Natural resource remote sensing image data quality improvement method

Abstract

The invention discloses a natural resource remote sensing image data quality improvement method, and particularly relates to the technical field of natural resource remote sensing image processing, comprising the following steps: acquiring remote sensing image data and auxiliary data of a natural resource scene, the remote sensing image data comprises a spectral imaging type image, a scanning imaging type image, a radar imaging type image and a photographic imaging type image. Through multi-dimensional differential comparative analysis and construction of a feature quality evaluation matrix, core dominant features of various images can be accurately extracted, overlapped redundant data are eliminated, and then standardized preprocessing and regional screening fusion are combined, so that the ratio of invalid data is greatly reduced, the data processing efficiency is remarkably improved, and meanwhile, the consumption of storage resources is reduced; the method greatly improves the quality of natural resource remote sensing image data, and effectively solves the problems of redundancy of a large amount of information, key feature loss and poor precision adaptability in traditional fusion.

Description

Technical Field

[0001] This invention relates to the field of natural resource remote sensing image processing technology, specifically to a method for improving the quality of natural resource remote sensing image data. Background Technology

[0002] Natural resources refer to all tangible and intangible substances in nature that humans can directly obtain for production and daily life, and that can be directly or indirectly used to meet human needs. These include air, water, land, forests, grasslands, wildlife, various minerals, and energy sources. Social development and scientific and technological progress necessitate the development and utilization of increasingly more natural resources. Therefore, accurate remote sensing of natural resources is becoming increasingly important. Remote sensing imagery refers to films or photographs that record the electromagnetic wave intensity of various land features. Through remote sensing imagery, we can quickly obtain information such as land use types, forest coverage, and mineral resource distribution, providing a data foundation for surveying.

[0003] Currently, in the existing natural resource remote sensing image processing, multi-source data stitching and overlay is the core method to improve image quality and enhance the dimensionality of image information. However, the combination methods generally rely on simple overlay fusion, while different types of data sources have different advantages and disadvantages and focuses. Furthermore, the fusion of a large amount of repetitive or valueless data not only leads to a significant increase in the information redundancy of the fused image, but also directly increases the data storage and processing costs, making it difficult to meet the accuracy and efficiency requirements of natural resource monitoring. Therefore, this application proposes a method for improving the quality of natural resource remote sensing image data. Summary of the Invention

[0004] The purpose of this invention is to provide a method for improving the quality of remote sensing image data of natural resources. By combining images of various imaging types and auxiliary data, abnormal data is initially processed and integrated to obtain preprocessed data, thereby significantly improving data processing efficiency. Furthermore, a feature evaluation matrix is ​​established to clarify the advantages and disadvantages of various types of images, extract the core technical features of each type of image data, remove redundant data overlapping with other images, and perform deep fusion of multiple features for further correction and processing, thereby obtaining high-quality remote sensing images. This effectively solves the problems of excessive information redundancy, loss of key features, and poor accuracy and adaptability in traditional fusion methods, thus addressing the aforementioned shortcomings of the technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for improving the quality of remote sensing image data of natural resources, comprising the following steps:

[0006] S1. Acquire remote sensing image data and auxiliary data of natural resource scenes. The remote sensing image data includes: spectral imaging images, scanning imaging images, radar imaging images and photogrammetric imaging images. The auxiliary data includes: digital elevation model data, atmospheric profile parameter data, land cover classification data and meteorological data.

[0007] S2. Perform multi-dimensional differential comparison analysis on the images of the remote sensing image data, and construct a feature evaluation matrix to clarify the image advantages and performance disadvantages of the remote sensing image data.

[0008] S3. Based on the feature evaluation matrix, extract the image advantage feature data of the remote sensing image data, and preprocess the corresponding image advantage feature data to obtain a preprocessed image.

[0009] S4. The preprocessed image is magnified using super-resolution technology to improve the spatial resolution of the preprocessed image and obtain a first image set. Based on the auxiliary data, a supplementary model for residual details is constructed. The supplementary model is trained by machine learning and supplements the missing details in the first image set to obtain a second image set.

[0010] S5. Set the image of the remote sensing image data as the base image, and fuse the base image with the second image set to obtain an enhanced image set;

[0011] S6. Divide the enhanced image set into regions, compare the image quality of the same region in the divided enhanced image set, select and retain high-quality image regions, remove low-quality image regions that are blurry or have lost geographical features, re-merge the selected high-quality image regions, and perform correction processing to obtain a preliminary enhanced remote sensing image.

[0012] S7. Perform quality verification on the preliminary enhanced remote sensing image. If the verification result does not reach the preset clarity, re-optimize the extraction range of the superior features based on the feature quality evaluation matrix, and repeat steps S3-S6 until the preset clarity is reached to obtain the improved quality remote sensing image of natural resources.

[0013] Preferably, the multi-dimensional differential comparative analysis includes radiance accuracy, spatial detail reproduction, environmental adaptability, spectral resolution, spatial positioning accuracy, and data redundancy. By comparing and analyzing the adaptation differences of the remote sensing image data in natural resource scenarios through six dimensions, the advantages and disadvantages of the remote sensing image data are determined, and a feature evaluation matrix is ​​established.

