Intelligent identification method and device for alluvial-proluvial fan deposition by remote sensing

Abstract

The application discloses a kind of alluvial fan deposition remote sensing intelligent identification method and device, the method includes: synthesis optical remote sensing image and topographic remote sensing image to obtain synthesis remote sensing image, and according to the synthesis remote sensing image production alluvial fan target sample data set;MASK R-CNN deep convolutional neural network is constructed, and the structure parameter of the network is adjusted according to the feature of the synthesis remote sensing image;According to the alluvial fan target sample data set and the MASK R-CNN deep convolutional neural network, model training and model testing are carried out;The model of training completion is packaged to obtain alluvial fan identification model.The present application fully exploits the spectral information and topographic information of alluvial fan deposition satellite remote sensing, establishes alluvial fan remote sensing intelligent identification model, quickly and accurately identifies alluvial fan deposition body, and provides basis for modern coarse-grained fan body deposition rule summary and fine description.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image intelligent recognition and modern sedimentology, in particular to a diluvial fan sediment remote sensing intelligent recognition method and device. BACKGROUND

[0002] Coarse fan body deposition has been an important reservoir in lithologic oil and gas reservoirs, and coarse fan body deposition is an important exploration object in current continental basins, so the research on the characteristics of onshore diluvial fan body deposition becomes very important. The research on modern deposition for ancient research is an important research idea of sedimentology, and the research on the distribution characteristics and rules of modern diluvial fan deposition has practical and theoretical significance for guiding the oil and gas exploration and development of coarse fan reservoirs in oil and gas basins and perfecting the theory of diluvial fan.

[0003] At present, the main research methods of modern diluvial fan deposition are sediment investigation and simulation experiment, the traditional modern sediment investigation has the defect of limited observation range, and the water tank and numerical simulation experiment are difficult to restore the real situation of diluvial fan deposition due to the limitation of simulation parameters. Remote sensing technology can make up for the defects of sediment investigation and simulation experiment with its macro imaging and detail highlighting advantages of earth observation, and becomes a new technical means for the research of modern diluvial fan deposition. How to accurately and quickly identify the diluvial fan deposition target under the support of large-scale multi-dimensional remote sensing information provides a research basis for the spatial distribution identification and sediment rule summary of modern diluvial fan deposition, which is the key and difficulty of modern sediment remote sensing observation research.

[0004] Generally, diluvial fans of different periods and different sources can show differences in particle size and color tone of fan gravel, on the other hand, the fan body is affected by topographic slope and gully erosion under cutting, and shows different degrees of dispersion characteristics. These characteristics can be reflected in various geomorphic parameter indexes extracted based on remote sensing topographic data, so these different period and different source diluvial fan geomorphic units can be identified and divided by combining remote sensing optical and topographic multiple indexes. At present, the methods widely used for remote sensing identification of diluvial fan at home and abroad mainly include traditional visual discrimination method and remote sensing spectral index analysis method: in 2001, Qiao Yanxiao analyzed and summarized the image characteristics of diluvial fan; in 2006, Cui Weiguo carried out the research on the evolution of Manasi River alluvial fan by using remote sensing interpretation method; in 2009, Hardgrove proposed a remote sensing classification method based on the thermal imaging characteristics of alluvial fan; in 2013, Yang Shuwen carried out the research on the extraction method of alluvial fan by using the spectral characteristics of ETM image. The previous research on modern diluvial fan mainly relies on the experience-based visual interpretation, and the existing remote sensing identification method of diluvial fan only stays at the level of simple establishment of spectral index, without fully utilizing the topographic feature differences of diluvial fan and without deeply mining the remote sensing image characteristics of diluvial fan, so the accuracy and efficiency of identification are not ideal. SUMMARY

[0005] The present application proposes a method and device for intelligent recognition of alluvial-proluvial fan deposition by remote sensing in order to solve at least one of the technical problems in the background art.

[0006] In order to achieve the above-mentioned purpose, according to one aspect of the present application, a method for intelligent recognition of alluvial-proluvial fan deposition by remote sensing is provided, which comprises:

[0007] synthesizing the optical remote sensing image and the terrain remote sensing image to obtain a synthesized remote sensing image, and preparing an alluvial-proluvial fan target sample data set according to the synthesized remote sensing image;

[0008] constructing a MASK R-CNN deep convolutional neural network, and adjusting structural parameters of the network according to features of the synthesized remote sensing image;

[0009] performing model training and model testing according to the alluvial-proluvial fan target sample data set and the MASK R-CNN deep convolutional neural network;

[0010] packaging the trained model to obtain an alluvial-proluvial fan recognition model, so as to perform alluvial-proluvial fan information recognition according to the alluvial-proluvial fan recognition model.

[0011] Optionally, the optical remote sensing image comprises three RGB bands; and the terrain remote sensing image comprises three bands of digital elevation model, slope and aspect.

[0012] The synthesized optical remote sensing image and the terrain remote sensing image to obtain a synthesized remote sensing image comprises:

[0013] increasing three bands of digital elevation model, slope and aspect of the terrain remote sensing image containing terrain information on the basis of the three bands of RGB of the optical remote sensing image, to generate a synthesized remote sensing image containing six-band features of spectrum and terrain information.

