A high-resolution remote sensing image water body extraction method and system

Abstract

The application discloses a high-resolution remote sensing image water body extraction method and system, which comprehensively utilizes the spectrum, area and shape and other characteristics of a water body object in a high-resolution remote sensing image to establish a decision tree, performs preliminary water body extraction from the high-resolution remote sensing image, and then constructs a lightweight convolutional neural network model to perform accurate extraction on the basis of the coarse water body extraction. Originating from the idea of ensemble learning, the decision tree is constructed by utilizing prior knowledge, and a deep learning method is introduced to combine weak classifiers into a strong classifier; after the image is subjected to the coarse water body extraction of the decision tree, the lightweight neural network model used for the accurate water body extraction has low complexity, small parameter quantity and high operation speed, can efficiently and accurately obtain water body patch information, and removes the interference of shadows, small targets and the like of the high-resolution remote sensing image.

Description

TECHNICAL FIELD

[0001] The application relates to a water body extraction method of remote sensing images, in particular to a high-resolution remote sensing image water body extraction method and system. BACKGROUND

[0002] Water is the necessary material basis for the survival of human beings and other creatures, and therefore understanding the distribution of water resources is of great guiding significance for the rational use and protection of water resources. Remote sensing technology has macroscopic, timely and real nature, and can obtain large-scale spectral image information of the earth's surface. In particular, the data obtained by various types of remote sensing means represented by satellite remote sensing have greatly improved spatial resolution and spectral resolution. Using satellite remote sensing data processing method to extract surface water body information has become a work with great practical value.

[0003] There are many methods for water body extraction on optical remote sensing satellite images. In the early stage, the threshold method was mainly used. The single-band threshold method uses the obvious absorption characteristics of water body in the near-infrared band to identify and extract water body, and is generally used for water body extraction in simple scene areas. The multi-band method uses spectral relationship and index method to extract water body in more complex scenes, but different background information has different effects on the accuracy of water body extraction, and the same index cannot be applied to all real scenes. In the actual extraction process, new extraction indexes need to be designed according to different background information, and the process has strong dependence on expert knowledge. In the later stage, machine learning type water body extraction method gradually achieved success. The decision tree method is used to select spectral and texture features first, and then set the decision rule to classify the scene to successfully extract the water body. However, the selection of features in complex scenes still has strong dependence on expert knowledge, and a very complex decision tree needs to be constructed. SVM and random forest methods are helpful to realize automatic extraction of water body, but due to their pixel-oriented processing characteristics, the extraction result is relatively fragmented and has more noise. In addition, there are methods for water body extraction using GIS vector assistance, but not every actual scene has GIS assistance information, so it is not suitable for water body extraction in all scenes. With the development of deep learning method, great progress has been made in water body information extraction on optical satellite images. VGG, ResNet and other classification networks and FCN and other segmentation networks have achieved good results in water body information extraction.

[0004] Deep learning method has achieved some good results for water body extraction, but this method needs a certain number of sample data, using neural network method for water body extraction, first needs artificial annotation data set, and with the complexity of neural network, in order to get better training effect, the number of training samples needed is more. High resolution remote sensing image spatial information is rich, and even the smallest water body target can be visually observed, so artificial annotation of water sample set in a complete scene needs a lot of time and labor cost. Therefore, how to reduce the labor cost and time cost, how to improve the efficiency and precision is an important development direction of high resolution remote sensing image water body extraction research. SUMMARY

[0005] The present application aims at overcoming the above-mentioned defects, and provides a high-resolution remote sensing image water body extraction method and system.

