A geological environment remote sensing interpretation method and device and related components

Abstract

The application discloses a geological environment remote sensing interpretation method and device and related components, and relates to the field of remote sensing image information processing.The method comprises the following steps: extracting features of a sample image by using a ResNet network to obtain shallow image features; performing multi-scale feature aggregation on the shallow image features by using a multi-scale feature extraction module to obtain a multi-scale geological element feature map; performing global feature fusion on the multi-scale geological element feature map by using a global feature fusion module to output a prediction probability map, so that a geological element classification model is constructed; training and optimizing the geological element classification model by using a loss function to obtain an optimized geological element classification model; and classifying a target image by using the optimized geological element classification model and outputting a classification result.The method effectively solves the problems of low accuracy of geological remote sensing interpretation tasks and few label samples.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote sensing image information processing, and in particular to a geological environment remote sensing interpretation method and device and related components. BACKGROUND

[0002] At present, with the rapid development of satellite sensors, spectral remote sensing data has been widely used in geological remote sensing interpretation, and has assisted geological survey and mapping. Geological remote sensing interpretation uses the differences in mineral composition, structure, physical properties, etc. of geological elements reflected on remote sensing images to identify lithology, soil, water, and other geological elements. Due to the need for a large amount of manpower and material resources for traditional field geological survey and mapping, especially in large-scale and difficult-to-reach areas, it is particularly important to build a remote sensing information processing deep interpretation model under such circumstances.

[0003] At present, geological environment remote sensing automatic classification uses various spectral satellite data to identify rocks, minerals, rock glaciers, geological disasters, etc. However, due to the characteristics of geological elements, the accuracy of current geological remote sensing automatic interpretation is still not high. Specifically, on the one hand, in remote sensing images, the inter-class difference of ground objects is small and the similarity is high. The geological elements are subjected to wind and sand erosion and transportation all year round, making the boundaries between elements blurred and the characteristics not obvious. Under the geological environment, the exposed mountains are subjected to glacial erosion and cutting all year round, and under the influence of regional faults, folds and other geological structures, the distribution of some geological elements is very fragmented and scattered, making the geological remote sensing interpretation result very fragmented and producing hole phenomenon. Traditional feature extractors cannot accurately extract element features.

[0004] In terms of accuracy, the current deep learning method represented by convolutional neural network has become the mainstream method of geological remote sensing intelligent interpretation. CNN model has very significant effect on the interpretation of single geological element, but the effect on the geological interpretation of all elements such as lithology, soil, rock glacier and water is poor. In the pixel-level semantic segmentation task, CNN model extracts local spatial information and is easily trapped in local optimum, resulting in serious fragmentation and holes in the classification result, i.e. the segmentation result has scattered misclassification. SUMMARY

[0005] The purpose of the present application is to provide a geological environment remote sensing interpretation method, device and related components, which aims to solve the problem of poor accuracy of existing neural network models in interpreting geological environment remote sensing images.

[0006] To solve the above technical problems, the purpose of the present application is achieved by the following technical scheme: a geological environment remote sensing interpretation method is provided, which comprises:

[0007] extracting features of the sample image by using a ResNet network to obtain shallow image features;

[0008] aggregating multi-scale features of the shallow image features by using a multi-scale feature extraction module to obtain a multi-scale geological element feature map;

[0009] fusing global features of the multi-scale geological element feature map by using a global feature fusion module to output a prediction probability map, thereby constructing a geological element classification model;

[0010] training and optimizing the geological element classification model by using a loss function to obtain an optimized geological element classification model;

[0011] inputting an unlabeled image into the optimized geological element classification model to generate a pseudo-labeled image, and performing self-supervised learning on the geological element classification model by using the pseudo-labeled image;

[0012] classifying a target image by using the optimized geological element classification model and outputting a classification result.

[0013] In addition, the technical problem to be solved by the present application is also to provide a geological environment remote sensing interpretation device, which comprises:

[0014] a shallow image feature extraction unit configured to extract features of a sample image by using a ResNet network to obtain shallow image features;

[0015] a multi-scale geological element feature map extraction unit configured to aggregate multi-scale features of the shallow image features by using a multi-scale feature extraction module to obtain a multi-scale geological element feature map;

[0016] a global feature fusion unit configured to fuse global features of the multi-scale geological element feature map by using a global feature fusion module to output a prediction probability map, thereby constructing a geological element classification model;

[0017] an optimization unit configured to train and optimize the geological element classification model by using a loss function to obtain an optimized geological element classification model;

[0018] a self-supervised learning unit configured to input an unlabeled image into the optimized geological element classification model to generate a pseudo-labeled image, and perform self-supervised learning on the geological element classification model by using the pseudo-labeled image;

[0019] a classification unit configured to classify a target image by using the optimized geological element classification model and output a classification result.

