High-resolution image ground object extraction method suitable for oil field area

Abstract

The invention relates to the technical field of ground feature extraction, in particular to a high-resolution image ground feature extraction method suitable for an oil field area, and the method comprises the steps: introducing a residual structure on the basis of a Unet framework, and employing a Strip Pooling strip-shaped pooling layer in a pooling module; and after the strip pooling mechanism is introduced, the feature extraction capability of the method is stronger, and the extracted features can better express image information, so that the extraction precision of the strip ground features in the oil field region is effectively improved. The invention discloses a high-resolution image ground object extraction method suitable for an oil field area. The method comprises the following steps: acquiring a to-be-researched high-resolution remote sensing image of the oil field area; constructing to obtain a sample data set; an SPMResUnet model is constructed; an oil field area ground feature classification optimal model is constructed; and obtaining an oil field area ground feature extraction result under the high-resolution image condition, and the like.

Description

Technical Field

[0001] The invention relates to the technical field of ground object extraction, and in particular to a high-resolution image ground object extraction method suitable for oil field areas. Background Art

[0002] In the process of oil extraction in oil fields, in order to accurately grasp the spatial distribution pattern and temporal and spatial change laws of key features in oil fields and ensure the smooth progress of oil production safety, technicians need to extract and confirm the surface information of oil fields. Among them, feature extraction technology, as a new technology for obtaining surface information using remote sensing images, has been widely studied and applied. Relevant scholars have also conducted a lot of experiments and research to improve the accuracy and efficiency of feature extraction.

[0003] In order to obtain the spatial distribution information of the object classification, the most accurate method is to use manual annotation. However, this method is costly and inefficient, and it is bound to be unable to be applied on a large scale. For the semantic segmentation task of multi-class object extraction from remote sensing images, the accuracy has also been greatly improved with the support of deep learning technology. For the extraction of multi-class objects in oilfield areas, high-precision and high-efficiency object information extraction has also become a hot research direction. The mainstream method gradually transitions from the more traditional image feature processing method to the method of building a convolutional neural network model based on deep learning. With the rapid development of deep learning image classification technology, many classification models in deep learning such as ResNet, DenseNet, Inception_v3, and AlexNet have been widely used in object extraction and have achieved good extraction results.

[0004] Specifically, for example, the patent title: A method for extracting weak information objects from remote sensing, application number: CN202211656515.8, records the following technical solutions: A method for extracting weak information objects from remote sensing, which quickly selects useful information related to the identification of weak information objects in electronic maps; uses feature selection and preparation methods to convert the aforementioned information into data content and format that can participate in the remote sensing extraction of weak information objects; uses object-oriented segmentation technology with small-scale thresholds to accurately generate object monomers that can match the edges of weak information objects; uses image processing technology processes to effectively extract weak information objects from remote sensing images with the help of the above information. The invention uses a combination of segmentation parameters with smaller size thresholds to generate object monomers that can match the edges of weak information objects, which can significantly improve the recognition accuracy of the boundaries of weak information objects; the invention uses a method process using spatial registration and image processing technology to complete the rapid remote sensing extraction of weak information objects, greatly improve work efficiency, and meet user needs.

[0005] However, after further research, the inventors found that the prior art, including the above patent application documents, still has the following defects: 1. Traditional ground feature extraction methods cannot fully utilize the high-quality texture features and spatial features of high-resolution images, resulting in misclassification, wrong classification, and missed classification. 2. The deep learning network structure can obtain the features of different depths of the image by training a large number of existing original images to achieve the purpose of segmentation. However, there are currently few data sets that can participate in model training for the extraction of special thematic features in oilfield areas. 3. The unique thematic features in the oil field area, such as oil well sites and well site roads, show certain shape regularities in remote sensing images, which are different from conventional features.

