample extraction.py" program, and the core process is as follows:

[0154] 1. Coordinate Projection Transformation: The program reads the sample point SHP file and the first-phase original multispectral image (used to obtain geographic transformation parameters), projects the geographic coordinates of the sample points to the image coordinate system, and calculates the pixel row and column number (pixel_row, pixel_col) corresponding to each sample point.

[0155] 2. Feature Sequence Extraction: Read the "image_fingerprint_data.txt" file generated by image preprocessing. This file stores a 150-dimensional integer sequence for each pixel (corresponding to 15 image periods, with 10 features per period). Based on the row and column numbers of the sample points, locate and extract the corresponding 150-dimensional integer sequence as the "original pixel image fingerprint" for that sample point.

[0156] 3. Sample data storage: Generate an independent sample file (e.g., "water_sample.txt") for each type of land cover. The file contains the spatial coordinates of the sample points, row and column numbers, and a 150-dimensional feature sequence to ensure the traceability of the sample data.

[0157] The training sample set is input into the deep autoencoder network for training. The trained autoencoder is used to learn features from all training samples to generate deep fingerprint features. The mean of all deep fingerprint features is processed to generate a standard pixel image fingerprint library. The distance distribution of each type of target object between the deep fingerprint features and the standard pixel image fingerprint is analyzed. A high confidence recognition threshold is set based on the distance distribution of each type of target object.

[0158] Specifically, the training sample set is input into a deep autoencoder network for training. The trained autoencoder then learns features from all training samples to generate deep fingerprint features. The mean of all deep fingerprint features is applied to generate a standard pixel image fingerprint database. The distance distributions of different target object categories between the deep fingerprint features and the standard pixel image fingerprints are analyzed. High-confidence recognition thresholds are set based on the distance distributions of various target objects, including:

[0159] A deep autoencoder network is constructed, with the pixel image encoding sequence of the training sample set as input. Low-dimensional depth features are extracted through the deep learning autoencoder, and the input is reconstructed through the deep learning autoencoder to train the deep autoencoder network and minimize the reconstruction error.

[0160] After training is completed, the trained autoencoder is used as the feature extractor. All training samples in the training sample set are input into the feature extractor for feature extraction to obtain the deep fingerprint features of each training sample.

[0161] The mean value of the depth fingerprint features of all samples of each type of target land cover is calculated, and the mean value is used as the standard pixel image fingerprint of each type of target land cover to generate a standard pixel image fingerprint library for each type of target land cover.

[0162] The distance between the depth fingerprint features of each training sample and the standard pixel image fingerprint of the corresponding class of target land cover is calculated to obtain the intra-class distance distribution. The distance between the training sample and the standard pixel image fingerprint of the target land cover of the opposite class is calculated to obtain the inter-class distance distribution. The high confidence recognition threshold of the class land cover is automatically set based on the statistics of the intra-class distance distribution.

[0163] Specifically, the high-confidence recognition threshold for a class is automatically set based on the statistics of the intra-class distance distribution to ensure that the confidence of the recognition results within the threshold meets the preset requirements. Specifically, the 95th percentile of the intra-class distance distribution can be used as the high-confidence recognition threshold for the class, or the threshold coefficient can be adjusted according to the accuracy of the validation set.

[0164] Specifically, in this embodiment, the construction and generation of the standard pixel image fingerprint database involves constructing and training a deep autoencoder network to compress high-dimensional discretized spectral sequences into low-dimensional depth features, thereby statistically generating standard fingerprints and adaptive recognition thresholds for various land features. The specific implementation steps are as follows:

[0165] 1. Construction and training of deep autoencoder networks:

[0166] (1) Network architecture design:

[0167] A symmetrical Deep Autoencoder (DAE) network is constructed, consisting of an encoder and a decoder. The aim is to map the input 150-dimensional discretized spectral encoded sequence (corresponding to 15 image periods × 10 feature bands) to a low-dimensional latent space and reconstruct the original input as losslessly as possible. Input layer: 150 nodes receive the 150-dimensional integer spectral encoded sequence for each pixel. Encoder part (feature compression): A multi-layer fully connected neural network structure is used for layer-by-layer dimensionality reduction. Specifically, the first hidden layer contains 64 neurons with Rectified Linear Unit (ReLU) activation; the second hidden layer contains 32 neurons with ReLU activation; and the bottleneck layer contains 16 neurons with Linear Activation. The 16-dimensional vector output by the bottleneck layer is the final generated "deep fingerprint feature," achieving efficient dimensionality reduction from 150 to 16 dimensions. Decoder part (signal reconstruction): The structure is strictly symmetrical with the encoder. The third hidden layer contains 32 neurons (ReLU activation), the fourth hidden layer contains 64 neurons (ReLU activation), and the output layer contains 150 neurons (Linear activation), which are responsible for restoring the 16-dimensional fingerprint features into a 150-dimensional spectral sequence.

