Remote sensing image cloud detection method with variable receptive field

Abstract

The invention relates to the technical field of remote sensing image processing, in particular to a variable receptive field remote sensing image cloud detection method, which comprises the following steps: acquiring an original remote sensing image, and performing boundary filling preprocessing on the original remote sensing image to generate a filling image; extracting all sub-blocks of the filling image based on a sliding window, carrying out cloud-existence classification on all the sub-blocks, and weighting a corresponding coverage window according to a classification result; mapping each sliding window detection result to a corresponding position in the original remote sensing image; and weighting results of all covered sliding windows of each pixel are accumulated, an accumulated weighting map is generated, threshold segmentation is further carried out, a binary cloud mask is obtained and post-processing is carried out, and a corresponding vector result is generated. According to the method, whether the image blocks have clouds or not is judged more effectively by adjusting the receptive field of the image through variable parameters, each small block in the image is classified step by step, and accurate recognition of the cloud layer area is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a remote sensing image cloud detection method with a variable receptive field. Background Art

[0002] Cloud detection technology is widely used in various remote sensing image analysis scenarios, especially in satellite remote sensing, drone imagery, and aerospace data processing. With the rapid increase in the amount of remote sensing image data, traditional cloud detection methods have been unable to meet the processing requirements of large-scale, high-resolution images. Specifically, existing optical remote sensing satellites obtain massive amounts of earth observation data every day, and the accurate and rapid identification of cloud ranges in remote sensing images is the key to high-quality reference of remote sensing data. However, the current cloud area recognition technology for optical images still has the following shortcomings: 1) False detection and missed detection problems in complex backgrounds: Traditional cloud detection methods usually rely on threshold-based or simple classification techniques, which are easily affected by the background environment, especially when there is spectral similarity between clouds and the ground. For example, bright ground objects such as snow, sand dunes, saline-alkali land, and urban buildings often have similar spectral characteristics to clouds, resulting in confusion between clouds and ground objects and inability to accurately distinguish them.

[0003] 2) Low detection accuracy for small-scale clouds: Traditional cloud detection methods often have difficulty accurately identifying small-scale, thin clouds or clouds with unclear boundaries, and are prone to miss detection or produce large errors. Many threshold-based methods cannot make accurate judgments in these complex situations.

[0004] 3) Lack of flexible granularity adjustment for cloud detection: Most traditional cloud detection methods usually do not consider the different requirements for cloud granularity in different application scenarios during the detection process. For example, some applications require rapid detection of large cloud areas, while other scenarios require high-precision judgment of clouds in small areas. Traditional methods often lack a flexible granularity adjustment mechanism and cannot provide efficient and accurate detection in different scenarios.

[0005] 4) Low efficiency in processing large-scale images: Traditional cloud detection methods often require global analysis or pixel-by-pixel processing of the entire image, which is inefficient in large-scale remote sensing images, especially in high-resolution images. The processing time is long and it is difficult to meet the needs of real-time analysis.

[0006] Based on this, technical personnel in this field urgently need to propose a new remote sensing image cloud detection method to overcome the technical problems existing in the above-mentioned prior art. Summary of the invention

[0007] Therefore, the technical problem to be solved by the present invention is to overcome the defects existing in the above-mentioned prior art, thereby providing a remote sensing image cloud detection method with a variable receptive field.

[0008] A remote sensing image cloud detection method with a variable receptive field comprises the following steps: Acquire the original remote sensing image, and perform boundary filling preprocessing on the original remote sensing image to generate a filled image; Extract all sub-blocks of the filled image based on the sliding window, classify all sub-blocks for cloudiness, and weight the corresponding covered windows according to the classification results; Map each sliding window detection result back to the corresponding position in the original remote sensing image; accumulate the weighted results of all sliding windows covering each pixel to generate an accumulated weight map, and further perform threshold segmentation to obtain a binary cloud mask; Post-process the binary cloud mask to generate the corresponding vector result.

[0009] Preferably, all sub-blocks of the filled image are extracted based on the sliding window, and all sub-blocks are classified for cloudiness, specifically as follows: Use the pre-trained network model to perform binary classification on each sub-block, and the binary classification results are cloud or no cloud; Among them, the pre-trained network model adopts the ResNet50 network model.

