Method for detecting airport runway in remote sensing image based on D-Unet

Abstract

The invention belongs to the technical field of airport runway detection, and particularly relates to a method for detecting an airport runway in a remote sensing image based on D-Unet, and the method comprises the following steps: S1, obtaining BARS data set sample pictures containing the airport runway in batches, and making a sample set; s2, training the sample set by adopting a D-Unet model, outputting an airport runway detection model and updating parameters; and S3, a remote sensing image to be detected is input into the updated airport runway detection model for detection and identification, a detected airport runway image is output and stored, and the detection further comprises the step of determining a detection threshold according to the updated airport runway model so as to judge whether pixels in the image belong to an airport runway or not. The airport runway detection efficiency is remarkably improved by combining a D-Unet deep learning algorithm, and meanwhile, the airport runway detection method has relatively high detection precision and is suitable for an airport runway automatic detection task in a complex remote sensing environment.

Description

Technical Field

[0001] The present invention relates to the technical field of airport runway detection, and in particular to a method for detecting an airport runway in a remote sensing image based on D-Unet. Background Art

[0002] In recent years, with the continuous development of remote sensing satellite technology, the spatial resolution of the acquired images has been significantly improved, and people can obtain richer and clearer remote sensing satellite images. High-resolution satellite data has become an important means of obtaining geographic information, and the use of remote sensing images to extract artificial objects has become one of the key applications in this field. By extracting objects, vector maps can be quickly generated, which has a wide range of use value in national economic construction and military applications. As an important artificial facility, airports are not only closely related to human daily life, but also a key element in military targets.

[0003] The current methods for airport extraction in remote sensing images include: technical processes based on scale transformation, edge detection, parallel line extraction, etc.; runway extraction methods based on bandpass filtering; runway extraction methods that eliminate non-airport features through image segmentation and combine feature extraction algorithms; and methods for airport runway detection using multi-scale Hough transform in image space. Although these technologies can detect airport targets, they are inefficient, often require a lot of time and effort, and are complex to operate. At the same time, for remote sensing images with complex backgrounds, the detection accuracy is not high and lacks versatility. Summary of the invention

[0004] 1. Technical issues to be solved

[0005] In view of the deficiencies of the prior art, the present invention provides a method for detecting airport runways in remote sensing images based on D-Unet, which solves the problems raised in the above-mentioned background technology.

[0006] (II) Technical solution

[0007] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:

[0008] The detection method of airport runways in remote sensing images based on D-Unet includes the following steps:

[0009] S1: Batch obtain sample images of the BARS dataset containing airport runways and create a sample set;

[0010] S2: Use the D-Unet model to train the sample set, output the airport runway detection model and update the parameters;

[0011] S3: Input the remote sensing image to be detected into the updated airport runway detection model for detection and recognition, output the detected airport runway image and save it, wherein the detection also includes determining the detection threshold according to the updated airport runway model to determine whether the pixels in the image belong to the airport runway.

[0012] Furthermore, the batch acquisition of sample images in S1 includes batch acquisition of remote sensing images containing complete airport runways from the BARS dataset, and batch acquisition of remote sensing data from other remote sensing data sources, such as Gaofen-1 or Gaofen-2 satellite images, to produce sample sets, and the storage format of the remote sensing images is .json.

[0013] Furthermore, in S1, while acquiring the sample images, the remote sensing images also need to be preprocessed, and after preprocessing, they need to be uniformly stored and managed in a format suitable for model input. The specific preprocessing content is: color channel standardization, histogram equalization and noise suppression of the sample images; to enhance image quality and improve model training effects.

[0014] Furthermore, the specific process of using the D-Unet model to train the sample set in S2 includes the following steps:

[0015] S2.1: Initialize the D-Unet network using the initialization parameters of the pre-trained model;

[0016] S2.2: Fine-tune the encoder and decoder modules of D-Unet based on the image features in the BARS dataset. During the fine-tuning process, focus on adjusting the feature extraction capability of the model to improve the sensitivity to specific runway features.