[0006] Therefore, it is urgent for those skilled in the art to design a high-resolution image object extraction method to meet and adapt to use in oil field areas. Summary of the invention

[0007] The present invention provides a method for extracting objects from high-resolution images in oilfield areas. The method introduces a residual structure based on a Unet framework and uses a Strip Pooling layer in a pooling module. After the strip pooling mechanism is introduced, the feature extraction capability of the present invention is stronger, and the extracted features can better express image information, thereby effectively improving the extraction accuracy of strip objects in oilfield areas.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions: A method for extracting objects from high-resolution images in oilfield areas, comprising the following steps: Step S1: obtaining high-resolution remote sensing images of the oilfield area to be studied and preprocessing the remote sensing images; Step S2: Based on the visual interpretation method, labels are drawn on the preprocessed remote sensing images to construct a sample data set; Step S3: constructing an SPM_ResUnet model; wherein the constructed SPM_ResUnet model at least includes: a Unet model framework, a residual module in ResNet, and a Strip Pooling module; Step S4: Based on the SPM_ResUnet model constructed in step S3, the sample data set constructed in step S2 is classified and verified; the optimal model for the classification of land features in the oilfield area is constructed; Step S5: According to the optimal model for classifying the oilfield area objects obtained in step S4, the oilfield area object extraction result under the condition of high-resolution image is obtained.

[0009] Preferably, the step of preprocessing the remote sensing image in step S1 can be specifically described as follows: The remote sensing images are processed in sequence through denoising, color correction, geometric correction and image enhancement, thereby reducing the impact of noise, geometric distortion, color distortion and other defects in the remote sensing images on their accuracy.

[0010] Preferably, step S2 can be specifically described as: S21: Based on the visual interpretation method, labels are drawn for the pre-processed remote sensing images; and classification training prediction results of the remote sensing images are obtained according to the drawn labels; S22: According to the classification training prediction results of the remote sensing image, a data enhancement processing operation is performed on the sample data to expand the sample data set; S23: Divide the sample data set after data enhancement processing into training set, test set and validation set according to a certain ratio.

[0011] Preferably, the residual module in the ResNet is formed by stacking a plurality of residual blocks with skip connections; wherein the stacking process of the residual blocks satisfies: F = W2σ(W1x) (Formula 1); y= F(x,{W i})+x (Formula 2); Among them, x represents the input of the residual block, σ represents the ReLU activation function, W1 and W2 represent the weights of each layer, F is the result after two weight layers and the first ReLU function activation, and y is the output of the residual block without the second activation.

[0012] Preferably, step S4 can be specifically described as: S41: input the training sample data in the sample data set constructed in step S2 into the SPM_ResUnet model constructed in step S3 to obtain a feature map of the training sample data; S42: performing downsampling module processing on the feature map of the training sample data, thereby compressing the number of channels of the high-level features therein to achieve feature extraction, and obtaining a high-level abstract semantic feature map; S43: Based on the fully connected layer in the SPM_ResUnet model, the high-level abstract semantic feature map obtained in step S42 is classified to achieve classification of the training sample data in the sample data set; S44: performing upsampling module processing on the feature map of the training sample data, and then performing 1*1 convolution dimensionality reduction processing to reduce the number of channels to a specific number, and obtaining a final classification target map; S45: The final classification target map obtained in step S44 is subjected to accuracy verification with the verification set in the sample data set, and finally an optimal model for classifying land features in the oilfield area is constructed whose extraction results can best meet the extraction accuracy requirements.

[0013] Preferably, the downsampling module is composed of two 3x3 convolutional layers and a Strip Pooling layer; wherein the Strip Pooling layer can be specifically described as: The Strip Pooling layer consists of two branches, one of which is pooled along the vertical direction and the other is pooled along the horizontal direction; the horizontal pooling and the vertical pooling respectively satisfy: Among them, the input x∈R H×w , the pooling kernel size is h×w; if it is assumed that H and W are divisible by h and w, the size after pooling is

[0014] Preferably, the pooling process of the Strip Pooling layer can be specifically described as follows: The feature map of the input training sample data is transformed into H×1 and 1×W through horizontal pooling and vertical pooling; After a 1D convolution with a convolution kernel of 3, the two outputs obtained by horizontal pooling and vertical pooling transformation are expanded and fused, so that the feature map dimensions of the two output branches after expansion and fusion are consistent, and we get The corresponding positions of the feature maps of the two output branches are summed, and then processed by 1×1 convolution and sigmoid to obtain a weight matrix; Multiply the original input by the corresponding position of the weight matrix to obtain the final output result Z = scale(x,σ(f(y))).