[0014] Preferably, the image advantages features of the remote sensing image data include:

[0015] The advantages of photographic imaging images are edge detail features, texture clarity features, and high pixel density features. The advantages of scanning imaging images are large-area continuous coverage features, fast imaging efficiency features, and multi-band compatibility features. The advantages of radar imaging images are microwave penetration features, echo amplitude features, polarization response features, and day / night cloud / rain adaptability features. The advantages of spectral imaging images are multi-band spectral curve features, ground object spectral difference features, and quantitative inversion support features.

[0016] Preferably, the preprocessing of the image advantage feature data includes format standardization, geometric registration, and invalid pixel removal. The format standardization is to convert image advantage feature data from different sources into a unified data format and metadata specification. The geometric registration is to eliminate geometric deviations of multi-source images by establishing a mathematical transformation model to achieve sub-pixel level spatial alignment. The invalid pixel removal is to remove abnormal pixels caused by sensor noise and atmospheric interference.

[0017] Preferably, the acquisition of the remote sensing image data and auxiliary data includes the following steps:

[0018] The spectral imaging image is generated by using a hyperspectral imaging satellite and a spectroscopic system to decompose the radiation of ground objects into multiple narrow bands, and the signals are synchronously received by a multi-channel detector.

[0019] By using pushbroom satellites and sensors to scan ground features point by point or line by line, electromagnetic radiation is received, converted into electrical signals, and then processed to generate the scanned imaging image.

[0020] The radar imaging image is generated by actively transmitting microwave signals and receiving ground object echo signals through synthetic aperture radar equipment and performing amplitude and phase analysis and inversion.

[0021] The photographic image is generated by using aerial photogrammetry equipment and capturing reflected light from ground objects with an optical lens, and then converting the light into a photoelectric image through a photosensitive medium or digital sensor.

[0022] The digital elevation model data is based on the geographic boundaries and imaging time of the remote sensing image data, using 30-meter resolution SRTM data or 10-meter resolution ASTERGDEM data. The atmospheric profile parameter data is based on the imaging time of the remote sensing image data, and real-time data within one hour before and after the imaging time are obtained from the Global Atmospheric Observation Network. The land cover classification data uses the latest version of GlobeLand30 data. The meteorological data is obtained from the nearest meteorological station in the target area, with hourly data for the day.

[0023] Preferably, the construction of the supplementary model includes the following steps:

[0024] The supplementary model is constructed using a random forest algorithm. The input features are the slope, aspect, and elevation variation coefficient of the digital elevation model data, the aerosol optical thickness, atmospheric transmittance, and water vapor content of the atmospheric profile parameter data, the land cover type and land cover degree of the land cover classification data, and the light intensity and land surface temperature of the meteorological data. The output labels are the detailed textures in historical high-resolution images. The supplementary model is then trained by computer learning.

[0025] When supplementing missing details in the first image set, a residual is calculated by comparing the output of the supplementation model with the pixel difference of the first image set. The residual is then superimposed onto the first image set to supplement the details, thus obtaining the second image set.

[0026] Preferably, the fusion of the base image and the second image set includes the following steps:

[0027] The spectral imaging image is set as the first base image, and the second image set is fused into the first base image using a wavelet transform fusion algorithm to obtain the first enhanced image;

[0028] The scanned imaging image is set as the second base image, and the second image set is fused into the second base image using a wavelet transform fusion algorithm to obtain the second enhanced image;

[0029] The radar imaging image is set as the third base image, and the second image set is fused into the third base image using a wavelet transform fusion algorithm to obtain the third enhanced image;

[0030] The photographic image is set as the fourth base image, and the second image set is fused into the fourth base image using a wavelet transform fusion algorithm to obtain the fourth enhanced image;

[0031] The enhanced image set is obtained by using the first enhanced image, the second enhanced image, the third enhanced image, and the fourth enhanced image.

[0032] Preferably, the region segmentation, comparison, and fusion of the enhanced image set includes the following steps:

[0033] The enhanced image set is divided into regions of size NxN. Taking 3x3 as an example, the regions of the first enhanced image are sequentially labeled as n11, n12, n13, n14...n19.

[0034] The regions of the second enhanced image are sequentially labeled as n21, n22, n23, n24…n29;

[0035] The regions of the third enhanced image are sequentially labeled as n31, n32, n33, n34…n39;

[0036] The regions of the fourth enhanced image are sequentially labeled as n41, n42, n43, n44…n49;

[0037] The image quality of the same divided regions is compared, such as n11, n21, n31 and n41, n12, n22, n32, n42, n13, n23, n33 and n43. High-quality image regions within the same divided regions are selected and retained, while low-quality image regions that are blurred or have lost geographic features are removed, resulting in nX1, nX2, nX3...nX9, where X represents the corresponding image in the enhanced image set. The nX1, nX2, nX3...nX9 images are then re-fused using a stitching fusion algorithm and a weighted fusion algorithm to obtain the target enhanced image.

[0038] Preferably, the correction processing of the target enhanced image includes color adjustment, geometric correction, radiometric correction, and noise reduction;

[0039] The color adjustment is based on a standard spectral database of natural resource features, combined with scene lighting conditions, to adjust color balance and saturation, ensuring color consistency in different areas. The geometric correction is based on a digital elevation model to eliminate geometric distortion and projection deviations generated during imaging, and to eliminate deformations caused by sensor attitude or terrain. The radiometric correction corrects the radiometric brightness deviation of the fused image through a radiometric calibration model. The noise processing adopts a multi-dimensional noise reduction strategy, including one or more of spatial domain filtering, frequency domain filtering, and adaptive noise reduction algorithms, to obtain a preliminary enhanced remote sensing image.