[0168] (2) Interlayer connectivity and parameter initialization:

[0169] The network uses a fully connected approach between adjacent layers, meaning that every neuron in the previous layer is connected to all neurons in the next layer. Let the output vector of the l-th layer be... Then the output calculation formula for the (l+1)th layer is: ,in This is the weight matrix. Let be the bias vector, and σ be the activation function. To avoid gradient vanishing or exploding, the weight matrix W uses He initialization, i.e., it starts from a value with a mean of 0 and a variance of 2 / n. in Sampling in a Gaussian distribution (n in (This represents the number of input neurons); the bias vector b is initialized to a zero vector.

[0170] (3) Loss function and optimization strategy:

[0171] Mean Squared Error (MSE) is used as the loss function L to measure the difference between the reconstructed sequence and the original input. For training samples with a batch size of N, the loss function is defined as follows: ;in, This represents the original 150-dimensional input vector of the i-th sample. This represents the reconstructed vector output by the network. This represents the Euclidean norm.

[0172] The model training employed the Adam Optimizer algorithm, with an initial learning rate of 0.001, a first-order moment estimation decay rate β1 = 0.9, and a second-order moment estimation decay rate β2 = 0.999. During training, the sample set was divided into a training set (80%) and a validation set (20%), with a batch size of 64. The maximum number of training epochs was set to 200, and an early stopping mechanism was introduced: if the validation set loss did not decrease within 10 consecutive epochs, training was terminated early to prevent overfitting.

[0173] 2. Standard fingerprint feature extraction and library generation:

[0174] (1) Feature Extractor Deployment: After training, the encoder part (from the input layer to the 16-dimensional bottleneck layer) in the network is truncated as the final deep feature extractor, while the decoder part is discarded. This extractor can directly map any 150-dimensional spectral encoding sequence into a 16-dimensional deep fingerprint feature vector.

[0175] (2) Standard fingerprint calculation: Input all labeled training samples into the feature extractor described above to obtain the corresponding 16-dimensional depth fingerprint feature set. For each type of land cover (such as farmland, forest land, water body, etc.), calculate the arithmetic mean of the depth feature vectors of all samples under that type, and define the mean vector as the "standard pixel image fingerprint" of that type. This step eliminates the random noise of a single sample and preserves the core spectral temporal regularity of that type of land cover.

[0176] 3. Adaptive recognition threshold setting:

[0177] To ensure the confidence level of the recognition results, this embodiment automatically sets a dynamic recognition threshold for each type of land cover. Intra-class distance distribution calculation: All training samples of a certain land cover type are traversed, and the Hamming distance (or Euclidean distance) between the depth fingerprint of each sample and the "standard fingerprint" of that type is calculated, forming the intra-class distance distribution for that land cover type. Inter-class distance distribution calculation: The distances between samples of this type and all other dissimilar standard fingerprints are calculated, forming the inter-class distance distribution, used to evaluate inter-class separability. Threshold determination: The 95th percentile of the intra-class distance distribution is taken as the high-confidence recognition threshold for that type. This means that as long as the distance between the pixel to be identified and the standard fingerprint is less than this threshold, the confidence level of its recognition result meets the preset requirement (i.e., covering 95% of normal samples). This threshold can be fine-tuned based on the overall accuracy of the validation set, thereby constructing a complete standard pixel image fingerprint library containing the "standard fingerprint vector" and the "dynamic recognition threshold" for subsequent rapid land cover matching and recognition.

[0178] Through the above process, this embodiment not only achieves effective dimensionality reduction and noise reduction of high-dimensional spectral data, but also establishes an adaptive threshold mechanism based on statistical principles, which significantly improves the robustness and accuracy of ground feature identification in multi-temporal multispectral images.

[0179] The algorithm iterates through the image encoding sequence of each pixel in the structured data file to be identified, extracts the deep fingerprint features to be identified using the trained autoencoder, calculates the similarity between the deep fingerprint features to be identified and the standard pixel image fingerprint, matches the similarity with a high confidence recognition threshold, and generates the target ground object identification result.

[0180] Specifically, the process iterates through the image encoding sequence of each pixel in the structured data file to be identified, extracts the depth fingerprint features using a trained autoencoder, calculates the similarity between the depth fingerprint features and the standard pixel image fingerprint, and matches the similarity with a high-confidence recognition threshold to generate target feature identification results, including:

[0181] Read the standard pixel image fingerprint features of one or more target land features specified by the user from the standard pixel image fingerprint database;

[0182] Traverse the structured data file to be identified, input the spectral encoding sequence of each pixel image into the trained autoencoder to obtain the deep fingerprint features to be identified;

[0183] Calculate the Hamming distance between the depth fingerprint feature to be identified and the standard pixel image fingerprint features of various target features, and use the Hamming distance as the similarity between the depth fingerprint feature to be identified and the standard pixel image fingerprint.

[0184] If the minimum similarity is less than or equal to the high confidence threshold of the corresponding target feature category, the pixel image is determined to be the corresponding target feature category and assigned a category label; otherwise, it is marked as unrecognized, and the target feature recognition result is obtained.