[0015] Preferably, the upsampling module is repeatedly composed of a layer of deconvolution, feature concatenation and two 3x3 convolution layers.

[0016] The present invention provides a method for extracting objects from high-resolution images in oilfield areas, which includes the steps of obtaining high-resolution remote sensing images of the oilfield area to be studied; constructing a sample data set; constructing an SPM_ResUnet model; constructing an optimal model for classifying objects in the oilfield area; and obtaining the object extraction results in the oilfield area under the condition of high-resolution images. Compared with the prior art, the method for extracting objects from high-resolution images in oilfield areas with the above-mentioned steps consumes relatively less manpower, material resources and time costs, and can obtain more accurate object extraction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A schematic diagram of a process flow of a method for extracting objects from high-resolution images of oilfield areas provided by the present invention; Figure 2 A schematic diagram of the structure of the SPM_ResUnet model in a method for extracting objects from high-resolution images in oilfield areas provided by the present invention; Figure 3 It is a schematic diagram of the structure of the residual module in the SPM_ResUnet model; Figure 4 This is a schematic diagram of the module of the strip pooling core in the SPM_ResUnet model; Figure 5 A schematic diagram of ground object prediction results of a high-resolution image ground object extraction method applicable to oil field areas provided by the present invention. DETAILED DESCRIPTION

[0018] The present invention provides a method for extracting objects from high-resolution images in oilfield areas. The method introduces a residual structure based on a Unet framework and uses a Strip Pooling layer in a pooling module. After the strip pooling mechanism is introduced, the feature extraction capability of the present invention is stronger, and the extracted features can better express image information, thereby effectively improving the extraction accuracy of strip objects in oilfield areas. Embodiment 1

[0019] The present invention provides a method for extracting objects from high-resolution images in oilfield areas. Figure 1 As shown, the following steps are included: Step S1: Obtain high-resolution remote sensing images of the oilfield area to be studied and preprocess the remote sensing images.

[0020] Among them, as a more preferred embodiment of the present invention, the preprocessing process of the remote sensing image can be further specifically described as: performing denoising, color correction, geometric correction, and image enhancement processing operations on the remote sensing image in sequence. The purpose of the above preprocessing of the remote sensing image is to reduce the influence of defects such as noise, geometric distortion, and color distortion in the remote sensing image on the accuracy of the remote sensing image, thereby improving the extraction accuracy of the ground object extraction method of the present invention.

[0021] After completing step S1, step S2 is further implemented: based on the visual interpretation method, labels are drawn on the pre-processed remote sensing images to construct a sample data set.

[0022] After completing step S2, further implement step S3: construct the SPM_ResUnet model; wherein, Figure 2 As shown in the figure, the constructed SPM_ResUnet model includes at least: Unet model framework, residual module in ResNet (the residual module structure in ResNet is as follows Figure 3 As shown), Strip Pooling module (Strip Pooling module structure as shown Figure 4 shown).

[0023] It should be noted that the residual module in ResNet in the SPM_ResUnet model is specifically composed of a plurality of residual blocks with skip connections stacked together. The stacking process of the residual blocks mentioned in the above steps satisfies: F = W2σ(W1x) (Formula 1); y= F(x,{W i})+x (Formula 2); Among them, x represents the input of the residual block, σ represents the ReLU activation function, W1 and W2 represent the weights of each layer, F is the result after two weight layers and the first ReLU function activation, and y is the output of the residual block without the second activation.