[0014] Optionally, the preparing of the alluvial-proluvial fan target sample data set according to the synthesized remote sensing image comprises:

[0015] annotating alluvial-proluvial fan target contour vectors in the synthesized remote sensing image;

[0016] cutting the synthesized remote sensing image into a plurality of image blocks according to a preset size;

[0017] performing intersection processing on all the image blocks and the alluvial-proluvial fan target contour vectors, and extracting alluvial-proluvial fan target contour vectors corresponding to each image block to form a contour vector small image vector set.

[0018] Optionally, the adjusting of the structural parameters of the network according to the features of the synthesized remote sensing image comprises:

[0019] On the basis of the basic network structure, the input channel is changed from RGB three bands to six bands of the synthetic remote sensing image.

[0020] Optionally, the model training and model testing according to the alluvial fan target sample data set and the MASK R-CNN deep convolutional neural network comprises:

[0021] According to the test sample set in the alluvial fan target sample data set, the trained model is tested to obtain the sample recognition accuracy of the trained model;

[0022] If the sample recognition accuracy is greater than a preset threshold, the model training is stopped, and the current trained model is taken as the trained model.

[0023] In order to achieve the above-mentioned purpose, according to another aspect of the present application, an alluvial fan deposition remote sensing intelligent recognition device is provided, which comprises:

[0024] A sample data set generation unit is configured to synthesize optical remote sensing images and terrain remote sensing images to obtain synthetic remote sensing images, and to make an alluvial fan target sample data set according to the synthetic remote sensing images;

[0025] A neural network construction unit is configured to construct a MASK R-CNN deep convolutional neural network, and to adjust the structure parameters of the network according to the characteristics of the synthetic remote sensing images;

[0026] A model training and testing unit is configured to train and test a model according to the alluvial fan target sample data set and the MASK R-CNN deep convolutional neural network;

[0027] A model packaging unit is configured to package the trained model to obtain an alluvial fan recognition model, so as to recognize alluvial fan information according to the alluvial fan recognition model.

[0028] In order to achieve the above-mentioned purpose, according to another aspect of the present application, a computer device is also provided, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps in the alluvial fan deposition remote sensing intelligent recognition method when executing the computer program.

[0029] In order to achieve the above-mentioned purpose, according to another aspect of the present application, a computer readable storage medium is also provided, which stores a computer program, wherein the computer program implements the steps in the alluvial fan deposition remote sensing intelligent recognition method when executed in a computer processor.

[0030] The beneficial effects of the present application are: the present application obtains a synthetic remote sensing image by synthesizing an optical remote sensing image and a terrain remote sensing image, and according to the synthetic remote sensing image, a target sample data set of a diluvial fan is made, and then a diluvial fan recognition model is trained based on a MASK R-CNN deep convolutional neural network to recognize diluvial fan information, which effectively improves the recognition accuracy and recognition efficiency compared with the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor. In the drawings:

[0032] Figure 1 is a flow chart of the diluvial fan sediment remote sensing intelligent recognition method of the embodiment of the present application;

[0033] Figure 2 is a technical route map of the diluvial fan sediment remote sensing intelligent recognition method of the embodiment of the present application;

[0034] Figure 3 is an anchor box (Anchor) generation rule diagram in the embodiment of the present application;

[0035] Figure 4 is a research area one diluvial fan extraction result diagram in the embodiment of the present application;

[0036] Figure 5 is a research area two diluvial fan extraction result diagram in the embodiment of the present application;

[0037] Figure 6 is a structural block diagram of the diluvial fan sediment remote sensing intelligent recognition device in the embodiment of the present application;

[0038] Figure 7 is a computer device diagram in the embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the person skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0040] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems, or computer program products. Accordingly, the application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code thereon.

[0041] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification and in the claims and the above drawings are intended to cover both the exclusive and the non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a list of steps or units does not necessarily include only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to such processes, methods, products, or apparatus.

[0042] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0043] The present application fully taps the spectral information and topographic information of satellite remote sensing, introduces remote sensing intelligent analysis technology into the field of modern sedimentology, and proposes a new automatic recognition method for modern alluvial fan deposition, which provides protection for large-area alluvial fan deposition investigation, provides basis for analyzing the spatiotemporal distribution rule of modern alluvial fan under the same climate condition and geomorphic environment, and summarizing and finely depicting the deposition characteristics of coarse fan body under specific geological conditions. It is helpful for geologists to carry out scientific and systematic theoretical research on modern deposition, and to establish a useful sedimentary model for oil and gas basin coarse fan reservoir exploration and development under the comparison between present and past, thereby improving the productivity and efficiency of oilfield exploration and development.

[0044] Figure 1 is a flow chart of the alluvial fan deposition remote sensing intelligent recognition method of the embodiment of the present application, Figure 2 is a technical route map of the alluvial fan deposition remote sensing intelligent recognition method of the embodiment of the present application, as shown in Figure 1 and Figure 2 The alluvial fan deposition remote sensing intelligent recognition method of the embodiment includes steps S1 to S4.