[0185] Specifically, the process iterates through each spectral feature vector in the tabular dataset to be identified, calculates the Hamming distance between the spectral feature vector to be classified and the spectral feature vectors in the spectral fingerprint template library, and uses a multi-round progressive threshold matching mechanism to progressively match the Hamming distance. By gradually relaxing the threshold in each round, spectral feature vectors with different confidence levels are covered to generate land cover identification results. Specifically, this includes reading standard spectral feature vectors of several types of land cover specified by the user from the spectral fingerprint template library as a reference sequence, and allowing the user to customize the matching rounds and round thresholds; according to the customized matching rounds, it iterates through the tabular dataset to be classified, and calculates the Hamming distance between each spectral feature vector and the spectral feature vectors in the spectral fingerprint template library of the user-specified type in turn; progressively matching the Hamming distance according to the multi-round progressive threshold matching mechanism, assigning the corresponding category label to the spectral feature vector that meets the minimum threshold in the first round, and the unmatched spectral feature vectors enter the subsequent rounds, and are rematched and re-identified in turn after relaxing the threshold in each round to generate complete land cover identification results.

[0186] Specifically, the matching process is initialized by customizing a progressively stricter threshold sequence and inputting it into the spectral fingerprint template library. A multi-round iterative process is then initiated. The first round iterates through all spectral feature vectors to be classified, using Hamming distance and the strictest threshold to determine which spectral feature vectors are successfully matched, generating a set of unmatched pixels. Each matching round processes the unmatched pixel set generated in the previous round, repeatedly performing matching determination and recognition using a progressively looser threshold sequence, dynamically updating the unmatched pixel set. This iterative loop continues until the number of rounds exceeds the preset threshold or the unmatched pixel set is empty, at which point the loop terminates. Finally, the successfully matched pixel recognition results from all rounds are integrated, and the remaining unmatched pixels are uniformly labeled to generate a complete land cover recognition result.

[0187] Specifically, if the Hamming distance does not exceed the threshold of the current round, the cell value of the spectral feature vector is replaced with the replacement value of the corresponding category, and it is regarded as a spectral feature vector successfully identified in this round, ending the subsequent matching of the spectral feature vector; if the Hamming distance exceeds the current matching threshold, it is determined that the spectral feature vector does not have sufficient similarity with any land cover template in the current round, the spectral feature vector is retained, and the spectral feature vector is input into the unmatched cell set for the next round of matching.

[0188] Specifically, such as Figure 8 The diagram illustrates the core steps of the entire process, corresponding to the feature comparison and identity authentication process in fingerprint recognition. The computationally efficient Hamming distance is used as the similarity metric, comparing the "spectral fingerprint" of the pixel to be classified against various user-specified standard templates in the template library one by one. A multi-level progressive threshold matching strategy is employed: the first round uses a strict threshold to ensure accurate identification of high-confidence pixels; subsequent rounds gradually relax the threshold, gradually incorporating edge pixels with certain spectral variations, until the preset recognition coverage or accuracy target is achieved. Finally, through label mapping and spatial reconstruction, a complete land cover identification result is generated.

[0189] This step uses Hamming distance as a similarity metric and employs multi-round progressive threshold matching to achieve accurate on-demand identification of ground features, balancing identification accuracy with coverage integrity.

[0190] I. Hamming distance calculation principle: For a 150-dimensional feature vector of a pixel to be classified and the standard spectral fingerprint vector of a certain type of land cover in the template library The Hamming distance is defined as the number of times corresponding positions of two objects differ, and the calculation formula is as follows: In the formula, Indicates Hamming distance; This indicates an indicator function, which takes a value of 1 when the condition within the parentheses is true, and 0 otherwise. This metric is computationally efficient, robust to noise, and suitable for rapid matching of large-scale multispectral images.

[0191] II. Parameter Configuration: Running the Program: Execute the "03match.py" program. The user needs to input the following core parameters. Sample Folder Path: Points to the generated standard spectral fingerprint template library folder. CSV File Path to be Matched: Points to the "merged_data.csv" file generated by image preprocessing. Output CSV File Path: Specifies the path to save the recognition results. Recognition Type Selection: The user can customize the target land cover type to be recognized. For example, entering "water body" means recognizing only water bodies; entering "artificial surface, farmland" means recognizing both artificial surfaces and farmland simultaneously, with other types considered unclassified. This embodiment uses the recognition of water body land cover type as an example. Matching Rounds and Thresholds: Set the number of matching rounds (7 rounds in this embodiment). Each round corresponds to the Hamming distance threshold and category replacement value, as shown in Table 2. The number of matching rounds is user-defined and set according to the recognition effect the user wants to achieve. As the number of matching rounds increases, the land cover recognition accuracy will also increase.

[0192] Table 2 Matching Rounds Table

[0193]

[0194] III. Iterative Matching Execution: Matching Process: The program reads the "merged_data.csv" file and iterates through the 150-dimensional feature sequence of each cell. According to the set matching rounds, it calculates the Hamming distance between the sequence to be classified and the standard spectral fingerprints of each category in the template library of the user-specified type round after round. If the Hamming distance is less than or equal to the threshold of the current round, the cell value is replaced with the corresponding replacement value for that category, and the matching for that pixel in the current round ends; if the Hamming distance is greater than the current matching threshold, the cell does not have similarity to the land cover type, no replacement is performed, and the next round of matching begins. Core Advantage: Through a "strict to lenient" threshold strategy (the threshold gradually widens from 15 to 45), the first round ensures accurate identification of high-confidence areas, and subsequent rounds include edge or variant pixels, improving recognition coverage and robustness.