[0006] The high-resolution remote sensing image water body extraction method and system designed by the present application has the special feature that it comprises the following steps:

[0007] A. Obtain high-resolution remote sensing images of the study area and perform preprocessing;

[0008] B. Segment the preprocessed high-resolution remote sensing images, use multiple prior knowledge to establish decision tree classification method, and perform rough extraction of water body information;

[0009] C. Efficiently produce sample set by combining decision tree classification with artificial visual inspection or automatic inspection by comparing with historical base map, which is used for convolutional neural network model training and does not need artificial sample drawing;

[0010] D. Design water body extraction algorithm based on convolutional neural network, named DTCNN, and complete network training and optimization, the specific process is as follows:

[0011] 1) Construct DTCNN network for extracting water body information, the network model comprises 4 convolutional layers and 2 fully connected layers, the operation of the convolutional layers comprises convolution-batch normalization-pooling in turn; a global average pooling layer is added after the last output feature layer, and then connected with two fully connected layers to output results;

[0012] 2) Adjust hyperparameters and optimize; in the convolutional neural network training process, the convolution kernel size is 3x3, the number of four-layer convolution kernel is set to 32 32 64-128 respectively; 2x2 size maximum pooling method is adopted, the convolution step and the pooling step are both set to 1, the learning rate is 0.0001, the gradient descent method is SGD, the classification function is Softmax, and the BatchSize is 32;

[0013] E. On the basis of the water body rough extraction of the decision tree classification method, the DTCNN is used to realize the fine extraction of the water body object.

[0014] Preferably, the preprocessing in step A includes radiation correction and cropping of the high-resolution remote sensing image.

[0015] Further, the decision tree in step B includes four decision parameters: a normalized water index parameter I NDWI , an area parameter I a of the water body object, a spectral parameter I b of the water body object, and a shape parameter I r of the water body object.

[0016] B1, the normalized water index parameter I NDWI , when I NDWI of the pixel is greater than or equal to T1, the target is classified as water, otherwise the target is classified as other ground objects, wherein T1 is the minimum value of the normalized water index NDWI;

[0017] B2, the area parameter I a of the water body object, when I a is within [T2, T3], it is classified as a water body object, T2 and T3 respectively represent the minimum area and the maximum area of the water body object in the decision tree;

[0018] B3, the spectral parameter I b of the water body object, when I b is within [T4, T5], it is classified as a water body object, T4 and T5 respectively represent the minimum brightness value and the maximum brightness value of the water body object in the decision tree;

[0019] B4, the shape parameter I r of the water body object, when I r is within [T6, T7], it is classified as a water body object, T6 and T7 respectively represent the minimum value and the maximum value of the aspect ratio of the water body object in the decision tree.

[0020] Preferably, the process of step C is as follows:

[0021] 1) The high-resolution remote sensing image is segmented with a suitable scale, classified with the decision tree, and the vector map patches are derived;

[0022] 2) The high-resolution remote sensing image segmentation vector map patches are manually visually inspected or automatically inspected by comparing with historical base maps, the water body objects and non-water body objects are screened out, and the sample data set is efficiently produced without manual sample drawing.

[0023] Based on the same inventive concept, the present scheme also designs a system for implementing a high-resolution remote sensing image water body extraction method, which includes the following modules,

[0024] a preprocessing module, configured to preprocess the obtained high-resolution remote sensing image of the study area;

[0025] a coarse water body information extraction module, configured to segment the preprocessed high-resolution remote sensing image, and extract coarse water body information by using a decision tree classification method based on multiple priori knowledge;

[0026] a sample set production module, configured to efficiently produce a sample set by combining a decision tree classification method with manual visual inspection or automatic inspection based on historical base maps, and use the sample set for training of a convolutional neural network model;

[0027] a fine water body information extraction module, configured to extract fine water body information based on the coarse water body information extracted by the decision tree classification method, and use a DTCNN neural network to achieve the fine water body information extraction, the DTCNN neural network comprising four convolutional layers and two fully connected layers, the operations of the convolutional layers comprising, in sequence, convolution, batch normalization and pooling; a global average pooling layer is added after the last output feature layer, and then connected to two fully connected layers to output results; the DTCNN neural network adjusts hyperparameters and optimizes the process as follows: in the training process of the convolutional neural network, the size of the convolution kernel is 3x3, the number of convolution kernels in the four layers is set to 32 32 64-128 respectively; a 2x2 maximum pooling method is used, the convolution step and the pooling step are both set to 1, the learning rate is 0.0001, the gradient descent method is SGD, the classification function is Softmax, and the BatchSize is 32.