[0045] Step S1: Synthesizing optical remote sensing images and topographic remote sensing images to obtain a synthetic remote sensing image, and making an alluvial fan target sample data set according to the synthetic remote sensing image.

[0046] Step S2: Constructing a MASK R-CNN deep convolutional neural network, and adjusting the structure parameters of the network according to the characteristics of the synthetic remote sensing image.

[0047] Step S3, model training and model testing are performed according to the alluvial-proluvial fan target sample data set and the MASK R-CNN deep convolutional neural network.

[0048] Step S4, an alluvial-proluvial fan recognition model is obtained by packaging the trained model, and alluvial-proluvial fan information recognition is performed according to the alluvial-proluvial fan recognition model.

[0049] In the optional embodiment of the present application, the scheme is introduced based on two research areas. The two research areas of the present application are located in Junggar of Xinjiang and Qilian Mountain area of Gansu, respectively. The western mountainous area is dry and less rainy, and the alluvial-proluvial fan is relatively developed. Remote sensing technology can be used to carry out remote sensing recognition of the alluvial-proluvial fan.

[0050] The first research area is located between 84°23'~85°50' east longitude and 45°~46°26' north latitude, is located in the northwest of Junggar Basin, and is mostly Yardan landform, belonging to a typical temperate continental climate. The temperature difference between summer and winter is large, it is dry and less rainy, there are more spring and autumn winds, the temperature difference between winter and summer is large, the snow cover is thin, the evaporation is fast, and the frozen soil is deep. The research area is on the concealed thrust fault zone from Kuitun in the west to Xiaozijie in the east. The formation of the fault zone is due to the subduction, reduction and even collision of the oceanic crust of Junggar-Turfan plate to the Kazakhstan plate in the late Paleozoic, so that the area becomes a collision uplift belt and a collision foreland type sedimentary depression adjacent to the uplift belt.

[0051] The second research area is located between 95°29'~95°48' east longitude and 39°4'~39°18' north latitude, is located in the Nanshan area of Subei Mongol Autonomous County, is located in the south side of the west edge of Qilian Mountain and the west end of Hexi Corridor, is located in the northeast edge of Qinghai-Tibet Plateau, and belongs to the inland river basin of Hexi. The high mountain area of the west section of Qilian Mountain has high terrain, high mountains, deep valleys and intermontane basins. The Nanshan area is located in the middle latitude and is deep in the Eurasian continent, and is less affected by the southward monsoon, being a typical inland alpine desert steppe climate. The west section of the northern edge of Qilian Mountain where the research area is located is located in the junction zone of Qinghai-Tibet Plateau block, Tarim block and Alashan block.

[0052] In the embodiment of the present application, the optical remote sensing image is a medium-high resolution remote sensing image containing red, green and blue visible light bands, and the terrain remote sensing image is a DEM digital elevation model. The synthetic remote sensing image obtained by synthesizing the optical remote sensing image and the terrain remote sensing image in step S1 is obtained by synthesizing bands containing target optical features and bands containing terrain features, and finally six-band image data is synthesized, that is, on the basis of the RGB three bands of the optical remote sensing image, three bands containing terrain information of the terrain remote sensing image, namely digital elevation model, slope and aspect, are added to generate a synthetic remote sensing image containing six-band features of spectrum and terrain information. Here, the LANDSAT7 ETM data of the research area in 2001 is selected, with a resolution of 30M, and the red (RED), green (GREEN) and blue (BLUE) bands are selected as the three bands of optical features; the digital elevation data DEM of GDEMV2 is selected as the terrain data, with a resolution of 30M, and the digital elevation model DEM, slope SLOP and aspect ASPECT bands are further calculated from the DEM digital elevation data as the three bands of terrain features.

[0053] In the embodiment of the present application, the slope SLOP and the aspect ASPECT can be calculated according to the digital elevation DEM. The slope is calculated as the maximum rate of change in the direction from each pixel to its neighboring pixel, and the smaller the slope value, the flatter the terrain; the larger the slope value, the steeper the terrain; the aspect is calculated as the downhill direction with the maximum rate of change in the direction from each pixel to its neighboring pixel, and the aspect can be regarded as the slope direction, and the calculated value of the slope can indicate the compass direction of the surface orientation at each pixel position, which will be measured in a clockwise direction, with an angle range of 0 (due north) to 360 (still due north), that is, a complete circle, and the flat area without a downhill direction is assigned a value of -1.

[0054] In the embodiment of the present application, the step S1 of preparing the alluvial fan target sample data set according to the synthetic remote sensing image is to label the target in the range of the synthetic remote sensing image, interpret the target contour vector, cut the synthetic image according to the target contour vector to form square image blocks of the same size, and generate a sample data set, which is divided into a training set, a validation set and a test set. In an optional embodiment of the present application, the step S1 of preparing the alluvial fan target sample data set according to the synthetic remote sensing image can specifically include the following steps: 1. target labeling is performed in the range of the synthetic image to obtain alluvial fan target contour vectors shape; 2. the synthetic image is cut to form square image blocks of multiple scales such as Figure 33. Traverse all image patches and perform intersection processing with the contour vector shape to extract the target contour vector shape corresponding to each image patch intersection to form a contour vector small graph vector set. The purpose is to reduce the operation load during model training and improve the training speed. A large number of square image blocks and contour vector small graphs form a sample data set. In an optional embodiment of the present application, 2000 target alluvial fans in northwest China are labeled in total, and 2260 multi-scale square image blocks are cut. Based on the above image blocks and vector sets, a sample data set is made, and the point information in the vector set is re-encoded into a json format for storage. Making a sample data set is divided into a model training sample set, a verification set for fine-tuning parameters in the training process, and a test set for testing the model. A small amount of samples are reserved as a test set, the remaining samples are randomly stored as a training set and a verification set in a ratio of 9:1, and finally a sample set of the training set, the verification set and the test set is formed, wherein the training samples are 1853, the verification samples are 206, and the test samples are 200.