[0040] Preferably, the quality verification of the preliminary enhanced remote sensing image includes the following steps:

[0041] The preset resolution quantification index includes spatial resolution compliance rate, detail texture recovery rate and ground feature recognition accuracy rate, wherein the spatial resolution compliance rate is ≥90%, the detail texture recovery rate is ≥88%, and the ground feature recognition accuracy rate is ≥95%.

[0042] The accuracy of the ground feature identification is automatically detected using a deep learning model. Based on the labeled ground feature samples, the proportion of ground features correctly identified by the model to the total number of samples is statistically analyzed.

[0043] The detailed texture restoration rate is calculated by comparing the texture features of the preliminary enhanced remote sensing image and the high-resolution reference image, extracting contrast and pixel grayscale indices using the grayscale co-occurrence matrix, and calculating the similarity ratio between contrast and pixel grayscale.

[0044] The spatial resolution compliance rate is calculated by measuring ground features of known size in the preliminary enhanced remote sensing image;

[0045] If any one of the following indicators—spatial resolution compliance rate, detail texture recovery rate, and ground feature recognition accuracy rate—failes to meet the standard, the verification result is deemed not to have reached the preset clarity.

[0046] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0047] This invention constructs a feature quality evaluation matrix through multi-dimensional differential comparative analysis, which can accurately extract the core advantageous features of various images and eliminate overlapping and redundant data. Combined with standardized preprocessing and regional screening and fusion, it significantly reduces the proportion of invalid data, significantly improves data processing efficiency, and reduces storage resource consumption. It greatly improves the quality of natural resource remote sensing image data and effectively solves the problems of large amount of information redundancy, loss of key features, and poor accuracy and adaptability in traditional fusion.

[0048] This invention achieves comprehensive optimization of image quality by integrating super-resolution technology, residual detail supplementation, and multi-dimensional error correction. The final output image achieves high standards in spatial resolution compliance rate, detail texture restoration rate, and ground feature recognition accuracy, which can accurately support refined natural resource management scenarios such as farmland rights confirmation, mineral monitoring, and ground feature classification.

[0049] This invention covers a variety of images, taking into account the advantages of multi-band resolution, wide coverage, all-weather observation and high detail restoration. It adapts to different natural resource scenarios through multi-dimensional differential analysis and incorporates quality verification and iterative optimization mechanisms to ensure that high-quality images can be stably output even under extreme conditions such as complex terrain and severe weather, thus meeting diverse needs such as natural resource surveying, dynamic monitoring and boundary delineation. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0051] Figure 1 This is a flowchart illustrating the steps of a method for improving the quality of remote sensing image data of natural resources according to the present invention. Detailed Implementation

[0052] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0053] like Figure 1 As shown, the present invention provides a method for improving the quality of remote sensing image data of natural resources, which includes the following seven steps: acquiring remote sensing image data and auxiliary data, comparative analysis and constructing a feature quality evaluation matrix, extracting image advantage feature data and preprocessing, magnification processing and supplementing details, preliminary fusion, dividing regions and re-fusion correction, and quality verification.

[0054] Specifically, for acquiring remote sensing image data and auxiliary data of natural resource scenarios, the remote sensing image data includes: spectral imaging images, scanning imaging images, radar imaging images and photogrammetric imaging images, and the auxiliary data includes: digital elevation model data, atmospheric profile parameter data, land cover classification data and meteorological data.

[0055] In this embodiment, the spectral imaging image is obtained by a hyperspectral imaging satellite and by using a beam splitting system (such as a grating or prism) to decompose the radiation of ground objects into tens to hundreds of continuous narrow bands (such as the 0.4-2.5 μm range) to form a spectral data cube; a typical example is the AVIRIS hyperspectral imager, which has an ultra-high spectral resolution of 224 bands and a bandwidth of 10 nm, and can capture the differences in reflection and radiation of ground objects at different wavelengths;

[0056] Scanning imaging images acquire ground object radiation signals line by line through electronic push-broom (such as SPOT satellite CCD array) or mechanical swipe scan (such as MSS multispectral scanner), forming a one-dimensional continuous and two-dimensional discrete digital image. Taking LandsatTM imagery as an example, it has 7 bands (including the newly added blue band and shortwave infrared band), a ground resolution of 30m (60m for thermal infrared band), and a single imaging coverage of an area of ​​185km×185km, meeting the needs of large-scale dynamic land use monitoring.

[0057] Radar imaging type images are generated by synthetic aperture radar equipment, which actively transmits microwave signals (such as 5cm in C-band and 23cm in L-band) and images are formed by receiving backscattered echoes from ground objects. A typical example is synthetic aperture radar (SAR). Taking RADARSAT-1 as an example, it has multiple imaging modes: standard mode with 25m resolution and wide swath mode with 100m resolution (covering a width of 450km). It supports all-weather and all-time observation and can penetrate vegetation canopy (such as L-band can penetrate 3-5m high forest) and dry soil (penetration depth 5-10cm).