[0020] In addition, the embodiment of the present application further provides a computer device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the geological environment remote sensing interpretation method of the first aspect when executing the computer program.

[0021] In addition, the embodiment of the present application further provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program causes the processor to execute the geological environment remote sensing interpretation method of the first aspect when executed by the processor.

[0022] The embodiment of the present application discloses a geological environment remote sensing interpretation method and device and related components, wherein the method comprises: extracting features of a sample image using a ResNet network to obtain shallow image features; aggregating multi-scale features of the shallow image features using a multi-scale feature extraction module to obtain a multi-scale geological element feature map; fusing global features of the multi-scale geological element feature map using a global feature fusion module to output a prediction probability map, thereby constructing a geological element classification model; training and optimizing the geological element classification model using a loss function to obtain an optimized geological element classification model; inputting an unlabeled image into the optimized geological element classification model and generating a pseudo-labeled image, and using the pseudo-labeled image to perform self-supervised learning on the geological element classification model; and classifying a target image using the optimized geological element classification model and outputting a classification result. This method effectively solves the problems of low accuracy of geological remote sensing interpretation tasks and few labeled samples. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0024] Figure 1 The geological environment remote sensing interpretation method provided by the embodiment of the present application is shown in the flowchart.

[0025] Figure 2 The schematic block diagram of the geological environment remote sensing interpretation device provided by the embodiment of the present application is shown.

[0026] Figure 3 The schematic block diagram of the computer device provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0027] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be clearly and completely described below, obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0028] It should be understood that the terms "comprising" and "including" as used in the specification and the appended claims indicate the presence of the described features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0029] It should also be understood that the terms used in the present application specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the present application specification and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0030] It should be further understood that the term "and / or" used in the present application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0031] Please refer to Figure 1 , Figure 1 The flowchart of the geological environment remote sensing interpretation method provided by the embodiments of the present application is shown in the figure.

[0032] As Figure 1 shown, the method comprises steps S101-S106.

[0033] S101, feature extraction is performed on a sample image by using a ResNet network to obtain a shallow image feature;

[0034] S102, multi-scale feature aggregation is performed on the shallow image feature by using a multi-scale feature extraction module to obtain a multi-scale geological element feature map;

[0035] S103, global feature fusion is performed on the multi-scale geological element feature map by using a global feature fusion module to output a prediction probability map, thereby constructing a geological element classification model;

[0036] S104, the geological element classification model is trained and optimized by using a loss function to obtain an optimized geological element classification model;

[0037] S105, input the unlabeled image into the optimized geological element classification model, and generate a pseudo-labeled image, and use the pseudo-labeled image to perform self-supervised learning on the geological element classification model;

[0038] S106, classifying the target image by using the optimized geological element classification model, and outputting a classification result.

[0039] In the embodiment, the sample image is multispectral landsat8 satellite image data. It should be noted that the satellite is equipped with an Operational Land Imager (OLI) and a Thermal Infrared Sensor (TIRS), covering 11 bands. In order to further reflect the color difference of lithology and other geological elements on the image, on the basis of the synthesis of the 432 band, the sample image is preprocessed by using a gamma stretching transformation function with a coefficient of 1.146 in the Arcgis software, so that the sample image is cropped to a size of 224x224.

[0040] In step S101, the ResNet50 network is used to extract features of the sample image. The ResNet101 network can also be used to extract features of the sample image, but the depth of the ResNet101 network is deeper, which increases the burden of model training. After obtaining the extracted shallow image features, the shallow image features are input into the target semantic based dilated pyramid spatial pooling module (ASP-OC) (i.e. the multi-scale feature extraction module) to obtain a multi-scale element feature map considering target semantic information. Then, the global feature fusion module GAM is used to fully consider the "global" interaction information of the channel and space of the feature map, and output a prediction probability map to construct a geological element classification model. In addition, the loss function is used to train and optimize the geological element classification model, so that the optimized geological element classification model has better classification performance. The geological element classification model generated by the present application has good interpretation / classification accuracy for the geological image of all elements.

