A method for automatic color correction of massive remote sensing images of sub-meter satellites
By using an improved ResNet50 model and histogram matching method, the problem of acquiring pairing data for massive sub-meter-level satellite remote sensing images was solved, achieving automated color correction and improving the quality and efficiency of the final image output. This makes the image suitable for applications such as precision agriculture and disaster assessment.
Patent Information
- Application Number
- CN202510582093.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-07
- Publication Date
- 2026-05-15
- Estimated Expiration
- 2045-05-07
AI Technical Summary
Existing technologies struggle to acquire paired data from massive amounts of sub-meter resolution remote sensing images, and in areas with complex land cover, a uniform correction strategy cannot effectively handle different land cover types, leading to inconsistent or distorted color correction.
By training an improved ResNet50 model and combining it with histogram matching methods, massive amounts of sub-meter high-resolution paired data are obtained. The single spatial dimension output of the fully connected layer is aggregated into the probability distribution of R, G and B channels to achieve automated color correction.
It improves the robustness and generalization ability of the model, enabling automated correction of remote sensing images without reference, improving the efficiency and quality of image delivery, reducing labor costs, and making it suitable for fields such as precision agriculture management and disaster assessment.
Smart Images
Figure CN120411262B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of automated color correction technology for massive sub-meter-level satellite remote sensing images, specifically to a method for automated color correction of massive sub-meter-level satellite remote sensing images. Background Technology
[0002] Sub-meter resolution satellite remote sensing imagery products have significant application value in mapping, disaster and emergency management, and precision agriculture and forestry. Color issues in finished remote sensing images are often the result of a combination of factors.
[0003] Furthermore, the reflective properties of the Earth's surface itself also affect color accuracy. Different land features (such as vegetation, water bodies, and urban buildings) reflect different wavelengths of light based on their inherent physical and chemical properties; this difference is reflected in remote sensing imagery as color variations. Without proper correction and calibration, these natural color variations may be misinterpreted or misrepresented. The selection and application of color correction algorithms are also crucial in processing remote sensing data. Inappropriate image processing techniques, such as radiometric correction anomalies, can further distort the original color information, especially when processing large amounts of data automatically. Algorithms not optimized for specific scenes or conditions may lead to inconsistent color output.
[0004] In remote sensing image processing, radiometric correction is a fundamental and crucial step. Its purpose is to eliminate the effects caused by sensor systematic errors, atmospheric interference, and the inherent characteristics of the Earth's surface, thereby making the image closer to the spectral reflectance of the real surface. However, the radiometric correction process itself may introduce new color problems, especially when the correction method is not precise enough or the original data itself has quality issues.
[0005] Radiometric correction typically requires adjusting image data based on the sensor's radiometric response function. If these parameters (such as sensor gain and offset) are not accurately measured or estimated, true colors in the image may be incorrectly enhanced or suppressed. For example, in the visible light band, over-correction of the red band can result in a reddish visual effect, neglecting the balance of other colors. A crucial aspect of radiometric correction is atmospheric correction, which removes the influence of the atmosphere on light propagation. Different atmospheric correction models are suitable for different atmospheric conditions, such as clear skies, cloudy skies, or fog. Incorrectly selecting an atmospheric correction model can lead to color distortion in the corrected data. For instance, correcting a foggy image using a clear skies model may over-enhance certain bands, resulting in unnatural colors. The diversity of surface materials and conditions requires radiometric correction to adapt to the characteristics of different land cover types. For areas with complex surface cover, such as areas with a mixture of urban and natural vegetation, a uniform correction strategy may not effectively handle all land cover types, leading to over-correction or under-correction of some land cover colors.
[0006] Color correction methods for remote sensing images can be mainly divided into two categories: unsupervised methods and supervised methods. Unsupervised color correction methods are independent of any external standard image and rely on processing the internal information of the raw data.