[0058] Photographic imaging is generated by aerial photogrammetry equipment, which focuses the reflected light from ground objects (400-1300nm band) through an optical lens, and then converts it into digital images through photoelectric conversion by a CCD / CMOS sensor. Examples include full-frame cameras carried by UAVs (ground resolution up to 0.1m) and WorldView-3 satellites (0.31m panchromatic resolution, 1.24m multispectral resolution).

[0059] Furthermore, the digital elevation model (DEM) data is based on the geographic boundaries and imaging time of the remote sensing image data, using SRTM data with a resolution of 30 meters or ASTERGDEM data with a resolution of 10 meters.

[0060] Atmospheric profile parameter data is based on the imaging time of remote sensing image data. Real-time data within one hour before and after the imaging time are obtained from the Global Atmospheric Observation Network (GAW) or the China Meteorological Administration radiosonde station database. It includes parameters such as temperature, humidity, air pressure and aerosol optical thickness (AOD) within the altitude range of 0-10km. The spatial range of atmospheric profile parameter data can be cropped to the geographic boundary of remote sensing image to ensure that each pixel can be matched with the corresponding atmospheric profile parameter.

[0061] For land cover classification data, the latest version of GlobeLand30 global land cover data (30-meter resolution) should be preferred, or 10-meter / 20-meter resolution land cover classification data released by provincial / municipal authorities. The data release time should be later than the remote sensing image imaging time (to ensure that the land cover type is consistent with the actual situation). The land cover classification system of the data can be converted into categories that are compatible with this remote sensing image (such as cultivated land, forest land, grassland, built land, water area and bare land, etc.) to facilitate subsequent fusion and calculation with image data.

[0062] Meteorological data is obtained from the nearest meteorological station in the target area (to avoid data distortion caused by stations that are too far away, and priority is given to meteorological stations that include observations of multiple elements). The real-time data includes light intensity (irradiance), wind speed, wind direction, rainfall and relative humidity, etc. Each meteorological element can be converted into a dimensionless value (such as normalized to the 0-1 range) to facilitate subsequent integration into calculations.

[0063] In a specific embodiment of the present invention, by performing multi-dimensional differential comparative analysis on the images of remote sensing image data and constructing a feature evaluation matrix, the image advantages and performance disadvantages of the remote sensing image data are identified.

[0064] Specifically, multi-dimensional differential comparative analysis includes radiance accuracy, spatial detail reproduction, environmental adaptability, spectral resolution, spatial positioning accuracy, and data redundancy. Radiance accuracy measures the degree of agreement between the radiance values ​​of ground objects recorded in images and the actual radiance characteristics of ground objects. It directly determines the authenticity of ground object reflection and emission information and is the core foundation for natural resource type identification (such as vegetation growth and mineral composition determination).

[0065] Spatial detail fidelity characterizes an image's ability to depict minute features and fine textures. It is quantified by spatial resolution (pixel size) and edge sharpness, and directly affects the accuracy of identifying fine resource elements such as farmland boundaries, mining pits, and small tributaries of water bodies. Environmental adaptability is the ability of an image to maintain effective information acquisition under different meteorological and surface environments. It focuses on characteristics such as resistance to clouds and fog, resistance to changes in lighting, and all-weather operation, and is a key indicator for improving the quality of extreme scenarios (such as disaster emergency response and severe weather monitoring).

[0066] Spectral resolution is the ability of an image to distinguish the spectral characteristics of different ground features. It is determined by the number of bands and the bandwidth of the bands. It is a core indicator for material composition identification (such as mineral alteration information and vegetation pest and disease determination). Spatial positioning accuracy is the degree of agreement between the pixel coordinates of ground features in the image and the actual geographic coordinates (latitude, longitude, and altitude). It is quantified by registration error and directly affects the accuracy of resource boundary delineation (such as farmland ownership confirmation and the delineation of nature reserve boundaries).

[0067] Data redundancy is the proportion of duplicate and invalid information in image data. It is related to band correlation and data compression ratio, and directly affects computational efficiency (such as fusion processing speed and storage cost). It is a key indicator for balancing quality and efficiency. Through quantitative comparison in six dimensions, the advantages and disadvantages of four types of images are determined. Subsequent comparative analysis can establish a feature evaluation matrix based on the performance of the indicators, which will facilitate the design of targeted optimization strategies, use the strengths to make up for the weaknesses, maximize the core value of each type of image, and improve the quality of natural resource remote sensing images.

[0068] In a specific embodiment of the present invention, based on the feature superiority evaluation matrix, image advantage feature data of remote sensing image data is extracted, and the corresponding image advantage feature data is preprocessed to obtain a preprocessed image.

[0069] Based on this, the image advantages of remote sensing image data include: for photogrammetric images, the advantages are edge detail features, texture clarity features, and high pixel density features; for scanned images, the advantages are large-area continuous coverage features, fast imaging efficiency features, and multi-band compatibility features; for radar images, the advantages are microwave penetration features, echo amplitude features, polarization response features, and day / night cloud / rain adaptability features; and for spectral images, the advantages are multi-band spectral curve features, ground object spectral difference features, and quantitative inversion support features.