[0017] S2.3: Input the preprocessed remote sensing image into the D-Unet model for primary airport runway detection to generate primary runway image results; then, the primary detection results are used to update the sample data to enhance the adaptability of the model to different environmental conditions;

[0018] S2.4: Use a pixel-by-pixel loss function, such as cross entropy loss or Dice coefficient, to train the fine-tuned D-Unet model; perform end-to-end training and further fine-tuning at the same time, and adjust the convolutional and deconvolutional layers in the network to make it have higher resolution in detecting the edges and details of the airport runway;

[0019] S2.5: Repeatedly iterate training and parameter adjustment are performed on the updated sample set to further fine-tune the weights of each layer of the D-Unet network so that the model can better adapt to different image resolutions and complex environments.

[0020] Furthermore, when the updated D-Unet model is used for detection in S3, considering that the size of the remote sensing image to be detected may be different, the method includes adapting the image size to ensure that it meets the input requirements of the D-Unet model. Specifically, the remote sensing image to be detected is first resized to match the expected input resolution of the D-Unet model. The D-Unet model outputs a probability map for each pixel, and the airport runway area and the non-runway area are distinguished by setting an appropriate threshold. The threshold is the sample verification accuracy during model training. When the threshold is too low, false detection may occur. When the threshold is too high, missed detection may occur. Therefore, the threshold size can be adjusted according to the size of the remote sensing image to be detected and the detection effect.

[0021] (III) Beneficial effects

[0022] Compared with the prior art, the present invention provides a method for detecting airport runways in remote sensing images based on D-Unet, which has the following beneficial effects:

[0023] 1. The present invention uses the D-Unet algorithm to train batch acquired sample images to generate an optimized airport runway detection model. The model can perform high-precision detection and identification of airport runways in remote sensing images. The fully convolutional structure of the D-Unet network can effectively segment images at the pixel level, thereby improving the accuracy of detection. At the same time, by repeatedly iteratively training the sample set, the robustness and detection capability of the model can be gradually improved.

[0024] 2. By applying the D-Unet model to remote sensing image data from different sources (such as BARS datasets, images collected by GF-1 or GF-2), the present invention can effectively solve the detection accuracy problem caused by the difference in image features between different data sources. Regardless of the image source, the D-Unet model can achieve efficient airport runway detection through its deep feature extraction capabilities.

[0025] 3. The present invention also includes preprocessing the acquired remote sensing images to adapt to the input requirements of the D-Unet model, such as color space conversion and noise suppression. These preprocessing steps ensure the good performance of the model under various remote sensing image conditions and improve the versatility of the detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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.

[0028] Example

[0029] like Figure 1 As shown, a method for detecting an airport runway in a remote sensing image based on D-Unet according to an embodiment of the present invention comprises the following steps:

[0030] S1: Batch obtain sample images of the BARS dataset containing airport runways and create a sample set;

[0031] The batch acquisition of sample images includes preprocessing the images collected in batches by BARS and capturing the airport runway area, and batch acquiring sample images containing the complete airport runway; as well as acquiring remote sensing images collected in batches by GF-1 or GF-2, and extracting the target area of ​​the airport runway from them.

[0032] Sample images can be obtained in batches by using remote sensing images collected in batches by Gaofen-1 or Gaofen-2 to generate sample data with higher resolution; or by processing and intercepting images collected in batches by BARS to improve the accuracy of subsequent D-Unet model training. In order to optimize the generalization ability and accuracy of the model, the preferred approach is to combine BARS image processing and Gaofen satellite image data.

[0033] At the same time, the remote sensing images also need to be preprocessed, specifically: the sample images collected in batches by GF-1 or GF-2 are subjected to color channel standardization, histogram equalization and noise suppression to meet the requirements of the D-Unet network for multi-channel input, and to help improve the color information of the image, thereby improving the detection accuracy of the D-Unet model for airport runways. In order to further eliminate the differences between data sources and enhance the generalization performance of the model, BARS collected images and GF-1 or GF-2 remote sensing images are used to jointly produce sample sets. D-Unet can effectively use these multi-source data for deep learning training and improve the recognition accuracy of airport runways.