[0055] In the embodiment of the present application, the MASK R-CNN deep convolutional neural network of step S2 is constructed, and first, a software environment is configured, for example: Tensorflow = 1.15, keras = 2.2.5, python = 3.7, etc. The structure parameters of the network are adjusted according to the characteristics of the synthetic remote sensing image in step S2. On the basis of the basic network structure, the input channel is changed from RGB three bands to six bands of the synthetic image. Specifically, the first convolutional layer of Backbone is modified to adapt to the change of input data, so that the data can smoothly enter the model.

[0056] In the embodiment of the present application, the basic structure of the MASK R-CNN deep convolutional neural network is Backbone, RPN network, Align layer pooling layer and head layer from bottom to top.

[0057] The backbone includes ResNet and FPN network. ResNet uses a residual network to extract image features, solves the problem of gradient disappearance and gradient explosion in the deep network during the backward propagation process due to the chain rule in the differentiation process, can train a deeper convolutional neural network, and this time adopts ResNet 101 network; FPN (feature pyramid networks) is a feature pyramid network, including a forward propagation process of a bottom-up neural network. The feature map is calculated by a convolution kernel, which usually becomes smaller and smaller; the other is from top to bottom, which up-samples the high-level feature map with more abstract and semantic features, and then connects the feature horizontally to the previous layer feature. Therefore, the high-level feature is enhanced, and the Feature Map used for prediction in each layer integrates features with different resolutions and different semantic strengths, which can detect objects with corresponding resolution sizes and ensure that each layer has appropriate resolution and strong semantic features.

[0058] The RPN (Region Proposal Network) network is used for foreground and background classification and candidate frame extraction with classification information. For each position of the feature image, 9 possible candidate windows (anchors) are considered. The foreground and background classification assigns a positive label to the anchor with the highest IoU overlap with a certain ground truth (GT) bounding box or to the anchor with an IoU overlap greater than a certain value with any GT bounding box, and assigns a negative label to the other.

[0059] The ROI Align pooling layer is a pooling layer that outputs the same size and accurate coordinate results. The bilinear interpolation method is used to obtain the image value of the pixel point with floating point coordinates, so as to convert the entire feature aggregation process into a continuous operation network.

[0060] The head layer is three parallel outputs, including the head part of the class, bound box, and mask branch. The class classification, regression, and mask classification of the MASK R-CNN network include the final classes and bounding boxes of object detection. This part classifies and regresses the full connection results of ROI. The prediction of the mask is also after the ROI, and the FCN (Fully Convolution Network) generates a mask with the same size as the class classification result. By traversing all the predicted masks, the class_id class in the real mask is found by comparing the pixel point information of the real mask and the predicted mask.

[0061] In the embodiment of the present application, step S2 adjusts the structure parameters of the network according to the characteristics of the synthetic remote sensing image, which is based on the basic network structure, and the input channel is changed from RGB three bands to six bands of the synthetic image. The first convolutional layer of Backbone is modified to adapt to the change of input data, so that the data can smoothly enter the model. The MASK R-CNN source code is modified, the image loading method is rewritten, the array dimension in the model building function is changed, the model weight loading function and the training function are modified, the modification of the first convolutional layer is completed to adapt to the change of input data, and the input channel is changed from three channels of RGB to six channels of the synthetic image.

[0062] In the embodiment of the present application, step S3 trains and tests the model according to the alluvial fan target sample data set and the MASK R-CNN deep convolutional neural network. The training sample enters the improved MASK R-CNN model, and sequentially experiences the four modules of Backbone, RPN network, ROI Align pooling layer and head layer for training.

[0063] The synthetic image training sample picture is subjected to convolution feature extraction in the ResNet 101 of Backbone, and the ResNet 101 network has five convolutional layers, each of which is subjected to three steps of feature extraction, i.e. convolution, batch normalization BN and activation function RELU. The set parameters are shown in Table 1, and different scale Feature Maps C1, C2, C3, C4 and C5 are obtained.

[0064]

[0065]

[0066] Table 1

[0067] The Feature Map enters the FPN network, the FPN network uses C2, C3, C4 and C5 to establish a pyramid structure of feature maps, and obtains the Feature Map--[P2, P3, P4, P5, P6] after fusion of each scale through convolution and pooling fusion operations, wherein P6 is obtained by performing maximum pooling operation on P5 at a certain step length.