[0041] It should be noted that the present application performs remote sensing intelligent interpretation of 9 types of geological elements of all elements on the target image, including sandstone, carbonate rock, granite, volcanic clastic rock, slate, schist, glacier, lake and soil body. Among them, sandstone and carbonate rock are sedimentary rocks; granite and volcanic clastic rock are magmatic rocks; slate and schist are metamorphic rocks.

[0042] Since the pixel-level sample labeling in the field of geological remote sensing is time-consuming and laborious, there are few labeled sample images and many unlabeled sample images. Therefore, the geological element classification model is subjected to self-supervised learning by the unlabeled data in step S105, so as to sufficiently utilize the unlabeled samples, improve the accuracy of semantic segmentation, reduce the dependence of the model on the labeled samples to a certain extent, and increase the generalization ability of the geological element classification model.

[0043] In a specific embodiment, the step S102 comprises the following steps:

[0044] S10, inputting the shallow image feature into a 1x1 convolution layer to obtain a first convolution feature map;

[0045] S11, sequentially inputting the shallow image feature into three 3x3 convolution layers to obtain corresponding second convolution feature maps, wherein the dilation rates of each 3x3 convolution layer are different;

[0046] S11, sequentially inputting the shallow image feature into a 3x3 convolution layer and a target semantic pooling layer to obtain a third convolution feature map;

[0047] S12, performing feature splicing on the first convolution feature map, the three second convolution feature maps, and the third convolution feature map to obtain a multi-scale geological element feature map.

[0048] In step S11, the dilation rates of the three 3x3 convolution layers are 12, 24, and 36 respectively, which are used to adjust the interval number of the convolution kernel; by performing feature splicing on the obtained first convolution feature map, the three second convolution feature maps, and the third convolution feature map, a multi-scale geological element feature map is obtained.

[0049] In a specific embodiment, the step S12 comprises the following steps:

[0050] S20, inputting the shallow image feature into a 3x3 convolution layer to obtain a transition convolution feature map;

[0051] S21, inputting the transition convolution feature map into a target semantic pooling layer, so that the target semantic pooling layer performs a target semantic estimation step and a target semantic mapping step on the transition convolution feature map to obtain a third convolution feature map;

[0052] The target semantic estimation step comprises:

[0053] The class of each pixel in the transition convolution feature map is obtained, and the similarity degree of the current pixel and the pixels belonging to the same class is calculated according to the following formula to obtain a pixel similarity probability map w:

[0054]

[0055] wherein, xp The vector x represents pixel p. i The vector f represents pixel i. q Represents query transformation, f k Represents key transformation, Z p This represents the feature similarity between the current pixel p and all remaining pixels;

[0056] In this embodiment, (x p ) T In this context, "T" represents transpose; where f q q and f k In this context, k is similar to the q-k transformation in attention mechanisms, but here it is replaced by a convolution with a kernel size of 1×1.

[0057] S24, The target semantic mapping step includes:

[0058] The target context information of the pixels is obtained by weighted aggregation of the feature similarities of all remaining pixels using the following formula:

[0059]

[0060] Among them, c p φ represents the target context information of pixel p, and φ represents the transformation of the input features into different feature spaces, which is replaced by convolution.

[0061] In this embodiment, information from other pixels belonging to the same target is used to classify the current pixel. In step S23, the similarity between the current pixel p and other pixels belonging to the same target category is calculated. For example, in a remote sensing image, a ground feature is composed of many pixels. Step S23 of this application calculates the similarity between one pixel and other pixels in the neighborhood that belong to the same category, and finally obtains a pixel similarity probability map.

[0062] After obtaining the pixel similarity probability map, step S24 is the process of weighted aggregation of all pixels except pixel p, mapping the entire target semantic map corresponding to pixel p onto a single pixel to obtain the target context information of the pixel.

[0063] In one specific embodiment, step S103 includes the following steps:

[0064] S30. Perform feature fusion on the multi-scale geological element feature map according to the following formula, and output the prediction probability map F3:

[0065]

[0066]

[0067] Where F1∈RC×H×W F1 represents the multi-scale geological element feature map, F2 represents the intermediate feature map, M c F1 represents the channel attention map, M s F1 represents the spatial attention map, F1 represents the multiplication operation by element.