[0034] S2: Use the D-Unet model to train the sample set, output the airport runway detection model and update the parameters;

[0035] The specific process of training the sample set using the D-Unet model includes the following steps:

[0036] S2.1: Initialize the D-Unet network using the initialization parameters of the pre-trained model;

[0037] S2.2: Fine-tune the encoder and decoder modules of D-Unet based on the image features in the BARS dataset. During the fine-tuning process, focus on adjusting the feature extraction capability of the model to improve the sensitivity to specific runway features.

[0038] S2.3: Input the preprocessed remote sensing image into the D-Unet model for primary airport runway detection to generate primary runway image results; then, the primary detection results are used to update the sample data to enhance the adaptability of the model to different environmental conditions;

[0039] S2.4: Use a pixel-by-pixel loss function, such as cross entropy loss or Dice coefficient, to train the fine-tuned D-Unet model; perform end-to-end training and further fine-tuning at the same time, and adjust the convolutional and deconvolutional layers in the network to make it have higher resolution in detecting the edges and details of the airport runway;

[0040] S2.5: Repeatedly iterate training and parameter adjustment are performed on the updated sample set to further fine-tune the weights of each layer of the D-Unet network so that the model can better adapt to different image resolutions and complex environments.

[0041] The preprocessed remote sensing image is input into the D-Unet-based airport runway model for primary detection, and the primary airport runway image is output. On this basis, the primary airport runway image is screened to return the sample set, and the sample set is iteratively trained again to improve the model's recognition accuracy of airport runways in remote sensing images. Finally, a model suitable for airport runway detection and recognition is trained, which can efficiently detect and extract airport runway information in remote sensing images.

[0042] Through its U-shaped structure, D-Unet retains global context information during the encoding process, and effectively fuses low-level and high-level features during the decoding process through skip connections, gradually restoring high-resolution spatial details. This fully convolutional architecture enables D-Unet to process input remote sensing images of any size, while taking advantage of multi-scale feature fusion to achieve accurate detection and segmentation of airport runways. By aggregating features of different scales, D-Unet can not only capture the global information in remote sensing images, but also accurately identify the boundaries and local details of airport runways, significantly improving the efficiency and accuracy of airport runway detection.

[0043] S3: Input the remote sensing image to be detected into the updated airport runway detection model for detection and recognition, output the detected airport runway image and save it, wherein the detection also includes determining the detection threshold according to the updated airport runway model to determine whether the pixels in the image belong to the airport runway.

[0044] To improve the detection efficiency, when using the updated D-Unet model for detection, the size of the remote sensing image to be detected may be different. The method includes adapting the image size to ensure that it meets the input requirements of the D-Unet model. Specifically, the remote sensing image to be detected is first resized to match the expected input resolution of the D-Unet model. The D-Unet model outputs a probability map for each pixel, and distinguishes the airport runway area from the non-runway area by setting an appropriate threshold. The threshold is the sample verification accuracy during model training. When the threshold is too low, false detection may occur. When the threshold is too high, missed detection may occur. Therefore, the threshold size can be adjusted according to the size of the remote sensing image to be detected and the detection effect. The threshold is determined based on the performance of the model in the training and verification stages to optimize the detection results and avoid false detection or missed detection. The acquisition and processing of sample images should ensure that the image contains complete airport runway features and details to improve the detection accuracy and the versatility of the model. If necessary, large-size images can be divided into regions to ensure the detection accuracy of each region. This can ensure the efficiency and accuracy of detection, while improving the performance of the model under different remote sensing image conditions.