[0068] The feature maps [P2, P3, P4, P5, P6] enter the RPN network, which is a network for extracting candidate boxes. Based on the feature maps [P2, P3, P4, P5, P6], three aspect ratios of SCALES=(32, 64, 128, 256, 512) and RATIOS=[0.5, 1, 2] are used for each layer. For each pixel point, 9 Anchors of different sizes are generated, for example, three aspect ratios of 32 for the first layer [16x16, 32x32, 64x64, 16x32, 32x16, 16x64, 64x16, 32x64, 64x32], as shown in the Anchor as Figure 3 The five feature layers P2 to P6 are traversed to generate a large number of Anchors, and a certain number of candidate boxes are extracted to reduce the load of the machine and enter the next step. Next, the selected Anchors are marked according to certain rules, for example, setting the threshold of IoU>0.7 as foreground value 1 and IoU<0.7 as background value 0, and saving the offset between the Anchors and the ground truth.

[0069] The forward propagation calculates the classification score (probability) and the coordinate point offset. In the classification and regression of the RPN network, each Anchor in each layer is divided into two categories of background and foreground, and four displacement amounts are regressed. The score (probability) of each Anchor as foreground or background needs to be calculated, and the same classification and regression operations are performed on all layers from P2 to P6.

[0070] The RPN network loss value is calculated and the weights are updated by back propagation. In the RPN network, the classification and regression information of Anchors is saved. The RPN classification uses a cross-entropy loss function based on the Softmax function, and the cross-entropy loss value is calculated by the scores of the positive and negative samples and the labels of the positive and negative samples. The RPN regression uses a SmoothL1Loss loss function, and the loss value is calculated by the offset calculated by the forward propagation and the offset calculated between the positive sample and the ground truth. The results obtained by the above two functions are used to guide the parameter update of the next classification and regression, and the updated weights are used to realize back propagation.

[0071] The ROI is generated according to the RPN adjustment. The scores of the foreground (or positive samples) calculated by the forward propagation of the Anchors through the RPN network are sorted from high to low, and a certain amount of Anchors with the highest scores are taken out. The offsets of the Anchors calculated by the forward propagation of the Anchors through the RPN network are added to the Anchor boxes to obtain more accurate box coordinates. Finally, before returning, a non-maximum suppression (NMS) operation is performed to remove duplicate boxes.

[0072] ROI Align is a regional feature aggregation method that uses a bilinear interpolation method to obtain image values at pixel points with floating-point coordinates, thereby converting the entire feature aggregation process into a continuous operation to avoid coordinate errors caused by quantization.

[0073] The class classification, regression, and mask classification of the head layer. The class classification and regression of the head layer are the same as those in the RPN network, and the loss functions are also based on Softmax cross-entropy and SmoothL1Loss. However, the RPN network only classifies foreground (positive class) and background (negative class), while the classification in the head layer is specific to a certain class (multi-class classification). The mask classification aligns the mask branch to a 14*14 size Feature Map in the ROI alignment operation, and the RPN network saves the class_id corresponding to each positive sample mask mask area. Specific implementation: 1. Forward propagation: transform the 14*14 Feature Map through deconvolution to obtain a mask area (28*28) corresponding to each class, i.e., the predicted mask; 2. Save each positive sample mask (28*28) and the real class class_id of each mask area in the RPN network, i.e., the real mask; 3. Traverse all predicted masks by the class class_id in the real mask, find the predicted mask corresponding to the class class_id, and compare the real mask with the predicted mask through the binary cross-entropy loss function (binary_cross_entropy). Only when the predicted probability is equal to the real label, the loss is 0, otherwise the loss is a positive number. Moreover, the greater the difference between the probability, the greater the loss, so the weight is changed according to the loss value to realize back propagation.

[0074] During the training process of all sample sets, the model output and the training sample label are continuously compared, and the weight parameters are continuously adjusted to adapt to the optimization of the model output result, and finally a MASK R-CNN model containing specific network weight parameters is trained.

[0075] In the embodiment of the present application, the training of the MASK R-CNN model is a process of training of samples and constant adjustment of network weight parameters: (1) the training samples enter the MASK R-CNN model, are subjected to feature extraction in the ResNet network, generate the regions of predicted ROIs in the RPN network, are subjected to feature pyramid integration in the FPN network to complete fusion of multi-scale features, and in the process that the RGB-D training samples enter the MASK R-CNN model, are subjected to feature extraction in the ResNet network in turn, are subjected to pooling in the Align Layer network to obtain results of the same size and keeping accurate coordinates, and finally enter two routes in parallel, one of which is subjected to feature mapping to the same size as the original picture in the FCN network to output the Mask result, and the other of which is subjected to classification and bound box regression. (2) in the process that all sample sets participate in the training, the network weight parameters are constantly adjusted according to the constant comparison between the model output and the training sample label to adapt to the optimization of the model output result, and finally the MASK R-CNN model containing specific network weight parameters is trained.