[0070] In this embodiment, the preprocessing of image advantage feature data includes format standardization, geometric registration, and invalid pixel removal. Format standardization converts image advantage feature data from different sources into a unified data format (GeoTIFF format supporting geographic coordinate embedding, and BigTIFF extended format for large images) and metadata specifications to ensure data structure consistency. Geometric registration eliminates geometric deviations of multi-source images by establishing a mathematical transformation model, thereby achieving sub-pixel-level spatial alignment. Invalid pixel removal removes abnormal pixels caused by sensor noise and atmospheric interference. Thus, preprocessing standardizes, improves the precision of, and eliminates noise in advantage feature data of different imaging types, ensuring that the preprocessed data can directly support subsequent super-resolution processing and feature fusion.

[0071] In a specific embodiment of the present invention, super-resolution technology is used to magnify the preprocessed image to improve its spatial resolution, resulting in a first image set. Based on auxiliary data, a supplementary model for residual details is constructed. The supplementary model is trained through machine learning and supplements the missing details in the first image set, resulting in a second image set. Thus, this step is a key transformation link to realize the transformation from low-resolution advantageous features to high-resolution complete features. By improving accuracy through super-resolution magnification and combining it with the strategy of supplementing details through the residual model, the problem of insufficient spatial resolution of the preprocessed image is solved, and the geographical details that may be missing after magnification (such as terrain undulations, edge textures of ground features, etc.) are filled in, providing high-resolution and high-complete basic data for subsequent fusion.

[0072] First, deep learning-based super-resolution algorithms (such as ESPCN, SRGAN, and EDSR) are prioritized, with the magnification factor adaptively adjusted according to the original resolution of the preprocessed image, typically 2-4 times (e.g., magnifying a 30-meter resolution scanning image to 15-7.5 meters), thereby improving spatial resolution. Then, based on the scene correlation of auxiliary data, a supplementary model for residual details is constructed. The supplementary model is trained through machine learning and uses residual calculation to fill in the detail gaps after super-resolution magnification (such as shadow details caused by changes in terrain slope, and weakening of ground texture due to atmospheric conditions).

[0073] Specifically, a random forest algorithm can be used, with slope, aspect, and elevation variation coefficient from digital elevation model data; aerosol optical thickness, atmospheric transmittance, and water vapor content from atmospheric profile parameters; land cover type and land cover coverage from land cover classification data; and light intensity and surface temperature from meteorological data as input features. Historically stored high-resolution images of the same region (such as 0.5-meter UAV images, WorldView satellite images, etc.) are used as samples to extract detailed textures (such as vegetation leaf texture, road cracks, and terrain undulations) as output labels to build a supplementary model. This model is trained by computer learning. When supplementing missing details in the first image set, the residual is calculated by comparing the output of the supplementary model with the pixel difference in the first image set. The residual is then superimposed (with superposition weights adjusted according to land cover type) onto the first image set to achieve detail supplementation, resulting in a second image set.

[0074] In a specific embodiment of the present invention, the image of the remote sensing image data is set as the base image, and the base image is fused with a second image set to obtain an enhanced image set;

[0075] Specifically, the fusion of the base image and the second image set includes the following steps:

[0076] A first enhanced image is obtained by fusing a second image set into a spectral imaging image as the first base image and using a wavelet transform fusion algorithm. A second enhanced image is obtained by fusing a second image set into a scanning imaging image as the second base image and using a wavelet transform fusion algorithm. A third enhanced image is obtained by fusing a second image set into a radar imaging image as the third base image and using a wavelet transform fusion algorithm. A fourth enhanced image is obtained by fusing a second image set into a photographic imaging image as the fourth base image and using a wavelet transform fusion algorithm.

[0077] In this embodiment, the core function of the wavelet transform fusion algorithm is to accurately integrate the high-resolution details and residual supplementary information of the second image set while preserving the core advantageous features of the base image (spectral, scanning, radar, and photographic imaging). This achieves a fusion effect where advantageous features are not lost and supplementary details are fully absorbed. Specifically, through multi-scale decomposition, the base image and the second image set are split into low-frequency contour components (core features) and high-frequency detail components (edges and textures). Targeted fusion is then performed, preserving the core advantages of the base image (such as spectrum and penetration) in the low frequencies, while integrating the details and residual information of the second image set in the high frequencies. Inverse transform reconstruction is then performed, and the fused components are inversely processed to generate an enhanced image that combines advantageous features with detail accuracy.

[0078] An enhanced image set is obtained by using the first enhanced image, the second enhanced image, the third enhanced image, and the fourth enhanced image.

[0079] In a specific embodiment of the present invention, the enhanced image set is divided into regions, and the image quality of the same region in the divided enhanced image set is compared. High-quality image regions are selected and retained, while low-quality image regions that are blurred or have lost geographical features are removed. The selected high-quality image regions are then re-fused and correction processing is performed to obtain a preliminary enhanced remote sensing image.

[0080] Specifically, enhancing the region segmentation of the image set includes the following steps:

[0081] The enhanced image set is divided into regions of size NxN. Taking 3x3 as an example, the regions of the first enhanced image are labeled as n11, n12, n13, n14...n19 in sequence.