[0028] Based on the same inventive concept, the present scheme also designs an electronic device, comprising:

[0029] one or more processors;

[0030] a storage device configured to store one or more programs;

[0031] When the one or more programs are executed by the one or more processors, the one or more processors implement the high-resolution remote sensing image water body extraction method.

[0032] Based on the same inventive concept, the present scheme also designs a computer readable medium having a computer program stored thereon, wherein the program is executed by a processor to implement the high-resolution remote sensing image water body extraction method.

[0033] The advantages of this invention are as follows: It comprehensively utilizes the spectral, area, and shape features of water objects in high-resolution remote sensing imagery to build a decision tree, performing preliminary water body extraction from the high-resolution remote sensing imagery. Then, a lightweight convolutional neural network model is constructed to perform precise extraction based on the coarse water body extraction. Originating from the concept of ensemble learning, it utilizes prior knowledge to build the decision tree and introduces deep learning methods to combine weak classifiers into strong classifiers. Furthermore, addressing the difficulty in obtaining a large number of training samples required for deep learning models, it employs a method combining decision tree classification with manual visual inspection or automatic inspection using historical base maps to efficiently create sample sets for training the convolutional neural network model. This eliminates the need for manual sample drawing, reducing the labor cost of creating training datasets. After the image undergoes coarse water body extraction via the decision tree, the lightweight neural network model used for fine water body extraction has low complexity and few parameters, resulting in fast computation speed. Therefore, it can efficiently and accurately obtain water body patch information and remove interference from shadows, small targets, etc., in the high-resolution remote sensing imagery. Attached Figure Description

[0034] Figure 1 This is a flowchart of the present invention.

[0035] Figure 2 This invention relates to the decision tree model.

[0036] Figure 3 The diagram below shows the structure of the convolutional neural network of this invention. Detailed Implementation

[0037] To make the technical means, creative features, workflow, usage methods, and achieved objectives and effects of the present invention easier to understand, the present invention will be further described below in conjunction with the accompanying drawings.

[0038] Example 1

[0039] The water body extraction method based on the collaboration of decision tree and convolutional neural network designed in this invention includes the following steps: A. Acquiring high-resolution remote sensing images of the study area and performing preprocessing such as radiometric correction and cropping on the high-resolution remote sensing images;

[0040] B. The specific method for constructing a classification decision tree using prior knowledge is as follows:

[0041] The features used in the prior decision tree are one of the key factors affecting the classification accuracy. Through the analysis of water body targets, multiple features were selected to form a quantitative expression, which constitutes a classification decision tree for water body information extraction, such as... Figure 2 As shown. Figure 2 The paper presents a prior decision tree model for water body extraction that combines spectral information, exponential features, and prior knowledge. This model consists of four parameters and their value ranges, including the exponential parameter I. NDWI , area parameter Ia, spectral parameter Ib, and shape parameter Ir.

[0042] B1 When I NDWI greater than or equal to T1, the target is classified as a water body, otherwise the target is classified as other ground objects, this step is used to remove non-water body objects, wherein T1 is the minimum value of the NDWI of the water body, and the empirical value is generally 0. The NDWI index calculation formula is as follows, wherein p green and p nir respectively represent the reflectivity of the green wave band and the near-infrared wave band,

[0043] .

[0044] B2 The water body object has a certain area range, and I a is used to represent the water body object area, when I a is within [T2, T3], it is classified as a water body object, and T2 and T3 respectively represent the minimum area and the maximum area of the water body object in the decision tree.

[0045] B3 The spectral parameter I b of the water body object, the water body usually has a relatively obvious brightness range, when I b is within [T4, T5], it is classified as a water body object, and T4 and T5 respectively represent the minimum brightness value and the maximum brightness value of the water body object in the decision tree.

[0046] B4 The shape parameter I r of the water body object, the aspect ratio of the water body object is usually within a range, when I r is within [T6, T7], it is classified as a water body object, and T6 and T7 respectively represent the minimum value and the maximum value of the aspect ratio of the water body object in the decision tree.