[0195] IV. Round-based Threshold Optimization Strategy: The threshold for each round of matching is different, calculated according to rules ranging from strict to lenient. The formula for calculating the round-based threshold is: In the formula, L=150 represents the length of the spectral eigenvector; The threshold coefficient is denoted as 0.10-0.15 for strict matching, 0.20-0.35 for general matching, and 0.40-0.50 for lenient matching, as shown in Table 4. Optimization logic: After each round of matching, the overall accuracy (OA) is calculated, and the trend of accuracy changes is tracked. As shown in Table 3, the threshold combination that achieves the optimal OA is selected, balancing classification accuracy and coverage integrity.

[0196] Table 3. Trend of Tracking Accuracy

[0197]

[0198] Table 4 Threshold Coefficients Value table

[0199]

[0200] Example: L=100, for strict matching: T=0.1×100=10; In this study, the feature vector to be matched is 150 bits long. Based on the strict matching principle, the initial threshold is set to 15. During the matching process, the feature values ​​of each land cover sample are compared one by one with the 150-bit feature vector corresponding to each pixel in the merged CSV file; if the Hamming distance between the two does not exceed the current threshold, the pixel is considered to belong to the land cover category and is assigned the corresponding category label (assigned a new value). These pixels that are successfully classified in the first round of matching have the highest confidence and constitute the most reliable land cover identification results in subsequent analysis.

[0201] This invention designs a refined, adaptive, and controllable multi-level progressive spectral matching process. This process aims to maximize classification coverage while ensuring classification accuracy in high-confidence regions through a strategy of multiple iterations and progressively relaxed thresholds, achieving effective identification of complex spectral variations and mixed pixels. I. Overall Architecture of the Matching Process: The process follows a basic framework of "initialization, multi-round iterative matching, and result integration." Each round of matching uses a unified "spectral fingerprint" template library, but with different Hamming distance thresholds. The process starts with the most stringent threshold and gradually relaxes it, forming a progressive identification mode "from core to edge, from certainty to probing."

[0202] II. Detailed Step-by-Step Breakdown:

[0203] Step 1: System Initialization and First Round of High-Precision Matching. Setting the initial threshold: Based on the spectral fingerprint encoding length (L) and strict matching principles, using the formula... ; Calculate the first round matching threshold In the formula, the threshold coefficient is... A small value, such as 0.1 to 0.15, is typically chosen to ensure extremely high matching confidence. The first round of matching is performed: all pixels to be classified are traversed, and the Hamming distance between their spectral feature vectors and the standard fingerprints of each class in the template library is calculated. If the distance between a pixel and a certain class template is less than or equal to... If a pixel is found to be in the correct category, it is classified as "classified". This round of matching results corresponds to the most typical and purest spectral features of the ground cover area in the image, possessing the highest classification confidence. A set of unmatched pixels is generated: all pixels that did not find any match in this round (i.e., whose distance to all templates is greater than a certain value) are included. The cells are recorded to form an "unmatched cell set", which serves as the input for the next round of matching.

[0204] Step 2: Multi-round progressive matching and threshold relaxation, threshold increment strategy: Set a threshold increment sequence ,in The increment can be set empirically (e.g., fixed step size) or dynamically adjusted based on the success rate of the previous matching round. Iterative matching round by round: For the... Wheel matching ( ≥2): Update threshold: Use a new, more lenient threshold. Limit the matching scope: Only perform matching calculations on pixels in the "unmatched pixel set". Pixels already classified in previous rounds will not participate in subsequent matching to avoid duplicate calculations and potential erroneous coverage. Perform matching and judgment: Calculate the Hamming distance between these unmatched pixels and each type of template. If the distance between a pixel and a certain type of template is less than or equal to... If a cell is found to belong to a certain category, it is removed from the "unmatched cell set" and added to the classified results. Update the set: After this round, update the "unmatched cell set," which only contains cells that have not found a class after k rounds of matching. Loop termination condition: Iterative matching continues until one of the following conditions is met: The preset maximum number of rounds n is reached. The "unmatched cell set" is empty (all cells have been classified). Further relaxing the threshold no longer significantly increases the number of matched cells.

[0205] Step 3: Post-processing and Output of Results, Final Category Integration: Integrate all successfully matched pixels from all rounds according to their classification round and category to form a complete classification result map. The "round" information of a pixel's classification can be retained as an indirect measure of its classification confidence (generally, the earlier the round, the higher the confidence). Handling Unclassified Pixels: For pixels that remain unclassified after all rounds of matching (i.e., elements in the final "unmatched pixel set"), special processing can be performed, such as assigning a "unclassified" or "unknown" label. Auxiliary methods such as nearest neighbor assignment and spatial context smoothing are used for estimation. Their original spectral features are retained for manual interpretation or further analysis. Output and Evaluation: Output the final land cover classification results and calculate evaluation indicators such as overall accuracy (OA). A "classification confidence layer" (based on matching round or minimum Hamming distance) can also be output simultaneously to provide users with spatial distribution information on the reliability of the results.