[0082] The regions of the second enhanced image are sequentially labeled as n21, n22, n23, n24…n29;

[0083] The regions of the third enhanced image are sequentially labeled as n31, n32, n33, n34…n39;

[0084] The regions of the fourth enhanced image are sequentially labeled as n41, n42, n43, n44…n49;

[0085] Enhancing the contrast and fusion of image sets includes the following steps:

[0086] Image quality is compared within the same partitioned region, such as n11, n21, n31 and n41, n12, n22, n32, n42, n13, n23, n33 and n43. High-quality image regions within the same partitioned region are selected and retained, while low-quality image regions that are blurred or have lost geographic features are removed, resulting in nX1, nX2, nX3...nX9, where X represents the corresponding image in the enhanced image set. nX1, nX2, nX3...nX9 are then re-fused using a stitching fusion algorithm and a weighted fusion algorithm to obtain the target enhanced image.

[0087] Subsequently, the enhanced image of the target can be corrected, including color adjustment, geometric correction, radiometric correction, and noise reduction.

[0088] Color adjustment is based on a standard spectral database of natural resources and features, combined with scene lighting conditions, to adjust color balance and saturation to ensure color consistency in different areas. Geometric correction is based on a digital elevation model to eliminate geometric distortion and projection deviations generated during imaging, as well as deformations caused by sensor attitude or terrain. Radiometric correction corrects the radiometric brightness deviation of the fused image through a radiometric calibration model. Noise processing adopts a multi-dimensional noise reduction strategy, including one or more of spatial domain filtering, frequency domain filtering, and adaptive noise reduction algorithms, to obtain a preliminary enhanced remote sensing image.

[0089] In this embodiment, taking the adaptive noise reduction algorithm as an example, the algorithm intelligently identifies local features of the image, such as edge regions, smooth regions, and noise-dense regions, and adjusts the noise reduction intensity accordingly. Specifically, the algorithm delineates a small local window and analyzes image features window by window, including judging the grayscale difference of pixels within the window (whether it is a smooth area), texture complexity (whether it is an edge or detail area), and noise intensity (density of noise). Based on the local features, the algorithm outputs a corresponding noise reduction strategy. In smooth regions (noise concentration areas), the noise reduction intensity is enhanced by using strong filtering such as mean and median. In edge or texture areas (with large grayscale differences), the noise reduction intensity is weakened or even not filtered to avoid blurring of details. In noise-sparse areas, a light weighted filtering is used to balance noise reduction and detail. In addition, the noise reduction result of each pixel can be fused with the feature weights of the surrounding windows to ensure that noise is effectively suppressed and that the processed pixels blend naturally with the surrounding environment, avoiding artificial traces (such as color blocks and blurred edges).

[0090] Based on this, the present invention integrates super-resolution technology, residual detail supplementation and multi-dimensional error correction to achieve comprehensive optimization of image quality. This results in high standards for the final output image in terms of spatial resolution compliance rate, detail texture restoration rate and ground feature recognition accuracy. It can accurately support refined natural resource management scenarios such as farmland rights confirmation, mineral monitoring and ground feature classification.

[0091] In a specific embodiment of the present invention, the quality of the preliminarily enhanced remote sensing image is verified, and the quality verification includes the following steps:

[0092] The preset resolution quantitative indicators include spatial resolution compliance rate, detail texture recovery rate, and ground feature recognition accuracy. These quantitative indicators can be set according to industry standards or historical experience. Specifically, the spatial resolution compliance rate is ≥90%, the detail texture recovery rate is ≥88%, and the ground feature recognition accuracy is ≥95%.

[0093] The accuracy of feature identification is achieved through automatic detection using a deep learning model. Based on labeled feature samples, the proportion of features correctly identified by the model out of the total number of samples is statistically analyzed. Natural resource surveys, land spatial planning, and other operations have gradually become reliant on automated feature identification. An accuracy rate of 95% can significantly reduce the workload of manual review. If the accuracy rate is below 90%, the manual correction rate will exceed 30%, rendering the automated process meaningless. When the accuracy rate is ≥95%, the manual review rate can be controlled within 10%, balancing efficiency and accuracy. Therefore, the 95% setting not only meets the upper limit of technical capabilities but also reserves a reasonable error margin for different feature types (such as sparse vegetation and small water bodies), avoiding the overall index being lowered by a few niche feature types.

[0094] The detail texture restoration rate is achieved by comparing the texture features of the preliminarily enhanced remote sensing image with the high-resolution reference image. The contrast and pixel grayscale indices are extracted using the grayscale co-occurrence matrix, and the similarity ratio of contrast and pixel grayscale is calculated. Dynamic monitoring of natural resources (such as changes in vegetation growth, land occupation of small projects, and monitoring of minor topographic relief) relies on the detailed texture of ground features. When the restoration rate is ≥88%, it can effectively identify subtle features such as vegetation leaf texture, road cracks, and changes in terrain slope. If it is below 85%, texture blurring will occur (such as confusion between shrubs and grasslands, and omission of minor mining points), thus failing to meet the needs of refined management. The residual detail supplementation model is based on the random forest algorithm and combined with auxiliary data, which can accurately predict 85%-90% of the detail gap. The 88% setting not only fully utilizes the model's capabilities but also takes into account the limitations of texture restoration in extreme scenarios (such as repair after thick cloud cover and complex terrain areas), avoiding the overall failure to meet the standard due to individual special areas.