[0045] According to the airport runway detection model based on D-Unet of the present invention, the remote sensing images collected by GF-1 and the BARS data set are jointly extracted to count the number of airports in Jiangsu Province. First, the pre-processed GF-1 2-meter resolution image and the BARS data set image are input, and the two data are subjected to feature fusion processing. For the GF-1 data, the D-Unet model is used for pixel-level detection, and the sliding window size is 700 pixels × 700 pixels, the step size is 600 pixels, and the threshold is set to 0.94 to extract the preliminary runway position. Then, the image of the BARS data set is fine-tuned and extracted, and the accuracy and robustness of the detection are improved by combining the preliminary runway detection results. After running the program, a total of 11 airports in Jiangsu Province are finally detected, and the specific locations are marked. After verification, there are 7 civil airports, 2 military airports, 1 abandoned airport, and 1 airport under construction in the plan, and 0 missed detections. By introducing the feature information of the BARS data set, the accuracy and efficiency of the detection are further improved, making the whole process faster and more accurate.

[0046] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting airport runways in remote sensing images based on D-Unet, characterized in that: The following steps are involved: S1: Batch obtain sample images of the BARS dataset containing airport runways and create a sample set; S2: Use the D-Unet model to train the sample set, output the airport runway detection model and update the parameters; S3: Input the remote sensing image to be detected into the updated airport runway detection model for detection and recognition, output the detected airport runway image and save it, wherein the detection also includes determining the detection threshold according to the updated airport runway model to determine whether the pixels in the image belong to the airport runway.

2. The method for detecting airport runways in remote sensing images based on D-Unet according to claim 1, characterized in that: The batch acquisition of sample images in S1 includes batch acquisition of remote sensing images containing complete airport runways from the BARS dataset, and batch acquisition of remote sensing data from other remote sensing data sources, such as Gaofen-1 or Gaofen-2 satellite images, to produce sample sets, and the storage format of the remote sensing images is .json.

3. The method for detecting airport runways in remote sensing images based on D-Unet according to claim 1, characterized in that: In S1, while acquiring sample images, the remote sensing images also need to be preprocessed, and after preprocessing, they need to be uniformly stored and managed in a format suitable for model input. The specific preprocessing content is: color channel standardization, histogram equalization and noise suppression of the sample images to enhance image quality and improve model training effects.

4. The method for detecting airport runways in remote sensing images based on D-Unet according to claim 1, characterized in that: The specific process of using the D-Unet model to train the sample set in S2 includes the following steps: S2.1: Initialize the D-Unet network using the initialization parameters of the pre-trained model; S2.2: Fine-tune the encoder and decoder modules of D-Unet based on the image features in the BARS dataset. During the fine-tuning process, focus on adjusting the feature extraction capability of the model to improve the sensitivity to specific runway features. S2.3: Input the preprocessed remote sensing image into the D-Unet model for primary airport runway detection to generate primary runway image results; then, the primary detection results are used to update the sample data to enhance the adaptability of the model to different environmental conditions; S2.4: Use a pixel-by-pixel loss function, such as cross entropy loss or Dice coefficient, to train the fine-tuned D-Unet model; perform end-to-end training and further fine-tuning at the same time, and adjust the convolutional and deconvolutional layers in the network to make it have higher resolution in detecting the edges and details of the airport runway; S2.5: Repeatedly iterate training and parameter adjustment are performed on the updated sample set to further fine-tune the weights of each layer of the D-Unet network so that the model can better adapt to different image resolutions and complex environments.

5. The method for detecting airport runways in remote sensing images based on D-Unet according to claim 1, characterized in that: When the updated D-Unet model is used for detection in S3, considering that the size of the remote sensing image to be detected may be different, the method includes adapting the image size to ensure that it meets the input requirements of the D-Unet model. Specifically, the remote sensing image to be detected is first resized to match the expected input resolution of the D-Unet model. The D-Unet model outputs a probability map of each pixel, and the airport runway area and the non-runway area are distinguished by setting an appropriate threshold. The threshold is the sample verification accuracy during model training. When the threshold is too low, false detection may occur. When the threshold is too high, missed detection may occur. Therefore, the threshold size can be adjusted according to the size of the remote sensing image to be detected and the detection effect.

Images