[0007] Supervised color correction methods rely on known reference imagery or ground-based data to correct the remote sensing imagery to be processed. Histogram matching techniques adjust the histogram of the image to be corrected to match the histogram of a selected reference image. Figure 1 This method achieves color correction by matching images, but it can only be performed one-to-one with a reference image. With the development of artificial intelligence, deep learning-based image color correction methods have become popular. These methods utilize large amounts of training data to train neural networks, automatically learning the complex mapping relationship between imagery and real-world ground conditions, demonstrating powerful capabilities in handling complex and heterogeneous data. However, for color correction of massive amounts of sub-meter resolution satellite imagery, obtaining a large number of high-resolution paired datasets remains one of the unsolved problems.
[0008] Therefore, automated color correction of massive amounts of sub-meter resolution remote sensing imagery still faces many challenges. One major challenge is acquiring massive amounts of high-resolution sub-meter resolution matching data, covering diverse surface types, different times, and lighting conditions. Furthermore, developing efficient algorithms is crucial for processing massive amounts of remote sensing data, and enhancing the robustness and generalization ability of these algorithms is also essential. Summary of the Invention
[0009] To address the problem that existing color correction methods struggle to acquire massive amounts of sub-meter resolution paired data, this invention provides an automated color correction method for massive amounts of sub-meter resolution satellite remote sensing images. By training an improved ResNet50 model, massive amounts of sub-meter resolution paired data are acquired, and then combined with a histogram matching method, automated color correction of massive amounts of sub-meter resolution satellite remote sensing images is achieved.
[0010] The method includes the following steps:
[0011] S1. Obtain N pre-correction sub-meter level remote sensing images, and perform histogram matching on the pre-correction sub-meter level remote sensing images to obtain the RGB histograms of the N pre-correction sub-meter level remote sensing images.
[0012] S2. Preprocess the N pre-corrected sub-meter level remote sensing images and their corresponding RGB histograms to obtain M pairs of map sheet data sets.
[0013] S3. Divide the M-pair map patch dataset into a training set and a validation set in an 8:2 ratio;
[0014] S4. Construct and improve the ResNet50 model to obtain the improved ResNet50 model:
[0015] S41. Add a fully connected module after the fully connected layer of the ResNet50 model;
[0016] S5. Train the improved ResNet50 model using the training and validation sets. When the convergence of the loss function tends to be stable, the training is complete, and the trained improved ResNet50 model is obtained.
[0017] S6. Correct sub-meter level remote sensing images using the improved ResNet50 model after training.
[0018] Furthermore, the preprocessing specifically involves cropping the image to a size of 224×224 pixels.
[0019] Furthermore, the fully connected module passes through the first, second, and third fully connected layers in parallel from input to output, and all three layers use the Sofmax activation function.
[0020] Furthermore, the first, second, and third fully connected layers are used to predict the probability distributions of the R, G, and B channels based on their spatial dimensions, respectively.
[0021] Furthermore, the specific steps for correcting sub-meter level remote sensing images using the trained ResNet50 improved model are as follows:
[0022] S51. Input the sub-meter level remote sensing image to be corrected into the trained ResNet50 improved model to obtain the RGB channel probability distribution of the sub-meter level remote sensing image after correction.
[0023] S52. Perform histogram matching between the RGB channel probability distribution of the corrected sub-meter level remote sensing image and the sub-meter level remote sensing image to be corrected to obtain the corrected sub-meter level remote sensing image.
[0024] Furthermore, the processor implements the steps of the above method when executing the computer program.
[0025] Furthermore, when the computer instructions are executed by the processor, they implement the steps of the above method.
[0026] The beneficial effects of this invention are as follows:
[0027] (1) In this invention, after the fully connected layer of the existing ResNet50 model, a fully connected module (the first fully connected layer, the second fully connected layer and the third fully connected layer) is introduced in parallel. The single spatial dimension aggregated value output by the fully connected layer is divided into the probability distribution of R, G and B channels, so that the RGB channel distribution information can be extracted, which improves the robustness and generalization ability of the ResNet50 basic model.
[0028] (2) This invention first obtains a massive, high-precision sub-meter level remote sensing image pairing dataset through histogram matching and cropping methods, covering a wide range of land surface types, different times and lighting conditions. The trained ResNet50 improved model can automatically correct remote sensing images without reference, and obtain the RGB channel probability distribution of the corrected sub-meter level remote sensing images.