[0024] After completing step S3, step S4 is further implemented: based on the SPM_ResUnet model constructed in step S3, the sample data set constructed in step S2 is classified and verified; and the optimal model for the classification of land features in the oilfield area is constructed.

[0025] After completing step S4, step S5 is further implemented: according to the optimal model of oilfield area feature classification obtained in step S4, the oilfield area feature extraction result under the condition of high-resolution image is obtained; specifically, the oilfield area feature extraction result can be referred to as follows Figure 5 shown. Embodiment 2

[0026] On the basis of the first embodiment, the second embodiment further provides a method for extracting objects from high-resolution images of oilfield areas provided by the present invention, wherein step S2 is specifically described as follows: Specifically, step S2 can be described as follows: S21: Based on the visual interpretation method, labels are drawn for the pre-processed remote sensing images; and classification training prediction results of the remote sensing images are obtained according to the drawn labels; S22: According to the classification training prediction results of the remote sensing image, a data enhancement processing operation is performed on the sample data to expand the sample data set; S23: Divide the sample data set after data enhancement processing into training set, test set and validation set according to a certain ratio.

[0027] It is worth noting that step S2 uses the method of visual interpretation to draw (higher precision data) labels for the preprocessed remote sensing images; and based on the classification training prediction results of the remote sensing images, the sample data set is divided into appropriate training sets, test sets and validation sets in proportion, thereby greatly improving the accuracy of the sample data set obtained after the remote sensing image operation and processing and the effectiveness of the model training. Embodiment 3

[0028] On the basis of the first embodiment, the third embodiment further provides a method for extracting objects from high-resolution images of oilfield areas provided by the present invention, wherein step S4 is specifically described as follows: Specifically, step S4 can be described as follows: S41: input the training sample data in the sample data set constructed in step S2 into the SPM_ResUnet model constructed in step S3 to obtain a feature map of the training sample data; S42: performing downsampling module processing on the feature map of the training sample data, thereby compressing the number of channels of the high-level features therein to achieve feature extraction, and obtaining a high-level abstract semantic feature map; S43: Based on the fully connected layer in the SPM_ResUnet model, the high-level abstract semantic feature map obtained in step S42 is classified to achieve classification of the training sample data in the sample data set; S44: performing upsampling module processing on the feature map of the training sample data, and then performing 1*1 convolution dimensionality reduction processing to reduce the number of channels to a specific number, and obtaining a final classification target map; S45: The final classification target map obtained in step S44 is subjected to accuracy verification with the verification set in the sample data set, and finally an optimal model for classifying land features in the oilfield area is constructed whose extraction results can best meet the extraction accuracy requirements.

[0029] It is worth noting that the specific structural form of the SPM_ResUnet model involved in this embodiment can be referred to as shown in the attached Figure 2 As shown in the figure, the adopted SPM_ResUnet framework is a symmetrical network structure, with a downsampling module on the left and an upsampling module on the right. The Skip Connection in the middle is four gray parallel lines, which superimpose the channels of the corresponding feature maps. The purpose is to fuse the feature maps in the downsampling process during the upsampling process, and to fill and superimpose the small feature maps.

[0030] Further, as a preferred embodiment of the present invention, the down-sampling module and the up-sampling module can be described as follows: The downsampling module consists of two 3x3 convolutional layers and a Strip Pooling layer; the StripPooling layer can be specifically described as: The Strip Pooling layer consists of two branches, one of which is pooled along the vertical direction and the other is pooled along the horizontal direction; the horizontal pooling and the vertical pooling respectively satisfy: Among them, the input x∈R H×w , the pooling kernel size is h×w; if it is assumed that H and W are divisible by h and w, the size after pooling is

[0031] Among them, the pooling process of the Strip Pooling layer can be specifically described as: The feature map of the input training sample data is transformed into H×1 and 1×W through horizontal pooling and vertical pooling; After a 1D convolution with a convolution kernel of 3, the two outputs obtained by horizontal pooling and vertical pooling transformation are expanded and fused, so that the feature map dimensions of the two output branches after expansion and fusion are consistent, and we get The corresponding positions of the feature maps of the two output branches are summed, and then processed by 1×1 convolution and sigmoid to obtain a weight matrix; Multiply the original input by the corresponding position of the weight matrix to obtain the final output result Z = scale(x,σ(f(y))).