[0206] III. The advantages of this invention lie in the balance between accuracy and coverage: This process achieves a dynamic balance between "strict matching to ensure accuracy" and "relaxed matching to improve coverage." The first round of high-threshold screening eliminates most uncertainties, while subsequent rounds gradually absorb pixels with similar spectral characteristics but reasonable variations. Combating "different spectra for the same object" and noise: For similar ground features whose spectra vary due to phenological, atmospheric, or lighting conditions (different spectra for the same object), a strict single threshold might exclude them. Progressive matching, by gradually relaxing tolerances, can "recapture" these varied pixels back to the correct category. Optimized computational efficiency: Since each round of matching only calculates for the remaining unidentified pixels from the previous round, repeated distance calculations for already identified pixels are avoided, effectively controlling the overall computational load. Providing decision transparency: The phased results of the process (identification maps for each round) and the final integrated results make the identification decision process more transparent, facilitating understanding and analysis of identification performance under different thresholds.

[0207] In summary, the multi-step progressive matching process proposed in this invention is not simply a relaxation of the threshold, but an intelligent identification decision with clear iterative logic, dynamic range, and termination mechanism. It fully utilizes the detailed information of multi-temporal spectral data, and through a layered and progressive approach, achieves robust, precise, and highly interpretable ground feature identification in complex remote sensing scenes. This is the distinguishing feature of this invention from traditional one-time classification algorithms.

[0208] The target feature identification results are reconstructed into a two-dimensional classification matrix, and the georeferenced information of the original multi-temporal multispectral images is reused to generate a classification result file.

[0209] Specifically, the target feature identification results are reconstructed into a two-dimensional classification matrix, and the georeferenced information of the original multi-temporal multispectral imagery is reused to generate classification result files, including:

[0210] Unidentified cells in the target feature identification results are uniformly assigned the value of zero to generate a pure two-dimensional classification matrix;

[0211] The classification result files include: the standard ENVI classification file and its header file;

[0212] By combining the two-dimensional classification matrix with the geographic transformation parameters and coordinate system information extracted from the original multi-temporal multispectral imagery, a standard ENVI classification file with geographic reference and its header file are generated.

[0213] Specifically, such as Figure 9 As shown, this step converts the matched CSV format recognition results into a standard remote sensing image format with georeferenced information, which facilitates subsequent visualization analysis and professional applications.

[0214] I. Data Format Cleaning and Conversion: The program to be run is "04envi.py". The process is as follows: Data cleaning involves reading the output recognition result table file, replacing unmatched cells (still the original 150-dimensional sequence or empty values) with 0, and setting the first cell of the file to 0 to ensure proper formatting; Format conversion involves converting the cleaned table file into a plain text matrix format (TXT format), with each row of the table corresponding to each row of the TXT file, and cell values ​​separated by spaces, forming a two-dimensional classification matrix.

[0215] II. Georeferenced Information Reuse and ENVI File Generation includes geographic information extraction, where the program provides an option to add geographic coordinates. If "Yes" is selected, the path to the original multispectral image TIFF file must be entered to read the geographic transformation parameters (affine transformation) and coordinate system information from the file. Image generation involves using the rasterio library to read the TXT format classification matrix, and based on the matrix shape (number of rows and columns) and the extracted geographic information, writing it into a new ENVI format raster file (.dat) and generating a corresponding header file (.hdr). The header file fully records geographic metadata such as data size, data type, coordinate system, cell size, and top-left corner coordinates, ultimately generating standard ENVI classification data with georeferenced information.

[0216] By using reclassification mapping rules, the classification result files are reclassified and merged, and visualization colors are configured to output target feature recognition products, thereby enabling the recognition of target features in multi-temporal and multispectral images.

[0217] Specifically, by using reclassification mapping rules, the classification result files are reclassified and merged, and visualization colors are configured to output target feature recognition products, thereby achieving the recognition of target features in multi-temporal and multispectral imagery, including:

[0218] Based on user-defined reclassification mapping rules, the category labels in the standard ENVI classification files and their header files are uniformly merged into a single category identifier, and the identifier values ​​of unrecognized areas are standardized to generate standardized classification data.

[0219] Configure a specified display color for each category in the normalized classification data, and write the display color along with georeferenced information into a new file to generate a target feature identification product for thematic mapping;

[0220] By comparing and analyzing the target feature recognition product with the verification sample and calculating the overall accuracy index, the recognition effect is quantitatively evaluated based on the overall accuracy index.

[0221] Specifically, such as Figure 10 As shown, this step generates an intuitive thematic map for identification through reclassification, color assignment, and accuracy verification, and verifies the effectiveness of the method.

[0222] I. Reclassification processing involves executing the program "05Reclassification.py", and the core process is as follows:

[0223] Image information reading: Read the generated ENVI.dat file and .hdr header file, and parse the image size, data type and geographic information; Reclassification rule setting: The user inputs the final number of land cover categories (4 categories in this example), and specifies the original numerical range for each category, as shown in Table 5.

[0224] Table 5 Final Number of Land Cover Categories

[0225]

[0226] Category merging: The program merges the category labels from different rounds in the .dat file into a single category identifier according to the above rules, and assigns a value of 0 to unclassified pixels.