[0029] (3) By combining histogram matching and the improved ResNet50 model, this invention can perform automated color correction on massive sub-meter level satellite remote sensing images, improve the efficiency and quality of remote sensing image delivery, and greatly reduce labor costs. It has important applications in fields such as precision agriculture management, disaster assessment, and environmental monitoring. Attached Figure Description
[0030] Figure 1 This is a schematic diagram of the improved ResNet50 model structure described in an embodiment of the present invention;
[0031] Figure 2 This is a schematic diagram of the Bottleneck module structure according to an embodiment of the present invention;
[0032] Figure 3 This is a flowchart of the method described in an embodiment of the present invention;
[0033] Figure 4 (a) is the Landsat 9 surface reflectance image according to the embodiment of the present invention; (b) is the image before color correction; (c) is the image after color correction. Detailed Implementation
[0034] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0035] This embodiment provides an automated color correction method for massive remote sensing images from sub-meter-resolution satellites. The flowchart of the method is as follows: Figure 3As shown, this embodiment relies on the "Jilin-1" series of sub-meter resolution remote sensing satellite images to acquire a large number of sub-meter resolution remote sensing image data sources. This embodiment collects 100 pre-correction "Jilin-1" high-resolution series sub-meter resolution remote sensing images, covering different landforms and features globally, thereby improving the model's generalization ability. To obtain the true surface reflectance of land features, this embodiment uses the official Landsat 8 / 9 surface reflectance data as a reference (data source: Google Earth Engine platform). Using a programming method, it downloads the corresponding regional data from the "Jilin-1" remote sensing images, obtaining 100 one-to-one corresponding 30m resolution surface reflectance images. Histogram matching is performed between the 100 pre-correction "Jilin-1" high-resolution series sub-meter resolution remote sensing images and the 100 one-to-one corresponding 30m resolution surface reflectance images. The colors of the pre-correction "Jilin-1" images are mapped to the surface reflectance colors, thus obtaining the RGB histograms of the 100 pre-correction sub-meter resolution remote sensing images.
[0036] Subsequently, the 100 uncorrected sub-meter-level remote sensing images and their corresponding RGB histograms were cropped into multiple map sheets of 224×224 pixels each, resulting in approximately 30,000 pairs of map sheets. Each pair of map sheets can be divided into pre-matching and post-matching sections.
[0037] The 30,000 map sheets were divided into a training set and a validation set in an 8:2 ratio.
[0038] Construct and improve the ResNet50 model to obtain the improved ResNet50 model (deep neural network model):
[0039] like Figure 1 As shown, a fully connected module is added after the fully connected layer of the ResNet50 model;
[0040] The fully connected module proceeds in parallel from input to output through the first, second, and third fully connected layers. All three layers utilize the Sofmax activation function. The first, second, and third fully connected layers predict the probability distributions of the R, G, and B channels based on their spatial dimensions, respectively.
[0041] The fully connected module divides the single spatial dimension aggregated value output by the fully connected layer into probability distributions of R, G, and B channels, thereby extracting RGB channel distribution information and improving the robustness and generalization ability of the ResNet50 base model.
[0042] The improved ResNet-50 model begins with an initial processing layer on the input image, consisting of a large 7×7 convolutional kernel (ConV7×7) to quickly extract low-level features and expand the receptive field. This is followed by a 3×3 max-pooling layer (MaXPool2d3×3) to further reduce the spatial resolution of the feature maps. Through this combination, the network effectively reduces computational cost while preserving key visual information.
[0043] The main body of the network consists of four stages, corresponding to the four convolutional parts conv2_x, conv3_x, conv4_x and conv5_x respectively. The number of residual modules in each convolutional part is 3, 4, 6 and 3 respectively, for a total of 16 Bottleneck residual blocks.
[0044] like Figure 2 As shown, each Bottleneck structure consists of three convolutional layers: first, a 1×1 convolution is used for channel compression, followed by a 3×3 convolution for spatial feature extraction, and finally, a 1×1 convolution is used to restore the number of channels to the original dimension. This "compression-convolution-expansion" structural design can significantly reduce the computational cost and the number of parameters while maintaining expressive power.