[0068] In this embodiment, CxHxW respectively represent the channel, height and width of the multi-scale geological element feature map; wherein, after the multi-scale feature extraction module considers the local spatial semantic information to obtain the multi-scale geological element feature, the invalid feature redundant information will also rise (this application combines the multi-scale features output by the three hole convolutions and the 1x1 point convolution with the target semantic context features output by the OCP in the channel, that is, the concat operation, so that it has certain information redundancy in the channel or space), at the same time, when the multi-scale feature extraction module focuses on the local information of the space, it will ignore the interaction information between the channel and the spatial features, thereby affecting the effect of geological element interpretation. Therefore, this application adopts the global attention GAM of “global” cross-dimension interactions in the channel and spatial dimensions to further fuse the channel features and spatial features, so that the final output prediction probability map has high interpretation effect.

[0069] The ASP-OC module of this application can fully extract multi-scale geological element features by combining different dilated rate hole convolutions and target semantic pooling mechanisms, while also considering the local spatial semantic information inside the geological element, which is beneficial to distinguish the boundary of the geological element and alleviate the problem of small inter-class difference, high similarity and unobvious features.

[0070] In a specific embodiment, the step S104 includes the following steps:

[0071] S40, converting the sample image with label into a label encoding map, and inputting the label encoding map and the prediction probability map into a preset discriminator to output a corresponding confidence map;

[0072] In this embodiment, the label encoding map obtained after the original label image is one-hot encoded is input into the preset discriminator together with the prediction probability map to output a corresponding confidence map; wherein, the sample for semantic segmentation training is pixel-level, including: an original rgb color image and a gray label image, it should be noted that each pixel has a corresponding label, which is represented by a gray value.

[0073] In this embodiment, the prediction probability map is obtained by classifying the input original image by the geological element classification model, wherein the sample image with label is labeled by artificial means.

[0074] It should be noted that the label encoding map of the present application is a Ground truth (label) label encoding map.

[0075] S41, wherein the loss function of the discriminator FCD is calculated as follows:

[0076]

[0077] wherein Y n represents the label encoding map, G(X n ) represents the predicted probability map, (h, w) represents the position of the pixel, represents the confidence map corresponding to the predicted probability map, represents the confidence map corresponding to the label encoding map, and:

[0078]

[0079]

[0080] S42, the cross-entropy loss L between the predicted probability map and the label encoding map is calculated as follows: ce

[0081]

[0082] S43, the adversarial loss L is calculated as follows: adv

[0083]

[0084] S44, the loss function of the geological element classification model is constructed as follows:

[0085] L = L FCD + L ce + L adv .

[0086] In the present embodiment, the geological element classification model is regarded as a generator, and an image X n with a size of HxWx3 is input, the label encoding map of the true semantic label is Y n , and G(X n ) is the predicted probability map output by the generator, wherein the size of the predicted probability map is HxWxC, and C represents the number of categories. In the present embodiment, the purpose of the generator is to make the predicted probability map continuously approximate (fit) the distribution of Y n , so as to achieve the purpose of confusing the discriminator. The input of the discriminator is the predicted probability map G(X n ) and the label encoding map Y n ​​, the output size of the discriminator is a confidence map of HxWx1, and the role of the discriminator is to distinguish the prediction probability map G(X n ) and the label encoding map Y n .

[0087] In this embodiment, the input of the discriminator has two types, that is, in the process of training, the input of the discriminator FCD can be G(X n ) or Y n . If y n = 0 indicates that the input is a prediction probability map, and y n = 1 indicates that the input is a label encoding map. In the formula of L FCD in step S40, the first term corresponds to the input of the prediction probability map, and the second term corresponds to the input of the one-hot encoded label map.

[0088] It should be noted that the purpose of the discriminator is to distinguish whether the input is the prediction probability map output by the geological element classification model or the label encoding map, and the purpose of the generator is to output a prediction probability map that is more and more realistic and label encoding map to confuse the discriminator. In other words, through the mutual confrontation between the discriminator and the generator, the geological element classification model is more realistic, and the interpretation effect of the image is better.