[0227] II. Thematic Map Generation and Output: Color assignment involves predefining visual colors for each final category and writing the color table into the ENVI header file. The output is the reclassified data saved as a new ENVI.dat file and a matching .hdr header file, forming the final thematic map, which can be directly opened in software such as ENVI and ArcGIS for easy visual analysis and cartographic applications.

[0228] III. Classification Accuracy Verification and Comparative Analysis. The verification metrics used were the confusion matrix and overall accuracy (OA) as the core evaluation indicators. The confusion matrix is ​​a C×C table (where C is the total number of categories), used to visually display the correspondence between the classification results and the true labels. Let the first... The row represents the true category, the first row... The columns represent the predicted categories, then the matrix elements Indicates that the reality belongs to the category. However, it was predicted by the classifier to be of category [missing information]. The number of samples. Diagonal elements. This represents the number of correctly classified samples in each category; off-diagonal elements reflect misclassifications. Overall accuracy (OA): the proportion of correctly classified samples out of the total sample size; overall accuracy is defined as the proportion of all correctly classified samples out of the total sample size, and its mathematical expression is:

[0229] ;

[0230] In the formula, OA represents the most intuitive accuracy indicator, but it may lead to misleading overestimation in cases of class imbalance.

[0231] Results verification: After 7 rounds of matching, the overall accuracy of this embodiment reached 92.31%, and the recognition effect was excellent, verifying the effectiveness of this method.

[0232] In summary, this invention effectively solves the identification problem caused by limited bands and temporal spectral variations in multi-temporal and multispectral data by constructing pixel image fingerprints and introducing deep learning autoencoders for feature learning and automatic threshold setting. It significantly improves the accuracy, robustness, and automation of ground feature identification, and supports users to select the type of target ground feature to be identified as needed. It can be widely used in high-precision ground feature identification and thematic mapping in fields such as agricultural monitoring, ecological environment assessment, and land and resources surveys.

[0233] According to another embodiment of the present invention, a multi-temporal multispectral image land cover recognition system based on pixel image fingerprints is also provided, the system comprising:

[0234] The image processing module is used to preprocess the original multi-temporal multispectral images to obtain surface reflectance data, calculate several characteristic index bands based on the surface reflectance data, integrate the original bands of the surface reflectance data with all the characteristic index bands to obtain a pixel image, discretize the pixel image to obtain a spectral coding sequence, and organize the spectral coding sequence into a structured data file.

[0235] The training sample set construction module is used to select sample points of various target features based on high-resolution reference images and visual interpretation, and extract the corresponding pixel image encoding sequences from structured data files using the spatial coordinates of the sample points to construct the training sample set.

[0236] The fingerprint database construction module is used to input the training sample set into the deep autoencoder network for training, use the trained autoencoder to learn features from all training samples to generate deep fingerprint features, perform mean processing on all deep fingerprint features to generate a standard pixel image fingerprint database, and analyze the distance distribution of each type of target object between the deep fingerprint features and the standard pixel image fingerprint, and set a high confidence recognition threshold based on the distance distribution of each type of target object.

[0237] The iterative matching module is used to traverse the image encoding sequence of each pixel in the structured data file to be identified, extract the deep fingerprint features to be identified using the trained autoencoder, calculate the similarity between the deep fingerprint features to be identified and the standard pixel image fingerprint, match the similarity with the high confidence recognition threshold, and generate the target ground object identification result.

[0238] The geocoding module is used to reconstruct the target feature identification results into a two-dimensional classification matrix and reuse the georeference information of the original multi-temporal multispectral imagery to generate a classification result file;

[0239] The productization module is used to reclassify and merge the classification result files through reclassification mapping rules, configure visualization colors, and output target feature recognition products to achieve the recognition of target features in multi-temporal multispectral images.

[0240] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying ground features in multi-temporal, multispectral images based on pixel image fingerprints, characterized in that, include: The original multi-temporal multispectral images are preprocessed to obtain surface reflectance data; several characteristic index bands are calculated based on the surface reflectance data, and the original bands of the surface reflectance data are integrated with all the characteristic index bands to obtain a pixel image; the pixel image is discretized to obtain a spectral coding sequence, and the spectral coding sequence is organized into a structured data file; Based on high-resolution reference images and visual interpretation, sample points of various target features are selected, and the corresponding pixel image encoding sequences are extracted from structured data files using the spatial coordinates of the sample points to construct a training sample set. The training sample set is input into the deep autoencoder network for training. The trained autoencoder is then used to learn features from all training samples to generate deep fingerprint features. The mean value of all deep fingerprint features is processed to generate a standard pixel image fingerprint database, and the distance distribution of each type of target object between the deep fingerprint features and the standard pixel image fingerprint is analyzed; a high confidence recognition threshold is set based on the distance distribution of each type of target object. The algorithm iterates through the image encoding sequence of each pixel in the structured data file to be identified, extracts the deep fingerprint features to be identified using the trained autoencoder, calculates the similarity between the deep fingerprint features to be identified and the standard pixel image fingerprint, matches the similarity with a high confidence recognition threshold, and generates the target ground object identification result. The target feature identification results are reconstructed into a two-dimensional classification matrix, and the georeferenced information of the original multi-temporal multispectral images is reused to generate a classification result file; By using reclassification mapping rules, the classification result files are reclassified and merged, and visualization colors are configured to output target feature recognition products, thereby enabling the recognition of target features in multi-temporal and multispectral images.