[0032] The upsampling module is composed of a layer of deconvolution, feature concatenation and two 3x3 convolution layers repeatedly.

[0033] It should be noted that the downsampling module processing process will perform processes such as convolution, batch processing, activation, etc. For example, if the input image size is 256*256*3, after a step size of 1, 3*3 convolution, a 128*128*64 feature map is obtained, and then after a residual and pooling process, a 128*128*64 feature map is obtained. The entire downsampling process is repeated for four times, and finally a 16*16*512 feature map is obtained. The upsampling process is repeated by a layer of deconvolution, feature concatenation, and two 3x3 convolution layers. After at least 4 such operations, it can be guaranteed to be consistent with the input target. Figure 1Then, the accuracy of the classified target map and the verification set is verified to ensure that the ground feature extraction model can make the extraction result meet the extraction accuracy requirements. I will not go into details here.

[0034] In addition, as a newly supplemented and more preferred implementation, the software environment used in the present invention is further explained as follows: First, the deep learning platform used in the method for extracting objects from high-resolution images in oilfield areas provided by the present invention is the Pytorch framework, and the cross entropy loss function is used in model training to measure the generalization ability of the model, and Adam optimization is used to update the model through back propagation. The learning rate of the weighted stochastic gradient descent uses a fixed value, and the initial learning rate is set to 0.001.

[0035] And, in order to further reflect the technical optimization of the method for extracting ground objects from high-resolution images of oilfield areas provided by the present invention, it is compared with other learning models as follows. Specifically, the existing learning models involved in the comparison are Unet and ResUnet.

[0036] After comparison, it is found that the total classification accuracy (OA), FWIOU, and Kappa coefficient of the high-resolution image feature extraction method for oilfield areas provided by the present invention are the highest when trained to convergence on the validation set, among which OA, FWIOU, and Kappa reach 0.9526, 0.9113, and 0.8322 respectively. The total classification accuracy (OA), Kappa coefficient, and FWIOU of different models on the validation set are shown in Table 1 below. Table 1 Classification method Accuracy / (%) Kappa FWOU Unet 0.9400 0.7805 0.8888 ResUnet 0.9509 0.8254 0.9083 SPM_ResUnet 0.9526 0.8322 0.9113

[0037] Thus, it is proved that the present invention provides a method for extracting objects from high-resolution images in oilfield areas and its technical advantages.

[0038] The present invention provides a method for extracting objects from high-resolution images in oilfield areas, which includes the steps of obtaining high-resolution remote sensing images of the oilfield area to be studied; constructing a sample data set; constructing an SPM_ResUnet model; constructing an optimal model for classifying objects in the oilfield area; and obtaining the object extraction results in the oilfield area under the condition of high-resolution images. Compared with the prior art, the method for extracting objects from high-resolution images in oilfield areas with the above-mentioned steps consumes relatively less manpower, material resources and time costs, and can obtain more accurate object extraction results.

Claims

1. A method for extracting objects from high-resolution images in oilfield areas, characterized in that: The steps include: Step S1: obtaining high-resolution remote sensing images of the oilfield area to be studied and preprocessing the remote sensing images; Step S2: Based on the visual interpretation method, labels are drawn on the preprocessed remote sensing images to construct a sample data set; Step S3: constructing an SPM_ResUnet model; wherein the constructed SPM_ResUnet model at least includes: a Unet model framework, a residual module in ResNet, and a Strip Pooling module; Step S4: Based on the SPM_ResUnet model constructed in step S3, the sample data set constructed in step S2 is classified and verified; the optimal model for the classification of land features in the oilfield area is constructed; Step S5: According to the optimal model for classifying the oilfield area objects obtained in step S4, the oilfield area object extraction result under the condition of high-resolution image is obtained.