[0089] Based on the idea of multi-task learning, the application adds an adversarial loss term on the basis of the labeled data cross-entropy loss, so as to constrain the gradient update direction of the segmentation network. The cross-entropy loss in step S42 is used to ensure the basic segmentation ability of semantic segmentation; based on the "min-max game" of GAN, in order to make the output prediction probability map of the geological element classification model confuse the discriminator as much as possible, the first term in the formula of L FCD in step S40 is maximized, and then L adv is minimized, that is, minimizing the adversarial loss of the geological element classification model L adv is equivalent to maximizing the first term in the formula of L FCD . In this way, the geological element classification model parameters are updated through back propagation, and the constraint of the discriminator on the gradient update direction of the geological element classification model is realized.

[0090] In a specific embodiment, the step S40 includes the following steps:

[0091] S50, input the label encoding map and the prediction probability map into 5 4x4 convolution layers, and output a confidence map with the same size as the original image through a LeakyRelu activation function and an upsampling operation, wherein the number of convolution kernels of the 5 4x4 convolution layers is {64, 128, 256, 512, 1}.

[0092] In the embodiment, the discriminator is a full convolutional discriminator (FCD), and since there is no restriction of full connection layer, the size of the label encoding map and the prediction probability map input by the model can be arbitrary. After the label encoding map and the prediction probability map are input into 5 convolutional layers and activated by a LeakyReLU activation function with a coefficient of 0.2, a gray image composed of values between 0 and 1 is output. Finally, an confidence map with the same size as the input is output by upsampling, wherein the value of each pixel in the confidence map represents the meaning of confidence, instead of a normal scalar value. The present application trains the generator, the discriminator and the labeled data (sample images with labels) together, and optimizes the direction of gradient update of the segmentation network and improves the performance of the semantic segmentation network through continuous adversarial learning.

[0093] It is further explained that the current remote sensing interpretation based on supervised learning needs a large number of labeled samples, and the labeling of semantic segmentation samples is very difficult, especially in regions with many elements and blurred boundaries. In the geological remote sensing scene, it is inevitable that there are few labeled samples and many unlabeled samples. In order to fully utilize the unlabeled samples, improve the accuracy of semantic segmentation, and reduce the dependence of the model on labeled samples to a certain extent and increase the generalization ability of the model, in a specific embodiment of the present application, after the step S104, the following steps are included:

[0094] S60, input a large number of unlabeled sample images into the geological element classification model, and obtain corresponding prediction probability maps through a softmax function;

[0095] S61, input the prediction probability map into the discriminator to obtain a corresponding training confidence map, and set a threshold T semi to screen samples with high confidence in the confidence map for self-supervised learning;

[0096] S62, calculate the semi-supervised loss L between the prediction probability map and the corresponding training confidence map according to the following formula: semi :

[0097]

[0098] wherein I represents an indicator function, and if D(G(X n ) (h，w) ) > T semi is True, then I is 1, and if D(G(X n ) (h，w) ) > T semi is False, then I is 0, represents the encoding of the label encoding map, and if then

[0099] Therefore, the total loss function is:

[0100] L = L ce +λ adv L adv +λ semi L semi .

[0101] In this embodiment, the unlabeled sample images are used to obtain a predicted probability map G(X) through a geological element classification model. n ) (h，w，c) The predicted probability map G(X) n ) (h，w，c) As input to the discriminator, a confidence map D(G(X) is generated. n ) (h，w) ), by setting the threshold T semi Binarization is performed, and regions with high confidence are selected for self-supervised learning, which provides additional supervision information for the classification model, thereby further optimizing the classification model, as shown in step S62.

[0102] Since the values ​​of the indicator function and the one-hot encoded tag vector can be considered as constants, L semi It can be approximated as a cross-entropy loss.

[0103] In this embodiment, a semi-supervised adversarial training (SSAL) approach is adopted, which utilizes a large amount of unlabeled data for semi-supervised learning. This effectively solves the problem of insufficient samples in geological remote sensing interpretation, improves interpretation accuracy, and reduces the model's dependence on samples.

[0104] Through continuous calculation of the loss function described above, the corresponding hyperparameter: λ adv , λ semi T semi By continuously optimizing hyperparameters and the direction of gradient updates in the geological element classification model, an optimal geological element classification model is trained. The adversarial learning architecture forces the geological element classification model to maintain high-order consistency between the predicted probability map and the label encoding map, thereby preserving the spatial continuity of the geological element segmentation results and alleviating, to some extent, the problems of blurred lithological boundaries and fragmented interpretation results. Furthermore, the introduction of semi-supervised learning provides additional supervisory information to the geological element classification model, further optimizing it and improving the interpretation effect.