[0095] The spatial resolution compliance rate is calculated by measuring known-sized features in preliminarily enhanced remote sensing images. In natural resource management, scenarios such as farmland rights confirmation, nature reserve boundary delineation, and mineral mining area monitoring require that the spatial scale of more than 90% of the image area be accurate. If the compliance rate is less than 90%, there will be insufficient resolution in some areas (such as blurred boundaries of some fields, omission of small features, etc.), which will lead to distortion of subsequent spatial analysis results and affect the accuracy of management decisions. In addition, the spatial resolution qualification line of mainstream international satellite imagery (such as WorldView-3 and Landsat-9) is ≥85%. The method of this invention has more significant technical advantages because it integrates multi-source auxiliary data and other technologies. Therefore, the standard is raised by 5 percentage points, which not only reflects the technological advancement but also conforms to industry understanding.

[0096] If any single indicator such as spatial resolution compliance rate, detail texture restoration rate, or ground feature recognition accuracy fails to meet the standard, it is determined that the verification result has not reached the preset clarity. The extraction range of advantageous features is then optimized again based on the feature quality evaluation matrix, and steps three through six are repeated until the preset clarity is achieved, resulting in a natural resource remote sensing image with improved quality.

[0097] Therefore, this invention, by encompassing a variety of images and taking into account the advantages of multi-band resolution, wide coverage, all-weather observation and high detail restoration, adapts to different natural resource scenarios through multi-dimensional differential analysis and incorporates quality verification and iterative optimization mechanisms to ensure that high-quality images can still be stably output under extreme conditions such as complex terrain and severe weather, thus meeting diverse needs such as natural resource surveying, dynamic monitoring and boundary delineation.

[0098] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for improving the quality of natural resource remote sensing image data, characterized in that, The method comprises the following steps: S1, obtaining remote sensing image data and auxiliary data of a natural resource scene, wherein the remote sensing image data comprises spectral imaging images, scanning imaging images, radar imaging images and photographic imaging images, and the auxiliary data comprises digital elevation model data, atmospheric profile parameter data, land cover classification data and meteorological data; S2, performing multi-dimensional differential comparative analysis on the images of the remote sensing image data, and constructing a feature advantage and disadvantage evaluation matrix to determine the image advantage features and performance short boards of the remote sensing image data; S3, based on the feature advantage and disadvantage evaluation matrix, extracting image advantage feature data of the remote sensing image data, and pre-processing the image advantage feature data to obtain pre-processed images; S4, using super-resolution technology to enlarge the pre-processed images to improve the spatial resolution of the pre-processed images, obtaining a first image set, and based on the auxiliary data, constructing a residual detail supplement model, wherein the supplement model is trained by machine learning, the supplement model supplements the missing details in the first image set to obtain a second image set; S5, setting the images of the remote sensing image data as base images, fusing the base images with the second image set to obtain an enhanced image set; S6, dividing the enhanced image set into regions, comparing the quality of the same region of the divided enhanced image set, screening and retaining high-quality image regions, and eliminating low-quality image regions with blurred or missing geographical features, re-fusing the screened high-quality image regions, and performing correction processing to obtain a preliminary enhanced remote sensing image; S7, verifying the quality of the preliminary enhanced remote sensing image, if the verification result does not reach a preset definition, re-optimizing the extraction range of the advantage features based on the feature advantage and disadvantage evaluation matrix, and repeating steps S3-S6 until the preset definition is reached, to obtain a natural resource remote sensing image with improved quality.

2. The method of claim 1, wherein, The multi-dimensional differential comparative analysis comprises radiation brightness accuracy, spatial detail restoration degree, environmental adaptability, spectral resolution capability, spatial positioning accuracy and data redundancy, and the adaptation differences of the images of the remote sensing image data in the natural resource scene are compared through six dimensions to determine the image advantage features and performance short boards of the remote sensing image data, and a feature advantage and disadvantage evaluation matrix is established.

3. The method of claim 1, wherein the method further comprises: The image advantage features of the remote sensing image data comprise: The image advantage features of the photographic imaging images are edge detail features, texture definition features and high pixel density features, the image advantage features of the scanning imaging images are large-area continuous coverage features, fast imaging efficiency features and multi-band compatibility features, the image advantage features of the radar imaging images are microwave penetration features, echo amplitude features, polarization response features and day-night cloud and rain adaptability features, and the image advantage features of the spectral imaging images are multi-band spectral curve features, ground object spectral difference features and quantitative inversion support features.

4. The method of claim 1, wherein, The preprocessing of the image advantage feature data includes format standardization, geometric registration and invalid pixel removal, wherein the format standardization is to convert image advantage feature data of different sources into a unified data format and metadata specification, the geometric registration is to eliminate geometric deviations of multi-source images by establishing a mathematical transformation model to realize sub-pixel level spatial position alignment, and the invalid pixel removal is to remove abnormal pixels caused by sensor noise and atmospheric interference.