[0047] C, a method of combining decision tree classification with artificial visual inspection or automatic inspection with historical base map is adopted to efficiently produce a sample set, and the specific operation steps are as follows:

[0048] 1) For high-resolution remote sensing images, suitable scales are used for segmentation, the above-mentioned decision tree is used for classification, and vector graph patches are derived.

[0049] 2) The segmented vector graph patches of the high-resolution remote sensing images are artificially visually inspected or automatically inspected with historical base maps, and the segmented correct water body objects and non-water body objects are screened out to produce a sample data set;

[0050] This method of combining decision tree classification with artificial visual inspection or automatic inspection with historical base map regards the segmented objects as samples and screens them out, without the need for manual sample drawing, so that the sample production workload is greatly reduced, thereby reducing time and labor costs and improving production efficiency.

[0051] D, Constructing DTCNN network for extracting water body information, and training and optimizing the network

[0052] 1) Constructing DTCNN network for extracting water body information

[0053] The convolutional neural network structure in the designed DTCNN model is as shown in Figure 3 , wherein "Con" represents a convolutional layer, "BN" represents a batch normalization layer, "ReLU" is an activation function, "Pooling" represents a pooling layer, "FC" represents a fully connected layer, and "GAP" represents a global average pooling. After the image segmentation result is input into the DTCNN model, it is first subjected to four groups of "convolution-batch normalization-pooling" operations, a global average pooling layer is added after the last output feature layer, and then two layers of fully connected layers are connected to output the result. The DTCNN model proposed in the present application has only 6 layers, 4 convolutional layers and 2 fully connected layers, and is a lightweight neural network model.

[0054] 2) Adjustment of hyperparameters and optimization

[0055] In order to make the classification result have higher precision, the hyperparameters in the network also need to be optimized. The hyperparameters of the convolutional layer mainly include the number of convolutional kernels, the size of the convolutional kernel, the padding mode and the step, and the parameters of the pooling layer mainly include the size of the pooling kernel and the pooling step. In the training process of the convolutional neural network, the present application selects a commonly used small convolutional kernel size of 3x3, the number of convolutional kernels of the four layers is set to 32 32 64-128 respectively, and a 2x2 size maximum pooling mode is adopted, and the convolution step and the pooling step are both set to 1. The learning rate is the step size of gradient descent in network training, which is set to 0.0001, the gradient descent method is SGD, the classification function is Softmax, and the BatchSize is 32.

[0056] D, Precise extraction of water body objects

[0057] On the basis of the water body rough extraction method of the decision tree classification method, the rough extraction result is input into the trained convolutional neural network DTCNN for precise extraction, so as to realize high-precision and high-efficiency extraction of water body information from high-resolution remote sensing images.

[0058] Case display

[0059] In order to verify the effectiveness and superiority of the proposed water body extraction method, the present application selects several commonly used and effective water body extraction methods and deep learning methods for comparative experiments, the selected comparison methods include support vector machine method SVM, random forest method RF and DT+Resnet34 model, wherein the DT+Resnet34 model also adopts the method of cooperation of decision tree and convolutional neural network, wherein the decision tree model is completely the same, and the neural network adopts 34-layer Resnet. The data used in the experiment is WorldView-2 high-resolution image data, and the water bodies extracted by the above methods are compared with the ground water body true value, the precision, recall, accuracy and F1 score are calculated, and the precision evaluation results are shown in Table 1. As can be seen from Table 1, although the precision of the SVM method and the RF method is not low; the DT+Resnet34 method and the method proposed in the present application can remove noise, building shadows and other sundries, and obtain results similar to the true value. Among them, the DTCNN network proposed in the present application has only 6 layers, which is much smaller than the 34 layers of Resnet. In order to compare the efficiency of the DT+Resnet34 method and the method proposed in the present application, the last column of Table 1 gives the time consumption of the two methods for extracting water body in a test area with a size of 4000x3000 pixels. From the test results, the water body extraction method proposed in the present application has all the precision evaluation indexes above 98% on high-resolution images, the highest precision, and the neural network model is small in size, and the efficiency of water body extraction is improved.