[0105] Therefore, in the process of discriminating the prediction probability map and the label encoding map, the discriminator is learning the high-order difference between the two, forcing the geological element classification model to maintain the high-order consistency between the two, thereby maintaining the spatial continuity of the geological element segmentation result. In addition, for unlabeled data, FCD also performs semi-supervised learning by setting a threshold to select pixels with high confidence.

[0106] In a specific embodiment, after the step S106, the following steps are included:

[0107] S70, post-processing the classification result by using a full connection CRF to obtain an optimized classification result.

[0108] In this embodiment, the full connection CRF (DenseCRF) uses a probability graph model to model the global spatial context information between each pixel and all other pixels, and then maximizes the posterior probability of the segmentation label by minimizing a predefined energy function, thereby performing fine segmentation and refinement on all pixels in the image.

[0109] That is, the present application smoothes the interpretation result of the geological element by using a full connection conditional random field, and alleviates the phenomenon of broken holes and isolated regions appearing in the prediction probability map.

[0110] The present application also provides a geological environment remote sensing interpretation device for executing any one of the embodiments of the geological environment remote sensing interpretation method. Specifically, please refer to Figure 2 , Figure 2 is a schematic block diagram of the geological environment remote sensing interpretation device provided by the embodiments of the present application.

[0111] As Figure 2 shown, the geological environment remote sensing interpretation device 500 includes:

[0112] The shallow image feature extraction unit 501 is configured to extract features of a sample image by using a ResNet network to obtain shallow image features.

[0113] The multi-scale geological element feature map extraction unit 502 is configured to aggregate multi-scale features of the shallow image features by using a multi-scale feature extraction module to obtain a multi-scale geological element feature map.

[0114] The global feature fusion unit 503 is configured to fuse global features of the multi-scale geological element feature map by using a global feature fusion module to output a prediction probability map, thereby constructing a geological element classification model.

[0115] The optimization unit 504 is configured to train and optimize the geological element classification model by using a loss function to obtain an optimized geological element classification model.

[0116] The self-supervised learning unit 505 is configured to input the unlabeled image into the optimized geological element classification model, and generate a pseudo-labeled image, and perform self-supervised learning on the geological element classification model by using the pseudo-labeled image.

[0117] The classification unit 506 is configured to classify the target image by using the optimized geological element classification model, and output a classification result.

[0118] The geological element classification model generated by the device has good interpretation / classification accuracy for geological images of all elements.

[0119] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0120] The above-mentioned geological environment remote sensing interpretation device can be realized in the form of a computer program, which can run on a computer device as shown in the figure. Figure 3 The computer device 1100 is a server, which can be a stand-alone server or a server cluster composed of multiple servers.

[0121] Please refer to Figure 3 , Figure 3 is a schematic block diagram of the computer device provided by the embodiment of the present application. The computer device 1100 is a server, which can be a stand-alone server or a server cluster composed of multiple servers.

[0122] Please refer to Figure 3 , the computer device 1100 includes a processor 1102, a memory and a network interface 1105 connected by a system bus 1101, wherein the memory can include a non-volatile storage medium 1103 and an internal memory 1104.

[0123] The non-volatile storage medium 1103 can store an operating system 11031 and a computer program 11032. The computer program 11032, when executed, can make the processor 1102 execute the geological environment remote sensing interpretation method.

[0124] The processor 1102 is configured to provide computing and control capabilities to support the operation of the entire computer device 1100.

[0125] The internal memory 1104 provides an environment for the execution of the computer program 11032 in the non-volatile storage medium 1103. The computer program 11032, when executed by the processor 1102, can make the processor 1102 execute the geological environment remote sensing interpretation method.

[0126] The network interface 1105 is configured to perform network communication, such as providing transmission of data information, etc. Those skilled in the art can understand that Figure 3 The structure shown in FIG. 11 is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device 1100 to which the scheme of the present application is applied. Specifically, the computer device 1100 can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0127] Those skilled in the art can understand that Figure 3 The embodiment of the computer device shown in FIG. 11 does not constitute a limitation on the specific structure of the computer device. In other embodiments, the computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. For example, in some embodiments, the computer device can only include a memory and a processor. In such embodiments, the structure and function of the memory and the processor are consistent with those of the memory 1101 and the processor 1102 shown in the embodiment, and will not be described here. Figure 3

[0128] It should be understood that, in the embodiments of the present application, the processor 1102 can be a central processing unit (CPU), and the processor 1102 can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic components, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0129] In another embodiment of the present application, a computer readable storage medium is provided. The computer readable storage medium can be a non-volatile computer readable storage medium. The computer readable storage medium stores a computer program, wherein the computer program is executed by a processor to implement the geological environment remote sensing interpretation method of the embodiments of the present application.