[0010] Preferably, the corresponding coverage window is weighted according to the classification result, specifically: For the classification results with clouds, the corresponding sliding window result weight is +1; For the cloud-free classification results, the corresponding sliding window results are weighted by -1.

[0011] Preferably, the cumulative weight graph expression is: ; in, is the cumulative weight graph; is the number of windows that cover the pixel; is the weight of each sliding window (+1 or -1); The range of is: ; is the width of the sliding window, is the sliding step of the sliding window; the number of sliding windows is N, N is a positive integer and 1≤i≤N; i is the i-th sliding window.

[0012] Preferably, the binary cloud mask expression is generated as follows: ; In the formula, To preset the threshold, the default value of T is set to 0; is a binary cloud mask.

[0013] Preferably, it also includes constructing a training sample set of the ResNet50 network model, specifically: The training sample set adopts a binary classification data set, classifying bright ground objects as cloudless categories and bright clouds as cloud categories.

[0014] The technical solution of the present invention has the following advantages: The present invention adjusts the image receptive field through variable parameters to more effectively determine whether there are clouds in the image block, gradually classifies each small block in the image, realizes accurate identification of cloud areas, and thus improves the efficiency and accuracy of the cloud detection process to meet the needs of real-time, large-scale remote sensing data analysis.

[0015] The present invention can efficiently and accurately identify and separate cloud areas through the image window classification method with variable receptive field, providing reliable support for subsequent data cleaning, image restoration, etc. This makes the technology of the present invention more advantageous in the fields of geographic information system (GIS), environmental monitoring, weather forecasting, etc., especially in areas with large cloud influence, which can effectively improve the quality of images and analysis accuracy.

[0016] In addition, the cloud detection technology of the present invention is also of great significance in applications in multiple industries. In the field of agricultural monitoring, clouds in remote sensing images often interfere with the extraction of agricultural information such as crop health and soil moisture. This technology can provide accurate farmland information while efficiently removing clouds, providing a reliable basis for crop growth monitoring and yield prediction. Disaster monitoring is also one of the important application scenarios of this technology. When natural disasters such as floods and forest fires occur, clouds may obscure the ground conditions in key areas, affecting disaster assessment and emergency response efficiency. The cloud detection technology of the present invention can quickly locate and remove clouds to ensure the accuracy of post-disaster images, thereby improving the efficiency of disaster monitoring and emergency decision-making. BRIEF DESCRIPTION OF THE DRAWINGS In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 The present invention is a flow chart of a remote sensing image cloud detection method with a variable receptive field. DETAILED DESCRIPTION

[0018] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0020] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0021] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0022] Example 1 like Figure 1 A remote sensing image cloud detection method with a variable receptive field is shown, which includes four stages, specifically including the following steps: Data preprocessing stage: obtain the original remote sensing image, and perform boundary filling preprocessing on the original remote sensing image to generate a filled image; Sliding window detection stage: extract all sub-blocks of the filled image based on the sliding window, classify all sub-blocks into cloud or not, and weight the corresponding coverage window according to the classification results; Result weighting and fusion: Map each sliding window detection result back to the corresponding position in the original remote sensing image; accumulate the weighted results of all sliding windows covering each pixel to generate a cumulative weight map, and further perform threshold segmentation to obtain a binary cloud mask; Post-processing stage: Post-process the binary cloud mask to generate the corresponding vector result.

[0023] Specifically: Data preprocessing stage: obtain the original remote sensing image, and perform boundary filling preprocessing on the original remote sensing image to generate a filled image, which specifically includes the following two steps: 1-1 Remote sensing image input: Input high-resolution original remote sensing image 𝐼 and set the sliding window width parameter to W.

[0024] 1-2 Boundary filling: To ensure that the pixels in the boundary area are covered by the complete window and reduce the edge effect, the boundary of the original remote sensing image is filled based on the width W of the sliding window, and the filling width is W to form a filled image 𝐼′.