[0060] Table 1 quantitative comparison and analysis of the method of the present application and other methods

[0061]

[0062] Example two

[0063] Based on the same inventive concept, the present embodiment provides a system for implementing a high-resolution remote sensing image water body extraction method, comprising the following modules:

[0064] The preprocessing module pre-processes the obtained high-resolution remote sensing image of the research area;

[0065] The water body information rough extraction module segments the pre-processed high-resolution remote sensing image, establishes a decision tree classification method using a plurality of priori knowledge, and performs water body information rough extraction;

[0066] The sample set making module efficiently makes a sample set by combining decision tree classification with manual inspection or automatic inspection using historical base maps;

[0067] The water body fine extraction module is based on the water body coarse extraction of the decision tree classification method, and realizes fine extraction of the water body object by using a DTCNN neural network, the neural network comprises four convolution layers and two full connection layers, the operation of the convolution layers comprises convolution, batch normalization and pooling in sequence, a global average pooling layer is added after the last output feature layer, and then two full connection layers are connected to output results, the DTCNN neural network adjusts hyperparameters and an optimization process are as follows: in the convolution neural network training process, the convolution kernel size is 3*3, the number of four convolution kernels is respectively set to 32 32 64-128; a 2*2 maximum pooling manner is adopted, the convolution step and the pooling step are both set to 1, the learning rate is 0.0001, the gradient descent method is SGD, the classification function is Softmax, and the BatchSize is 32.

[0068] Embodiment three

[0069] Based on the same inventive concept, the application further provides an electronic device comprising one or more processors; a storage device for storing one or more programs; and when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in embodiment one.

[0070] Since the device introduced in embodiment three of the application is an electronic device used to implement the high-resolution remote sensing image water body extraction method based on homology in embodiment one of the application, the specific structure and modifications of the electronic device can be understood by those skilled in the art based on the method introduced in embodiment one of the application, and therefore will not be described here again. Any electronic device used in the method in embodiment one of the application belongs to the scope of protection of the application.

[0071] Embodiment four

[0072] Based on the same inventive concept, the application further provides a computer readable medium having a computer program stored thereon, and the program is executed by a processor to implement the method described in embodiment one.

[0073] Since the device introduced in embodiment four of the application is a computer readable medium used to implement the high-resolution remote sensing image water body extraction method based on homology in embodiment one of the application, the specific structure and modifications of the electronic device can be understood by those skilled in the art based on the method introduced in embodiment one of the application, and therefore will not be described here again. Any electronic device used in the method in embodiment one of the application belongs to the scope of protection of the application.

[0074] The specific embodiments described herein are merely illustrative of the spirit of the application. Modifications or supplements to the specific embodiments described or the use of similar means instead of the specific embodiments described can be made by those skilled in the art without departing from the spirit of the application or exceeding the scope defined by the appended claims.

Claims

1. A method for water body extraction from high-resolution remote sensing images, characterized in that: The method steps include: A. Obtain high-resolution remote sensing images of the study area and perform preprocessing; B. Segmenting the pretreated high-resolution remote sensing image, using various prior knowledge to establish a decision tree classification method to coarsely extract water body information, wherein the decision tree includes four decision parameters: a normalized water body index parameter I NDWI ; an area parameter I a of the water body object; a spectral parameter I b of the water body object; and a shape parameter I r of the water body object; The specific values of the four decision parameters are as follows: B1, Normalized Difference Water Index Parameter I NDWI When the I of the pixel NDWI When T1 is greater than or equal to T1, the target is classified as a water body; otherwise, the target is classified as other land features. T1 is the minimum value of the Normalized Difference Water Index (NDWI). B2, area parameter I of the water body object a When I a is divided into a water body object within [T2, T3], T2 and T3 respectively represent the minimum area and the maximum area of the water body object in the decision tree; B3, spectral parameter I of the water body object b When I b is divided into a water body object within [T4, T5], T4 and T5 respectively represent the minimum brightness value and the maximum brightness value of the water body object in the decision tree; B4, shape parameter I of the water body object r When I r is divided into a water body object within [T6, T7], T6 and T7 respectively represent the minimum and maximum of the aspect ratio of the water body object in the decision tree; C. A method combining decision tree classification and manual visual inspection or automated inspection using historical base maps is used to screen out correctly segmented water body objects and non-water body objects, and to produce a sample data set; D. Design a water body extraction network based on a convolutional neural network, complete the training and optimization of the network, and the specific process is as follows: 1) Build a network for extracting water body information, which includes 4 convolutional layers and 2 fully connected layers. The operations of the convolutional layers include convolution, batch normalization, and pooling in turn. A global average pooling layer is added after the last output feature layer, followed by two fully connected layers to output the results. 2) Adjust the hyperparameters and optimize them; E. On the basis of the decision tree classification method for water body rough extraction, the water body extraction network is used to realize the fine extraction of water body objects. 2.The method according to claim 1, wherein: The preprocessing in step A includes radiation correction and cropping of the high-resolution remote sensing images. 3.The method according to claim 1, characterized in that: The activation function in step D is ReLU, and the loss function is cross-entropy.