[0045] In each residual block, the input features, while undergoing three convolutional layers in the main path, are also directly passed to the output via a bypass path and added to the output of the main path. This direct cross-layer connection effectively alleviates the gradient vanishing and performance degradation problems that occur in deep networks during training, allowing the network depth to be further increased without compromising learning ability.
[0046] Furthermore, each convolutional operation is followed by batch normalization (BN) and ReLU activation functions to improve training stability and non-linear modeling capabilities. After all residual modules are completed, the improved ResNet-50 model aggregates the spatial dimension of each channel into a single value through a global average pooling layer and a fully connected layer. Then, through a fully connected module, the aggregated value of the single spatial dimension is divided into probability distributions for R, G, and B channels.
[0047] The improved ResNet50 model was trained using the training and validation sets. The learning rate was set to 5e-5, the weight decay was set to 5e-4, the batch size was set to 256, and the training iterations were performed 200 times across the entire dataset. The training was completed when the convergence of the loss function tended to be stable, and the trained improved ResNet50 model was obtained.
[0048] like Figure 3 As shown, the sub-meter level remote sensing image is corrected using the trained ResNet50 improved model.
[0049] S51. Input the sub-meter level remote sensing image to be corrected into the trained ResNet50 improved model to obtain the RGB channel probability distribution of the sub-meter level remote sensing image after correction.
[0050] S52. Perform histogram matching between the RGB channel probability distribution of the corrected sub-meter level remote sensing image and the sub-meter level remote sensing image to be corrected to obtain the corrected sub-meter level remote sensing image.
[0051] This embodiment tested the method described in this invention in a selected area, and the tested images were sourced from the "Jilin-1" satellite data source. The area is taken as an example: the region near Beijing Capital International Airport. Figure 4 As shown, (a) represents the Landsat 9 surface reflectance image, (b) represents the image before correction (original images from the "Jilin-1" series), and (c) represents the image after correction (color-corrected "Jilin-1" satellite image). Here, comparing the colors of the airport and trees, the airport appears yellowish in the image before correction, while the trees are darker. After color correction, these problems can be effectively solved. From the above experimental results, it can be seen that the method described in this invention can effectively perform color correction on sub-meter level remote sensing satellite images and achieve ideal results, reduce manual delineation costs, and improve the efficiency and quality of image delivery.
Claims
1. An automated color correction method for massive remote sensing images from sub-meter-level satellites, characterized in that, The method includes the following steps: S1. Obtain N pre-correction sub-meter level remote sensing images, and perform histogram matching on the pre-correction sub-meter level remote sensing images to obtain the RGB histograms of the N pre-correction sub-meter level remote sensing images. S2. Preprocess the N pre-corrected sub-meter level remote sensing images and their corresponding RGB histograms to obtain M pairs of map sheet data sets. S3. Divide the M-pair map patch dataset into a training set and a validation set in an 8:2 ratio; S4. Construct and improve the ResNet50 model to obtain the improved ResNet50 model: S41. Add a fully connected module after the fully connected layer of the ResNet50 model; The fully connected module passes through the first, second, and third fully connected layers in parallel from input to output. The first, second, and third fully connected layers all use the Sofmax activation function. The first, second, and third fully connected layers are used to predict the probability distributions of the R, G, and B channels based on their spatial dimensions, respectively. S5. Train the improved ResNet50 model using the training and validation sets. When the convergence of the loss function tends to be stable, the training is complete, and the trained improved ResNet50 model is obtained. S6. Correct sub-meter level remote sensing images by improving the trained ResNet50 model; The specific steps for correcting sub-meter level remote sensing images using the trained ResNet50 improved model are as follows: S61. Input the sub-meter level remote sensing image to be corrected into the trained ResNet50 improved model to obtain the RGB channel probability distribution of the sub-meter level remote sensing image after correction. S62. Perform histogram matching between the RGB channel probability distribution of the corrected sub-meter level remote sensing image and the sub-meter level remote sensing image to be corrected to obtain the corrected sub-meter level remote sensing image.
2. The automated color correction method for massive remote sensing images from sub-meter-level satellites according to claim 1, characterized in that, The preprocessing specifically involves: according to... The pixel size is cropped.
3. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-2.
4. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-2.