[0076] In the embodiment of the present application, the improved MASK R-CNN model is also used for precision evaluation of the test samples in step S3. Indicators such as accuracy (Accuracy), precision (Precision), recall (Recall) and mean intersection over union (mIoU) are used to quantitatively evaluate the extraction results of the model. The accuracy refers to the proportion of the positive examples extracted from all positive and negative examples, the precision refers to the proportion of the positive examples in all extracted results, and the recall refers to the proportion of the positive examples extracted from all positive examples. The formulas of the accuracy, the precision and the recall are as follows:

[0077] Accuracy = (TP + TN) / (FP + TP + FN + TN)

[0078] Precision = TP / (FP + TP)

[0079] Recall = TP / (FN + TP)

[0080] Wherein, TP is the number of correctly extracted positive examples, FP is the number of negative examples incorrectly extracted as positive examples, FN is the number of positive examples not extracted, and TN is the number of negative examples not extracted. The confusion matrix of the predicted results and the true values is shown in Table 2.

[0081]

[0082] Table 2

[0083] The mean intersection over union (mIoU) is a standard metric for semantic segmentation, which calculates the ratio of the intersection and union of two sets, the ground truth and the predicted segmentation in the problem of semantic segmentation. The greater the value of the mean intersection over union, the higher the accuracy. The formula for calculating the mean intersection over union is:

[0084]

[0085] Where TP is the intersection of the true value and the predicted value, that is, the number of correctly extracted positive examples; FP+FN+TP is the union of the true value and the predicted value.

[0086] According to the recognition results of the improved MASK R-CNN network model on the verification set, the confusion matrix information of each pixel point is obtained by comparing the vector contour interpreted by artificial interpretation, and the accuracy, precision, recall rate and mean intersection over union are calculated, the evaluation of the floodplain recognition model is completed, and when the recognition accuracy of the verification set is greater than the preset threshold, the model training is completed.

[0087] In the embodiment of the application, the test of the MASK R-CNN model is to calculate the accuracy of the training MASK R-CNN model in identifying samples, and the test average accuracy greater than the preset threshold (for example, 80%) is considered as successful model training: (1) input the sample to be tested, input the test picture into the trained model to obtain the final head layer recognition result, including target type recognition result, target bounding box and target contour vector; (2) evaluate the recognition result, FN: judged as negative sample, but in fact it is a positive sample; FP: judged as positive sample, but in fact it is a negative sample; TN: judged as negative sample, and in fact it is also a negative sample; TP: judged as positive sample, and in fact it is also a positive sample; On this basis, the accuracy A=(TP+TN) / (FP+TP+FN+TN) and the precision P=TP / (TP+FP) are calculated, the recall rate R=TP / (TP+FN) and the mean intersection over union mIoU are calculated. When the accuracy of the test set recognition is greater than 80%, the model training is successful.

[0088] In the embodiment of the application, the model packaging of step S4 is to package the MASK R-CNN model file containing the network weight parameters obtained by training steps S2 and S3 through the support of the software and hardware environment, to form a floodplain intelligent recognition tool (model). The tool (model) is used to realize the intelligent automatic recognition of the contour information of the target area remote sensing terrain synthesis image floodplain, and the automatic recognition results of the floodplain of the study area one and the study area two are as shown in Figure 4 and Figure 5 The recognition accuracy evaluation results are shown in Table 3.

[0089] Study Area Accuracy Precision Recall mIoU Study Area One 0.966848519 0.904295363 0.937914661 0.853221598 Study Area Two 0.952508944 0.895966066 0.935637394 0.843950358

[0090] Table 3

[0091] As can be seen from the above embodiments, the present application provides an effective, operable modern alluvial-proluvial fan deposition remote sensing intelligent identification method, which greatly improves the scale, timeliness and accuracy of modern deposition analysis. Compared with the previous remote sensing research techniques and methods of modern alluvial-proluvial fan deposition, the present application has two significant features: 1) on the basis of the existing methods of alluvial-proluvial fan remote sensing identification, the intelligent analysis technology is first introduced innovatively, and the alluvial-proluvial fan intelligent identification model based on the MASK R-CNN deep convolutional neural network is constructed, so that it is transformed into a fast and accurate technical means for specific regional modern alluvial-proluvial fan deposition investigation; 2) the present application innovatively proposes to synthesize the terrain information and the remote sensing spectral information, and adds a new band containing spatial information on the basis of the RGB three bands, which makes up for the defect that the remote sensing spectral data is difficult to reflect the topographic slope characteristics of the alluvial-proluvial fan, fully excavates the spectral and three-dimensional spatial characteristics of the alluvial-proluvial fan, and improves the model recognition credibility.

[0092] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown.

[0093] Based on the same inventive concept, the present application also provides an alluvial-proluvial fan deposition remote sensing intelligent identification device, which can be used to implement the alluvial-proluvial fan deposition remote sensing intelligent identification method described in the above embodiments, as described in the following embodiments. Since the principle of solving problems of the alluvial-proluvial fan deposition remote sensing intelligent identification device is similar to that of the alluvial-proluvial fan deposition remote sensing intelligent identification method, the embodiments of the alluvial-proluvial fan deposition remote sensing intelligent identification device can refer to the embodiments of the alluvial-proluvial fan deposition remote sensing intelligent identification method, and the repeated parts will not be described herein. The term "unit" or "module" used below can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware or a combination of software and hardware is also possible and is conceived.