2. The multi-temporal multispectral image land cover recognition method based on pixel image fingerprints according to claim 1, characterized in that, The original multi-temporal multispectral images are preprocessed to obtain surface reflectance data; several characteristic index bands are calculated based on the surface reflectance data, and the original bands of the surface reflectance data are integrated with all the characteristic index bands to obtain a pixel image; The pixel image is discretized to obtain a spectral coding sequence. The spectral coding sequence is then organized into a structured data file, including: The metadata file in the original multi-temporal multispectral image is read, the radiance multi-gain, multi-offset, and reflectance scaling coefficients corresponding to each band in the metadata file are parsed, and the radiance multi-gain coefficient is multiplied with the original quantization value of the image. The product result is summed with the multi-offset coefficient to perform radiometric calibration and obtain the spectral radiance of the top atmosphere. The planetary reflectance of the top atmosphere is calculated by combining the spectral radiance of the top atmosphere with the Earth-Sun distance, solar zenith angle, and solar irradiance parameters. Based on the latitude and longitude and season of the image, the standard atmospheric model and aerosol model are automatically matched. The water vapor content of the atmospheric column is inverted using the water vapor absorption band. Then, the neighborhood effect correction is performed by the atmospheric correction algorithm based on the radiative transfer model, and the planetary reflectance of the top atmosphere is converted into the surface reflectance. A pre-defined image is selected as the reference. The same feature points between the image to be registered and the reference image are automatically extracted using the scale-invariant feature transformation algorithm. After removing mismatched points, a multinomial transformation model is constructed. The bicubic convolution interpolation method is used to resample all original multi-temporal multispectral images to obtain the geometrically registered spectral image. The geometrically registered spectral images were processed for cloud, cloud shadow and snow masking based on quality assessment bands and spectral characteristics to obtain surface reflectance data. The aforementioned characteristic indices include: Normalized Difference Vegetation Index, Normalized Difference Water Index, and Normalized Difference Building Index. The normalized vegetation index, normalized water index, and normalized building index bands are calculated based on the surface reflectance data. The original bands of the surface reflectance dataset are then stitched together with all the characteristic index bands in sequence to form a pixel image. The characteristic band raster data of all period pixel images are read band by band, converted into a pixel value matrix, and organized into temporary files according to row and column positions. The continuous pixel values ​​of each band in the temporary file are discretized using the quantile discretization method. The continuous values ​​are then converted into integer values ​​according to the mapping rules to obtain the spectral coding sequence of each pixel. The spectral coding sequence of each pixel is horizontally spliced ​​according to the pixel row and column positions to generate a structured data file.

3. The multi-temporal multispectral image land cover recognition method based on pixel image fingerprints according to claim 2, characterized in that, The geometrically registered spectral image is processed for cloud, cloud shadow, and snow masking based on quality assessment bands and spectral characteristics to obtain surface reflectance data, including: The quality assessment bands inherent in the geometrically registered spectral image are decoded, and key marker bits are extracted through bit operations. These key marker bits include dilatant clouds, cirrus clouds, high-confidence clouds, cloud shadows, and snow or ice, generating the quality assessment band decoding results. Normalized differential snow index is calculated for the geometrically registered spectral image, and snow is distinguished from bright clouds by combining the low reflectance characteristics of the near-infrared band. Low-temperature cloud tops are identified by using the thermal infrared band temperature threshold, and morphological dilation of cloud pixels in the geometrically registered spectral image is performed based on the solar azimuth angle to obtain spectral enhancement discrimination results. The quality assessment band decoding results and the spectral enhancement discrimination results are logically ORed to generate a composite mask image. After traversing the spectral images after multiple geometric registrations, the pixel positions marked as clouds, cloud shadows, or snow in the composite mask image are assigned invalid values, and the output is surface reflectance data containing only valid surface information.

4. The multi-temporal multispectral image land cover identification method based on pixel image fingerprints according to claim 2, characterized in that, The formula for calculating the normalized vegetation index is as follows: NDVI = (NIR - Red) / (NIR + Red); The formula for calculating the normalized water index is as follows: NDWI=(Green-NIR) / (Green+NIR); The formula for calculating the normalized building index is as follows: NDBI=(SWIR1-NIR) / (SWIR1+NIR); In the formula, NDVI represents the normalized vegetation index; NDWI represents the normalized water index; NDBI represents the normalized building index; NIR represents the near-infrared reflectance; Red represents the red band; Green represents the green band; and SWIR1 represents the shortwave infrared band.

5. The multi-temporal multispectral image land cover identification method based on pixel image fingerprints according to claim 1, characterized in that, The process of selecting sample points for various target features based on high-resolution reference images and visual interpretation, and extracting corresponding pixel image encoding sequences from structured data files using the spatial coordinates of the sample points to construct a training sample set includes: By visually interpreting and importing vector sample points and high-resolution reference images in batches, the vector sample points and high-resolution reference images are automatically aligned in coordinate system and space, and the pixel row and column number of each sample point on the original multi-temporal multispectral image is extracted. Based on the pixel row and column number of each sample point, the complete pixel image encoding sequence of each sample point is located and extracted from the digitized structured data file; The extracted pixel image encoding sequences are automatically organized according to land cover categories and output as structured sample files to form a training sample set.