[0025] It should be noted that this embodiment can flexibly adjust the receptive field and detection result granularity of cloud detection. By adaptively setting the sliding window size and step size, the detection granularity can be dynamically adjusted according to specific needs to adapt to cloud detection tasks of different scales. In practical applications, whether it is necessary to accurately locate the cloud boundary or to quickly determine the cloud area over a large range, this embodiment can meet different application needs and provide a flexible solution.

[0026] Sliding window detection stage: Based on the sliding window, all sub-blocks of the filled image are extracted, all sub-blocks are classified as having clouds or not, and the corresponding coverage windows are weighted according to the classification results. Specifically, the following steps are included: 2-1 Window sliding: Set the sliding step size to L, that is, the size of the sliding window for detection is a square block of W*W. The slider slides to the right each time by a distance of L. After reaching the rightmost side of the image, slide down by a step size of L, and continue to repeat the above sliding process from the leftmost side of the image until it reaches the lower right corner of the image, extracting all the image sub-blocks of the image. It should be noted that the sliding window classification method of this embodiment greatly improves the processing efficiency by gradually processing each small block in the image. The classification and discrimination within each window is more concentrated, which avoids redundant calculations of the entire image and improves the processing speed of large-scale images. And by reasonably setting the window size and step size, the technical solution of this embodiment can significantly improve the processing efficiency while ensuring the detection accuracy.

[0027] 2-2 Sub-block classification: Use the pre-trained ResNet network model to classify each window sub-block Classification is performed, and the classification results are clouded or cloudless, and the corresponding coverage window is weighted according to the classification results.

[0028] In this embodiment, a pre-trained network model is used to classify each sub-block, and the classification result is cloud or no cloud; Among them, the pre-trained network model adopts the ResNet50 network model.

[0029] In addition, in this embodiment, the corresponding coverage window is weighted according to the classification result, specifically: For the classification results with clouds, the corresponding sliding window result weight is +1; For the cloud-free classification results, the corresponding sliding window results are weighted by -1.

[0030] It should be noted that this embodiment constructs a binary classification data set, classifies bright objects as "no cloud" and bright clouds as "cloudy", which can effectively solve the problem of false detection and missed detection in complex backgrounds, accurately distinguish the overall texture details, and reduce misidentification. The sliding window method is used to classify and distinguish each window in the image, and the overall texture details of the image block can be analyzed to determine whether the area is a cloud area. Finally, by superimposing the classification results of all sliding windows, the distribution of cloud areas in large-scale remote sensing images can be accurately obtained.

[0031] Result weighting and fusion: Map each sliding window detection result back to the corresponding position in the original remote sensing image; accumulate the weighted results of all sliding windows covering each pixel to generate an accumulated weight map, and further perform threshold segmentation to obtain a binary cloud mask, which specifically includes the following steps: 3-1 Window overlay: Map the detection results of each window back to its corresponding original image position. Accumulate the weighted results of all overlay windows for each pixel to generate a cumulative weight map : ; in, is the number of windows that cover the pixel; is the weight of each window (+1 or -1); The range of is: ;in is the width of the window; is the window sliding step; the number of sliding windows is N, N is a positive integer and 1≤i≤N; i is the i-th sliding window.

[0032] 3-2 Threshold segmentation: setting the threshold , generate a binary cloud mask : ; The default value is set to 0 and can be adjusted according to actual needs.

[0033] Post-processing stage: Post-process the binary cloud mask to generate the corresponding vector result, which includes the following steps: 4-1 Cropping the image: Cropping the detection result mask back to the original size of the image to ensure that the detection result is consistent with the original image result range.

[0034] 4-2 Generate the final cloud result: Output the final binary cloud mask and generate the corresponding vector result as needed.