[0094] Figure 6 is a structural block diagram of the alluvial-proluvial fan deposition remote sensing intelligent identification device of the present application embodiment, as Figure 6 shown, the alluvial-proluvial fan deposition remote sensing intelligent identification device of the present application embodiment comprises:

[0095] a sample data set generation unit, configured to synthesize the optical remote sensing image and the terrain remote sensing image to obtain a synthesized remote sensing image, and to make an alluvial-proluvial fan target sample data set according to the synthesized remote sensing image;

[0096] a neural network construction unit, configured to construct a Mask R-CNN deep convolutional neural network and adjust structural parameters of the network according to features of the synthetic remote sensing image;

[0097] a model training and testing unit, configured to perform model training and model testing according to the alluvial fan target sample data set and the Mask R-CNN deep convolutional neural network;

[0098] a model packaging unit, configured to package the trained model to obtain an alluvial fan identification model, so as to perform alluvial fan information identification according to the alluvial fan identification model.

[0099] In an embodiment of the present application, the optical remote sensing image comprises RGB three bands; the terrain remote sensing image comprises digital elevation model, slope and aspect three bands. The sample data set generation unit comprises:

[0100] an image synthesis module, configured to add the digital elevation model, slope and aspect three bands containing terrain information of the terrain remote sensing image on the basis of the RGB three bands of the optical remote sensing image, to generate a synthetic remote sensing image containing six-band features of spectral and terrain information.

[0101] In an embodiment of the present application, the sample data set generation unit comprises:

[0102] a labeling module, configured to label the alluvial fan target contour vector in the synthetic remote sensing image;

[0103] a cutting module, configured to cut the synthetic remote sensing image into a plurality of image blocks according to a preset size;

[0104] a contour vector small image vector set extraction module, configured to perform intersection processing on all image blocks and the alluvial fan target contour vector, and extract the alluvial fan target contour vector corresponding to the intersection of each image block to form a contour vector small image vector set.

[0105] In an embodiment of the present application, the neural network construction unit comprises:

[0106] a structural parameter adjustment module, configured to change the input channel from RGB three bands to six bands of the synthetic remote sensing image on the basis of the basic network structure.

[0107] In an embodiment of the present application, the model training and testing unit comprises:

[0108] a sample recognition accuracy calculation module, configured to test the trained model according to a test sample set in the alluvial fan target sample data set, to obtain a sample recognition accuracy of the trained model;

[0109] The model training completion module is configured to stop model training when the sample recognition accuracy is greater than a preset threshold, and use the current trained model as a trained model.

[0110] To achieve the above object, according to another aspect of the present application, a computer device is further provided. As shown in the figure, the computer device comprises a memory, a processor, a communication interface and a communication bus, wherein the memory stores a computer program executable on the processor, and the processor implements the steps in the above-mentioned embodiment method when executing the computer program. Figure 7

[0111] The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, or combinations thereof.

[0112] The memory is a non-transitory computer readable storage medium, which can be used to store non-transitory software programs, non-transitory computer executable programs and units, such as the corresponding program units in the above-mentioned method embodiments. The processor executes various functions and work data processing of the processor by running the non-transitory software programs, instructions and modules stored in the memory, that is, implements the method in the above-mentioned method embodiments.

[0113] The memory can include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required by a function; the data storage area can store data created by the processor and the like. In addition, the memory can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory remotely arranged with respect to the processor, and these remote memories can be connected to the processor through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network and combinations thereof.

[0114] The one or more units are stored in the memory and executed by the processor to implement the method in the above-mentioned embodiments.

[0115] ​The above computer device specific details can be understood in correspondence with the relevant description and effects in the above embodiments, and will not be repeated here.

[0116] To achieve the above object, according to another aspect of the present application, a computer readable storage medium is provided, which stores a computer program, and the computer program, when executed in a computer processor, implements the steps in the above alluvial fan deposition remote sensing intelligent identification method. Those skilled in the art can understand that all or part of the processes in the above embodiments can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium, and when executed, the program can include the processes of the above embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD), etc. The storage medium can also include a combination of the above types of memories.

[0117] Obviously, those skilled in the art should understand that the above modules or steps of the present application can be realized by general computing devices, which can be concentrated on a single computing device or distributed on a network composed of multiple computing devices, and optionally, they can be realized by program codes executable by computing devices, so that they can be stored in storage devices and executed by computing devices, or they can be respectively manufactured into individual integrated circuit modules, or multiple modules or steps among them can be manufactured into a single integrated circuit module to realize. Thus, the present application is not limited to any specific combination of hardware and software.