6. The multi-temporal multispectral image land cover recognition method based on pixel image fingerprints according to claim 5, characterized in that, The batch import of vector sample points and high-resolution reference images automatically aligns the vector sample points with the high-resolution reference images in coordinate system 1 and space, and extracts the pixel row and column numbers of each sample point on the original multi-temporal multispectral image, including: Based on the land feature identification standards and referring to high-resolution reference images, sample points of various typical land features are selected through visual interpretation. Record and store the spatial location and category label of the sample points as vector point sample points with geographic coordinates; The vector point sample points are uniformly transformed to a spatial reference system consistent with the original multi-temporal multispectral image, and the pixel row and column number corresponding to each sample point is calculated.

7. The multi-temporal multispectral image land cover recognition method based on pixel image fingerprinting according to claim 1, characterized in that, The training sample set is input into the deep autoencoder network for training, and the trained autoencoder is used to learn features from all training samples to generate deep fingerprint features. The mean value of all deep fingerprint features is processed to generate a standard pixel image fingerprint database, and the distance distribution of each category of target land cover between the deep fingerprint features and the standard pixel image fingerprint is analyzed. High-confidence identification thresholds are set based on the distance distribution of various target features, including: A deep autoencoder network is constructed, with the pixel image encoding sequence of the training sample set as input. Low-dimensional depth features are extracted through the deep learning autoencoder, and the input is reconstructed through the deep learning autoencoder to train the deep autoencoder network and minimize the reconstruction error. After training is completed, the trained autoencoder is used as the feature extractor. All training samples in the training sample set are input into the feature extractor for feature extraction to obtain the deep fingerprint features of each training sample. The mean value of the depth fingerprint features of all samples of each type of target land cover is calculated, and the mean value is used as the standard pixel image fingerprint of each type of target land cover to generate a standard pixel image fingerprint library for each type of target land cover. The distance between the depth fingerprint features of each training sample and the standard pixel image fingerprint of the corresponding class of target land cover is calculated to obtain the intra-class distance distribution. The distance between the training sample and the standard pixel image fingerprint of the target land cover of the opposite class is calculated to obtain the inter-class distance distribution. The high confidence recognition threshold of the class land cover is automatically set based on the statistics of the intra-class distance distribution.

8. The multi-temporal multispectral image land cover identification method based on pixel image fingerprinting according to claim 1, characterized in that, The process involves traversing the image encoding sequence of each pixel in the structured data file to be identified, extracting the depth fingerprint features to be identified using a trained autoencoder, calculating the similarity between the depth fingerprint features to be identified and the standard pixel image fingerprint, and matching the similarity with a high-confidence recognition threshold to generate the target land cover identification result, including: Read the standard pixel image fingerprint features of one or more target land features specified by the user from the standard pixel image fingerprint database; Traverse the structured data file to be identified, input the spectral encoding sequence of each pixel image into the trained autoencoder to obtain the deep fingerprint features to be identified; Calculate the Hamming distance between the depth fingerprint feature to be identified and the standard pixel image fingerprint features of various target features, and use the Hamming distance as the similarity between the depth fingerprint feature to be identified and the standard pixel image fingerprint. If the minimum similarity is less than or equal to the high confidence threshold of the corresponding target feature category, the pixel image is determined to be the corresponding target feature category and assigned a category label; otherwise, it is marked as unrecognized, and the target feature recognition result is obtained.

9. The method for identifying ground features based on pixel image fingerprints in multi-temporal and multispectral imagery according to claim 1, characterized in that, The process of reconstructing the target feature identification results into a two-dimensional classification matrix and reusing the georeferenced information of the original multi-temporal multispectral imagery to generate a classification result file includes: Unidentified cells in the target feature identification results are uniformly assigned the value of zero to generate a pure two-dimensional classification matrix; The classification result file includes: the standard ENVI classification file and its header file; By combining the two-dimensional classification matrix with the geographic transformation parameters and coordinate system information extracted from the original multi-temporal multispectral imagery, a standard ENVI classification file with geographic reference and its header file are generated.

10. The multi-temporal multispectral image land cover identification method based on pixel image fingerprints according to claim 1, characterized in that, The process of reclassifying and merging the classification result files using reclassification mapping rules, configuring visualization colors, and outputting target feature recognition products to achieve the recognition of target features in multi-temporal and multispectral images includes: Based on user-defined reclassification mapping rules, the category labels in the standard ENVI classification files and their header files are uniformly merged into a single category identifier, and the identifier values ​​of unrecognized areas are standardized to generate standardized classification data. Configure a specified display color for each category in the normalized classification data, and write the display color along with georeferenced information into a new file to generate a target feature identification product for thematic mapping; By comparing and analyzing the target feature recognition product with the verification sample and calculating the overall accuracy index, the recognition effect is quantitatively evaluated based on the overall accuracy index.

Images