5. The method of claim 1, wherein, The acquisition of the remote sensing image data and auxiliary data includes the following steps: The spectral imaging type image is generated by a hyperspectral imaging satellite and using a light splitting system to decompose ground object radiation into multiple narrow wave bands, and by a multi-channel detector to synchronously receive signals; The scanning imaging type image is generated by a push-broom satellite and using a sensor to scan ground objects point by point or line by line, to receive electromagnetic radiation converted into electrical signals for further processing; The radar imaging type image is generated by a synthetic aperture radar device and by actively emitting microwave signals to receive ground object echo signals, and by amplitude and phase analysis inversion; The photographic imaging type image is generated by an aerial photogrammetry device and by using an optical lens to capture ground object reflected light, and by a photosensitive medium or a digital sensor to complete photoelectric conversion; The digital elevation model data is based on the geographical boundary and imaging time of the remote sensing image data, and uses 30-meter resolution SRTM data or 10-meter resolution ASTER GDEM data, the atmospheric profile parameter data is centered on the imaging time of the remote sensing image data, and real-time data within one hour before and after the imaging time is obtained from a global atmospheric observation network, the land cover classification data uses the latest version of GlobeLand30 data, and the weather data is obtained from the nearest weather station in the target area within the daily hourly data.

6. The method of claim 1, wherein the method further comprises: The construction of the supplementary model includes the following steps: The random forest algorithm is used to construct the model, the slope, slope direction and elevation variation coefficient of the digital elevation model data, the aerosol optical depth, atmospheric transmittance and water vapor content of the atmospheric profile parameter data, the ground object type and ground object coverage of the land cover classification data, and the illumination intensity and ground surface temperature of the weather data are used as input features, the detailed texture in the historical retained high-resolution image is used as output label, the supplementary model is established, and the supplementary model is trained by computer learning; When supplementing the missing details in the first image set, the residual error is calculated by the difference between the output result of the supplementary model and the pixels of the first image set, the residual error is added to the first image set to realize detail supplement, and the second image set is obtained.

7. The method of claim 1, wherein the method further comprises: The fusion of the basic image and the second image set includes the following steps: The spectral imaging type image is set as the first basic image, the second image set is fused into the first basic image by using the wavelet transform fusion algorithm, and the first enhanced image is obtained; The scanning imaging type image is set as the second basic image, the second image set is fused into the second basic image by using the wavelet transform fusion algorithm, and the second enhanced image is obtained; The radar imaging type image is set as a third base image, the second image set is fused into the third base image by using a wavelet transform fusion algorithm, and a third enhanced image is obtained; The photographic imaging type image is set as a fourth base image, the second image set is fused into the fourth base image by using a wavelet transform fusion algorithm, and a fourth enhanced image is obtained; The first enhanced image, the second enhanced image, the third enhanced image and the fourth enhanced image are used to obtain the enhanced image set.

8. The method of claim 7, wherein the method further comprises: The region division, contrast and fusion of the enhanced image set include the following steps: The enhanced image set is divided into region graphs with a size of NxN, and the divided regions of the first enhanced image are sequentially marked as n11, n12, n13, n14...n19 by taking 3x3 as an example; The divided regions of the second enhanced image are sequentially marked as n21, n22, n23, n24...n29; The divided regions of the third enhanced image are sequentially marked as n31, n32, n33, n34...n39; The divided regions of the fourth enhanced image are sequentially marked as n41, n42, n43, n44...n49; The same divided regions are compared in terms of image quality, such as n11, n21, n31 and n41, n12, n22, n32, n42, n13, n23, n33 and n43, high-quality image regions in the same divided region are selected and retained, and low-quality image regions with blurring or missing geographical features are removed, to obtain nX1, nX2, nX3...nX9 respectively, wherein X represents the corresponding image in the enhanced image set, the nX1, nX2, nX3...nX9 are re-fused by using a splicing fusion algorithm and a weighted fusion algorithm, and a target enhanced image is obtained.

9. The method of claim 8, wherein the method further comprises: The correction processing of the target enhanced image includes color adjustment, geometric correction, radiation correction and noise processing; The color adjustment adjusts color balance and saturation according to a standard spectral database of natural resource ground objects and in combination with scene lighting conditions, to ensure color consistency in different regions, the geometric correction eliminates geometric distortion and projection deviation generated in the imaging process based on a digital elevation model, and eliminates deformation caused by sensor posture or terrain, the radiation correction corrects radiation brightness deviation of the fused image through a radiation calibration model, and the noise processing adopts a multi-dimensional noise reduction strategy, including one or more of spatial domain filtering, frequency domain filtering and an adaptive noise reduction algorithm, to obtain a preliminary enhanced remote sensing image.

10. The method of claim 1, wherein the method further comprises: The quality verification of the preliminary enhanced remote sensing image includes the following steps: A quantization index of definition is preset, and the quantization index includes a spatial resolution compliance rate, a detail texture recovery rate and a ground object identification accuracy rate, wherein the spatial resolution compliance rate is greater than or equal to 90%, the detail texture recovery rate is greater than or equal to 88%, and the ground object identification accuracy rate is greater than or equal to 95%; The ground object identification accuracy rate is detected automatically by using a deep learning model, and the proportion of the number of ground objects correctly identified by the model to the total number of samples is counted based on the labeled ground object samples. The detail texture recovery rate is calculated by comparing the texture features of the preliminary enhanced remote sensing image and a high-resolution reference image, using a gray level co-occurrence matrix to extract contrast and pixel gray level indicators, and calculating the similarity proportion of the contrast and the pixel gray level; The spatial resolution compliance rate is calculated by a known size of a ground object in the preliminary enhanced remote sensing image; If the single indicators of the spatial resolution compliance rate, the detail texture recovery rate and the ground object recognition accuracy rate do not meet the standards, it is determined that the verification result does not reach the preset definition.

Images