[0130] The storage medium is an entity, non-transient storage medium, for example, can be a U disk, a mobile hard disk, a read-only memory (ROM), a magnetic disk or an optical disk, etc. Various entity storage media that can store program codes.

[0131] ​Those skilled in the art can clearly understand the specific working process of the devices, apparatuses and units described above can refer to the corresponding process in the foregoing method embodiments, and will not be described here.

[0132] The above merely describes specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements shall be encompassed within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A method for remote sensing interpretation of geological environment, characterized in that, The method comprises the following steps: feature extraction of a sample image is performed by using a ResNet network to obtain shallow image features; multi-scale feature aggregation is performed on the shallow image features by using a multi-scale feature extraction module to obtain multi-scale geological element feature maps; global feature fusion is performed on the multi-scale geological element feature maps by using a global feature fusion module to output a prediction probability map, thereby constructing a geological element classification model; the geological element classification model is trained and optimized by using a loss function to obtain an optimized geological element classification model; a no-label image is input into the optimized geological element classification model to generate a pseudo-label image, and self-supervised learning of the geological element classification model is performed by using the pseudo-label image; the target image is classified by using the optimized geological element classification model, and a classification result is output; wherein, the multi-scale feature aggregation performed on the shallow image features by using the multi-scale feature extraction module to obtain the multi-scale geological element feature maps comprises: the shallow image features are input into a 1x1 convolution layer to obtain a first convolution feature map; the shallow image features are sequentially input into three 3x3 convolution layers to obtain corresponding second convolution feature maps, wherein the dilation rates of each 3x3 convolution layer are different; the shallow image features are sequentially input into a 3x3 convolution layer and a target semantic pooling layer to obtain a third convolution feature map; the first convolution feature map, the three second convolution feature maps and the third convolution feature map are spliced to obtain the multi-scale geological element feature maps; wherein, the sequential input of the shallow image features into the 3x3 convolution layer and the target semantic pooling layer to obtain the third convolution feature map comprises: the shallow image features are input into a 3x3 convolution layer to obtain a transition convolution feature map; the transition convolution feature map is input into a target semantic pooling layer, and the target semantic pooling layer performs a target semantic estimation step and a target semantic mapping step on the transition convolution feature map to obtain the third convolution feature map; the target semantic estimation step comprises: the class of each pixel in the transition convolution feature map is obtained, and the similarity degree of the current pixel and the pixels belonging to the same class is calculated according to the following formula to obtain a pixel similarity probability map: where x p represents the vector of pixel p, x i represents the vector of pixel i, f q represents the transformation of query, f k represents the transformation of key, Z p represents the feature similarity of the current pixel p with all the remaining pixels; the target semantic mapping step comprises: the feature similarity of all the remaining pixels is weighted and aggregated according to the following formula to obtain the target context information of the pixel: where c p represent the target context information of the pixel p, and φ represents the transformation of the input features into different feature spaces.

2. The geologic environmental remote sensing interpretation method of claim 1, wherein, the global feature fusion of the multi-scale geological element feature maps by using the global feature fusion module (GAM) to output the prediction probability map, thereby constructing the geological element classification model, comprises: the feature fusion of the multi-scale geological element feature maps is performed according to the following formula to output the prediction probability map F3: Wherein, F1∈R C×H×W Represent the multi-scale geological element feature map, F2 represents the intermediate feature map, M c Represent the channel attention map, M s Represent the spatial attention map.

3. The geologic environmental remote sensing interpretation method according to any one of claims 1-2, characterized in that, the training and optimization of the geological element classification model by using the loss function to obtain the optimized geological element classification model, comprises: the sample image with labels is converted into a label encoding image, and the label encoding image and the prediction probability map are input into a preset discriminator to output a corresponding confidence map; wherein, the loss function of the discriminator FCD is calculated according to the following formula: position, representing a confidence map corresponding to said predicted probability map, representing a confidence map corresponding to said label encoding map, and: The cross-entropy loss L between the predicted probability map and the label encoding map is calculated as follows ce : The adversarial loss L is calculated as follows adv : the loss function of the geological element classification model is constructed according to the following formula: L = L FCD + L ce + L adv .