2. The method for extracting objects from high-resolution images in oilfield areas according to claim 1, characterized in that: The step of preprocessing the remote sensing image in step S1 can be specifically described as: The remote sensing images are processed in sequence through denoising, color correction, geometric correction and image enhancement, thereby reducing the impact of noise, geometric distortion, color distortion and other defects in the remote sensing images on their accuracy.

3. The method for extracting objects from high-resolution images in oilfield areas according to claim 1, characterized in that: The step S2 can be specifically described as: S21: Based on the visual interpretation method, labels are drawn for the pre-processed remote sensing images; and classification training prediction results of the remote sensing images are obtained according to the drawn labels; S22: According to the classification training prediction results of the remote sensing image, a data enhancement processing operation is performed on the sample data to expand the sample data set; S23: Divide the sample data set after data enhancement processing into training set, test set and validation set according to a certain ratio.

4. The method for extracting objects from high-resolution images in oilfield areas according to claim 1, characterized in that: The residual module in the ResNet is formed by stacking multiple residual blocks with skip connections; wherein the stacking process of the residual blocks satisfies: F = W2σ(W1x) (Formula 1); y = F(x, {W i [[ID=2}]}) + x (Equation 2); Among them, x represents the input of the residual block, σ represents the ReLU activation function, W1 and W2 represent the weights of each layer, F is the result after two weight layers and the first ReLU function activation, and y is the output of the residual block without the second activation.

5. The method for extracting objects from high-resolution images in oilfield areas according to claim 1, characterized in that: The step S4 can be specifically described as: S41: input the training sample data in the sample data set constructed in step S2 into the SPM_ResUnet model constructed in step S3 to obtain a feature map of the training sample data; S42: performing downsampling module processing on the feature map of the training sample data, thereby compressing the number of channels of the high-level features therein to achieve feature extraction, and obtaining a high-level abstract semantic feature map; S43: Based on the fully connected layer in the SPM_ResUnet model, the high-level abstract semantic feature map obtained in step S42 is classified to achieve classification of the training sample data in the sample data set; S44: performing upsampling module processing on the feature map of the training sample data, and then performing 1*1 convolution dimensionality reduction processing to reduce the number of channels to a specific number, and obtaining a final classification target map; S45: The final classification target map obtained in step S44 is subjected to accuracy verification with the verification set in the sample data set, and finally an optimal model for classifying land features in the oilfield area is constructed whose extraction results can best meet the extraction accuracy requirements.

6. The method for extracting objects from high-resolution images in oilfield areas according to claim 5, characterized in that: The downsampling module consists of two 3x3 convolutional layers and a Strip Pooling layer; wherein the Strip Pooling layer can be specifically described as: The Strip Pooling layer consists of two branches, one of which is pooled along the vertical direction and the other is pooled along the horizontal direction; the horizontal pooling and the vertical pooling respectively satisfy: Among them, the input x∈R H×w , the pooling kernel size is h×w; if it is assumed that H and W are divisible by h and w, the size after pooling is 7. The method for extracting objects from high-resolution images in oilfield areas according to claim 6, characterized in that: The pooling process of the Strip Pooling layer can be specifically described as: The feature map of the input training sample data is transformed into H×1 and 1×W through horizontal pooling and vertical pooling; After a 1D convolution with a convolution kernel of 3, the two outputs obtained by horizontal pooling and vertical pooling transformation are expanded and fused, so that the feature map dimensions of the two output branches after expansion and fusion are consistent, and we get The corresponding positions of the feature maps of the two output branches are summed, and then processed by 1×1 convolution and sigmoid to obtain a weight matrix; Multiply the original input by the corresponding position of the weight matrix to obtain the final output result Z = scale(x,σ(f(y))).

8. The method for extracting objects from high-resolution images in oilfield areas according to claim 5 is characterized in that: The upsampling module is repeatedly composed of a layer of deconvolution, feature concatenation and two 3x3 convolution layers.

Images