[0118] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for intelligent recognition of alluvial fan deposits by remote sensing, characterized in that, The application relates to a method for identifying a diluvial fan, and belongs to the technical field of remote sensing image processing. The method comprises the following steps: synthesizing remote sensing images by synthesizing optical remote sensing images and terrain remote sensing images, and preparing a diluvial fan target sample data set according to the synthesized remote sensing images, wherein the optical remote sensing images contain three RGB bands; the terrain remote sensing images contain three bands of a digital elevation model, slope and slope direction, and the slope and slope direction are calculated according to the digital elevation model; A MASK R-CNN deep convolutional neural network is constructed, and the structure parameters of the network are adjusted according to the characteristics of the synthesized remote sensing images, wherein the basic structure of the MASK R-CNN deep convolutional neural network is, from bottom to top, a Backbone network, an RPN network, an Align layer pooling layer and a head layer; Model training and model testing are performed according to the diluvial fan target sample data set and the MASK R-CNN deep convolutional neural network; A diluvial fan identification model is obtained by packaging the trained model, and diluvial fan information is identified according to the diluvial fan identification model. The synthesized remote sensing images are obtained by synthesizing the optical remote sensing images and the terrain remote sensing images, and the method comprises the following steps: Three bands of the digital elevation model, the slope and the slope direction containing terrain information are added to the three bands of the RGB of the optical remote sensing images to generate synthesized remote sensing images containing six bands of spectral and terrain information; The structure parameters of the network are adjusted according to the characteristics of the synthesized remote sensing images, and the method comprises the following steps: The source code of the MASK R-CNN deep convolutional neural network is modified, the image loading method is rewritten, the array dimension in the model establishment function is changed, the model weight loading function and the training function are modified, the first convolutional layer of the Backbone network is modified to adapt to the change of input data, and the input channel is changed from three channels of RGB to six channels of the synthesized remote sensing images; The diluvial fan target sample data set is prepared according to the synthesized remote sensing images, and the method comprises the following steps: The diluvial fan target contour vector is labeled in the synthesized remote sensing images; The synthesized remote sensing images are cut into a plurality of image blocks of multiple scales according to a preset multi-scale width and height; All the image blocks are traversed to perform intersection processing with the diluvial fan target contour vector, the diluvial fan target contour vector corresponding to each image block is extracted to form a contour vector small image vector set, a sample data set is prepared based on the image block and the vector set, and point information in the vector set is re-encoded into a json format and stored. 2.The alluvial fan deposit remote sensing intelligent identification method according to claim 1, characterized in that, The model training and model testing are performed according to the diluvial fan target sample data set and the MASK R-CNN deep convolutional neural network, and the method comprises the following steps: The trained model is tested according to the test sample set in the diluvial fan target sample data set, and the sample recognition accuracy of the trained model is obtained; If the sample recognition accuracy is greater than a preset threshold, the model training is stopped, and the current trained model is taken as the trained model.

3. A device for intelligent recognition of alluvial fan deposits from remote sensing, characterized in that, The application relates to a method for identifying a diluvial fan, and belongs to the technical field of remote sensing image processing. The sample data set generation unit is configured to synthesize optical remote sensing images and terrain remote sensing images to obtain synthetic remote sensing images, and to make a proluvial fan target sample data set according to the synthetic remote sensing images, wherein the optical remote sensing images contain three RGB bands; and the terrain remote sensing images contain three bands of a digital elevation model, slope and aspect, which are calculated according to the digital elevation model. The neural network construction unit is configured to construct a MASK R-CNN deep convolutional neural network, and to adjust structure parameters of the network according to features of the synthetic remote sensing images, wherein the basic structure of the MASK R-CNN deep convolutional neural network is, from bottom to top, a Backbone network, an RPN network, an Align layer pooling layer and a head layer. The model training and testing unit is configured to perform model training and model testing according to the proluvial fan target sample data set and the MASK R-CNN deep convolutional neural network. The model packaging unit is configured to package the trained model to obtain a proluvial fan identification model, so as to identify proluvial fan information according to the proluvial fan identification model. The sample data set generation unit comprises: The image synthesis module is configured to add the three bands of the digital elevation model, slope and aspect containing terrain information of the terrain remote sensing images to the three RGB bands of the optical remote sensing images to generate synthetic remote sensing images containing six-band features of spectral and terrain information. The neural network construction unit comprises: The structure parameter adjustment module is configured to modify source code of the MASK R-CNN deep convolutional neural network, rewrite an image loading method, change array dimensions in a model establishment function, modify a model weight loading function and a training function, complete modification of a first convolutional layer of the Backbone network to adapt to changes in input data, and realize change of input channels from three channels of RGB to six channels of the synthetic remote sensing images. The sample data set generation unit comprises: The labeling module is configured to label proluvial fan target contour vectors in the synthetic remote sensing images. The cutting module is configured to cut the synthetic remote sensing images into a plurality of image blocks of multiple scales according to a preset multiple-scale width and height. The contour vector small image vector set extraction module is configured to perform intersection processing on all image blocks and the proluvial fan target contour vectors, extract the proluvial fan target contour vectors corresponding to the intersection of each image block to form a contour vector small image vector set, and make a sample data set based on the image blocks and the vector set, and re-encode point information in the vector set into a json format for storage.

4. The alluvial fan deposit remote sensing intelligent identification device according to claim 3, characterized in that, The model training and testing unit comprises: The sample recognition accuracy calculation module is configured to test the trained model according to a test sample set in the proluvial fan target sample data set to obtain a sample recognition accuracy of the trained model. The model training completion module is configured to stop model training when the sample recognition accuracy is greater than a preset threshold, and to take the current trained model as a trained model.

5. 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 method of claim 1 or 2.

6. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 5. The computer program, when executed in a computer processor, implements the method as claimed in claim 1 or 2.

Images