4. The geologic environmental remote sensing interpretation method of claim 3, wherein, The prediction probability map corresponding to the label coding map and the sample image is input into a preset discriminator, and a corresponding confidence map is output. The label coding map and the prediction probability map are input into 5 4x4 convolution layers, and a confidence map with the same size as the original image is output through a LeakyRelu activation function and an upsampling operation, wherein the number of convolution kernels of the 5 4x4 convolution layers is {64, 128, 256, 512, 1}.

5. The geologic environmental remote sensing interpretation method of claim 3, wherein, The unlabeled image is input into the optimized geological element classification model, and a pseudo-label image is generated, and the pseudo-label image is used for self-supervised learning of the geological element classification model. A large number of unlabeled sample images are input into the geological element classification model, and a corresponding prediction probability map is obtained through a softmax function. inputting the prediction probability map into the discriminator to obtain a corresponding training confidence map, and setting a threshold T semi to screen samples with high confidence in the confidence map for self-supervised learning; The semi-supervised loss L between the prediction probability map and the corresponding training confidence map is calculated as follows semi : where I represents an indicator function, I is 1 if D(G(X n ) (h，w) ) > T semi is True, and I is 0 if D(G(X n ) (h，w) ) > T semi is False, represents a code of a label coding map, and if then 6. The geologic environmental remote sensing interpretation method of claim 1, wherein, After the target image is classified by using the optimized geological element classification model and a classification result is output, the following steps are included. The classification result is post-processed by using a fully connected CRF to obtain an optimized classification result.

7. A geological environment remote sensing interpretation device that executes the geological environment remote sensing interpretation method according to any one of claims 1 to 6, characterized by, It includes: a shallow image feature extraction unit configured to extract features of a sample image by using a ResNet network to obtain shallow image features; a multi-scale geological element feature map extraction unit configured to aggregate multi-scale features of the shallow image features by using a multi-scale feature extraction module to obtain a multi-scale geological element feature map; a global feature fusion unit configured to fuse global features of the multi-scale geological element feature map by using a global feature fusion module to output a prediction probability map, thereby constructing a geological element classification model; an optimization unit configured to train and optimize the geological element classification model by using a loss function to obtain an optimized geological element classification model; a self-supervised learning unit configured to input an unlabeled image into the optimized geological element classification model to generate a pseudo-label image, and use the pseudo-label image for self-supervised learning of the geological element classification model; a classification unit configured to classify a target image by using the optimized geological element classification model and output a classification result; The multi-scale geological element feature map extraction unit includes: input the shallow image features into a 1x1 convolution layer to obtain a first convolution feature map; input the shallow image features into three 3x3 convolution layers in sequence to obtain corresponding second convolution feature maps, wherein the hole rates of each 3x3 convolution layer are different; input the shallow image features into a 3x3 convolution layer and a target semantic pooling layer in sequence to obtain a third convolution feature map; concatenate the first convolution feature map, the three second convolution feature maps, and the third convolution feature map to obtain a multi-scale geological element feature map; input the shallow image features into a 3x3 convolution layer and a target semantic pooling layer in sequence to obtain a third convolution feature map, which includes: input the shallow image features into a 3x3 convolution layer to obtain a transition convolution feature map; inputting the transition convolution feature map into a target semantic pooling layer, and causing the target semantic pooling layer to perform a target semantic estimation step and a target semantic mapping step on the transition convolution feature map to obtain a third convolution feature map; the target semantic estimation step comprises: obtaining the class of each pixel in the transition convolution feature map, and calculating the similarity degree of a current pixel and pixels belonging to the same class as the current pixel according to the following formula to obtain a pixel similarity probability map: where x p represents the vector of pixel p, x i represents the vector of pixel i, f q represents the transformation of query, f k represents the transformation of key, Z p represents the feature similarity of the current pixel p with all the remaining pixels; the target semantic mapping step comprises: performing weighted aggregation on the feature similarity of all remaining pixels according to the following formula to obtain the target context information of the pixel: where c p φ represents the target context information of the pixel p, and φ represents the conversion of the input features into different feature spaces.

8. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the geological environment remote sensing interpretation method of any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program causes the processor to execute the geological environment remote sensing interpretation method of any one of claims 1-7 when executed by the processor.

Images