4. The high-resolution remote sensing image water body extraction method according to claim 1, characterized in that: In the convolutional neural network training process of step D, the 2d convolution with a kernel size of 3x3, the number of four-layer kernels is set to 32 32 64 – 128; and a 2x2 size maximum pooling method is adopted, the convolution step and the pooling step are both set to 1, the learning rate is 0.0001, the gradient descent method is SGD, the classification function is Softmax, and the BatchSize is 32.

5. The method of claim 1, wherein: The normalized difference water index NDWI index is calculated according to the following formula, where p(G) and p(NIR) represent the reflectance of the green band and near-infrared band, respectively. Green NIR respectively.​ 。 6. A system for implementing the method of extracting water bodies from high-resolution remote sensing images according to any one of claims 1-5, characterized in that it comprises: The method comprises the following steps: A. Preprocessing module, preprocessing the obtained high-resolution remote sensing images of the study area; The water body information rough extraction module uses multiple priori knowledge to establish a decision tree classification method to segment the preprocessed high-resolution remote sensing image and perform water body information rough extraction, wherein the decision tree comprises four decision parameters: a normalized water body index parameter I NDWI ; an area parameter I a of the water body object; a spectrum parameter I b of the water body object; and a shape parameter I r of the water body object. The specific values of the four decision parameters are as follows: B1, Normalized Difference Water Index Parameter I NDWI When the I of the pixel NDWI When T1 is greater than or equal to T1, the target is classified as a water body; otherwise, the target is classified as other land features. T1 is the minimum value of the Normalized Difference Water Index (NDWI). B2, an area parameter I of the water body object a When I a is divided into a water body object within [T2, T3], T2 and T3 respectively represent the minimum area and the maximum area of the water body object in the decision tree; B3, spectral parameter I of the water body object b When I b is divided into a water body object within [T4, T5], T4 and T5 respectively represent the minimum brightness value and the maximum brightness value of the water body object in the decision tree; B4, shape parameter I of the water body object r When I r is divided into a water body object within [T6, T7], T6 and T7 respectively represent the minimum and maximum of the aspect ratio of the water body object in the decision tree; Sample set production module, a method combining decision tree classification and manual visual inspection or automated inspection using historical base maps is used to efficiently produce sample data for convolutional neural network model training; Water body fine extraction module, on the basis of the decision tree classification method for water body rough extraction, the DTCNN neural network is used to realize the fine extraction of water body objects. The neural network includes 4 convolutional layers and 2 fully connected layers. The operations of the convolutional layers include convolution, batch normalization, and pooling in turn. A global average pooling layer is added after the last output feature layer, followed by two fully connected layers to output the results. Then the DTCNN neural network hyperparameters are adjusted and the neural network is optimized.

7. An electronic device, comprising: The method comprises: One or more processors; Storage device for storing one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1-5.

8. A computer readable medium having stored thereon a computer program, characterized in that: The program is executed by the processor to implement the method of any one of claims 1-5.

Images