[0035] In this embodiment, a training sample set for constructing a ResNet50 network model is also included. The training sample set uses a binary classification data set, classifies bright ground objects as cloudless, and classifies bright clouds as clouded. Specifically: Building a high-quality data set is the basis for realizing remote sensing image cloud detection. First, in this embodiment, the "Jilin-1" wide-band series of satellite remote sensing images are collected. Satellite remote sensing images need to cover a variety of landform scenes, such as cities, forests, deserts and oceans, and include different weather conditions such as sunny days, partial cloud coverage and complete cloud coverage, as well as time and seasonal changes. Subsequently, the satellite images are annotated using the labeling tool LabelImg and divided into two categories: "cloud" and "cloudless". At the same time, the data is divided according to the ratio of 70% training set, 20% validation set and 10% test set. In the preprocessing stage, all satellite remote sensing images are normalized as original remote sensing images, and the generalization ability of the model is improved by data enhancement operations such as rotation, cropping, and adjusting brightness. In addition, the boundaries of the original remote sensing images are symmetrically filled to reduce edge effects.

[0036] Network model training: In the training phase, the ResNet50 network is used as the main structure, and its classification head is a binary classifier, which is used to determine whether the input window contains clouds. In the network design, the cross entropy loss function is used as the optimization target. The training parameters of the model include using the Adam optimizer, the initial value of the learning rate is set to 0.001, the batch size is 32 or 64, and the number of training rounds is 50 to 100 rounds until the validation set loss converges. By randomly sampling small blocks of images in the training set and inputting them into the network, the classification probability of the forward propagation is calculated, and the loss is calculated in combination with the true label. The network parameters are updated in the back propagation, and the accuracy, recall rate, and F1 score of each round of training are recorded. At the same time, to prevent overfitting, L2 regularization and Dropout layers are added. The model parameters are continuously adjusted through evaluation on the validation set to achieve the best performance.

[0037] Actual parameters for cloud detection: sliding window size: 128×128 pixels; step size: 64 pixels; threshold is 0, and the cumulative weight map is converted into a binary cloud mask map. After generating the cloud mask, morphological operations such as dilation or erosion are used to remove isolated noise points, smooth the cloud boundaries, and finally crop the mask back to the original image size to output the final result.

[0038] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.

Claims

1. A remote sensing image cloud detection method with a variable receptive field, characterized in that: The following steps are involved: Acquire the original remote sensing image, and perform boundary filling preprocessing on the original remote sensing image to generate a filled image; Extract all sub-blocks of the filled image based on the sliding window, classify all sub-blocks for cloudiness, and weight the corresponding covered windows according to the classification results; Map each sliding window detection result back to the corresponding position in the original remote sensing image; accumulate the weighted results of all sliding windows covering each pixel to generate an accumulated weight map, and further perform threshold segmentation to obtain a binary cloud mask; Post-process the binary cloud mask to generate the corresponding vector result.

2. The remote sensing image cloud detection method with variable receptive field according to claim 1, characterized in that: Based on the sliding window, all sub-blocks of the filled image are extracted and classified for cloudiness or not. Specifically: Use the pre-trained network model to perform binary classification on each sub-block, and the binary classification results are cloud or no cloud; Among them, the pre-trained network model adopts the ResNet50 network model.

3. The remote sensing image cloud detection method with variable receptive field according to claim 1, characterized in that: The corresponding coverage window is weighted according to the classification result, specifically: For the classification results with clouds, the corresponding sliding window result weight is +1; For the cloud-free classification results, the corresponding sliding window results are weighted by -1.

4. The remote sensing image cloud detection method with variable receptive field according to claim 1, characterized in that: The cumulative weight graph expression is: ； in, is the cumulative weight graph; is the number of windows that cover the pixel; is the weight of each sliding window (+1 or -1); The range of is: ; is the width of the sliding window, is the sliding step of the sliding window; the number of sliding windows is N, N is a positive integer and 1≤i≤N; i is the i-th sliding window.

5. The remote sensing image cloud detection method with variable receptive field according to claim 1, characterized in that: The expression for generating a binary cloud mask is: ； In the formula, To preset the threshold, the default value of T is set to 0; is a binary cloud mask.

6. The remote sensing image cloud detection method with variable receptive field according to claim 2, characterized in that: It also includes a training sample set for building a ResNet50 network model, specifically: The training sample set adopts a binary classification data set, classifying bright ground objects as cloudless categories and bright clouds as cloud categories.

Images