An optimal water body index threshold determination method and a remote sensing image water surface extraction method

Abstract

The present application provides a kind of based on convolutional neural network's adaptive optimal water body index threshold determination method, system and remote sensing image water surface extraction method, medium, equipment, belongs to water color remote sensing technical field.The adaptive optimal water body index threshold determination method, with the reflectivity data of green light band and near infrared band and NDWI statistical information as input, constructs convolutional neural network and carries out training, determines adaptive optimal NDWI index threshold.The system includes pre-processing module, operation module, statistical module and model construction and training module.The water surface extraction method is implemented according to adaptive optimal NDWI index threshold.The present application only needs to input green light band reflectivity data, near infrared band reflectivity data and NDWI statistical information, and the output is single adaptive optimal NDWI index threshold, without outputting planar water body and non-water body information, without increasing training memory occupation, effectively reduces the consumption of computing resources and training difficulty.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water color remote sensing, and in particular to a method and system for determining an adaptive optimal water index threshold value based on a convolutional neural network, a method for extracting a water surface from a remote sensing image, a medium and an equipment. BACKGROUND

[0002] Water surface extraction is an important technical prerequisite and preprocessing step for water color remote sensing. Most current water surface extraction methods are based on deep learning algorithms or NDWI index methods. The water surface extraction algorithm based on deep learning algorithms requires a large amount of labels and computer power to label and train image data, consuming a lot of computing resources. The NDWI index method uses the strong reflection peak of water in the green light band and the strong absorption valley feature of the near-infrared band to extract the water surface, calculates the NDWI index (normalized difference water index) value by subtraction and division, and divides the water surface with a threshold value of 0.

[0003] However, there are some buildings around inland water bodies (especially narrow rivers), and due to the reflection of the building materials on the top of these buildings to green light and strong absorption of near-infrared or the shadow of these buildings causing water surface reflectivity signal interference, the NDWI index is not significantly divided between the water body and some special buildings, and the NDWI threshold value needs to be constantly adjusted for water surface extraction. In addition, the NDWI index threshold values of different rivers are not the same. Therefore, under the condition of not consuming too much computer power, it is extremely urgent to propose a method for determining an adaptive NDWI index threshold value. SUMMARY

[0004] The present application aims to provide a method and system for determining an adaptive optimal water index threshold value based on a convolutional neural network, and a method for extracting a water surface from a remote sensing image, a medium and an equipment, which only needs to input green light band reflectivity data, near-infrared band reflectivity data and NDWI statistical information, and outputs a single adaptive optimal NDWI index threshold value, without the need to output planar water and non-water body information, without the need to increase training memory occupation, effectively reducing the consumption of computing resources and the difficulty of training.

[0005] The technical solution adopted by the present application is as follows:

[0006] A method for determining an adaptive optimal water index threshold value based on a convolutional neural network, comprising the following steps:

[0007] Step S1, preprocessing the acquired remote sensing image;

[0008] Step S2, selecting the green light band reflectivity data and the near-infrared band reflectivity data of the preprocessed remote sensing image for operation to obtain an NDWI index;

[0009] Step S3, manually extracting the water surface range data and obtaining NDWI statistical information of the NDWI index in the water surface range data; wherein the NDWI statistical information includes NDWI value spatial distribution, NDWI index maximum value, NDWI index minimum value, NDWI index mean value and NDWI index standard deviation;

[0010] Step S4, constructing a convolutional neural network and training the same with the green waveband reflectivity data, the near-infrared waveband reflectivity data and the NDWI statistical information as input and the optimal NDWI index threshold as output to obtain an optimal convolutional neural network model and determine an adaptive optimal NDWI index threshold.

[0011] Further, the preprocessing in step S1 includes performing waveband registration, orthophoto mosaic and radiation calibration on the remote sensing image.

[0012] Further, the manual extraction of the water surface range data in step S3 is performed by manual labeling or an interactive tool to extract real water surface range data, which is a vector polygon or a raster mask.

[0013] Further, the convolutional neural network in step S4 includes an input layer, two normalization layers, five convolutional layers, five pooling layers, four deconvolutional layers, two ReLU activation function layers, a channel attention map, a fully connected layer and an output layer; the green waveband reflectivity data, the near-infrared waveband reflectivity data and the NDWI statistical information are input into the input layer as input data, the input data is normalized by the normalization layers to ensure that the data is on the same scale and facilitate network learning; the normalized data is subjected to two convolution-pooling-deconvolution operations to realize feature extraction and recovery, then a ReLU activation function is used to introduce nonlinearity, important feature channels are highlighted by channel attention mechanism for weighted processing, and the network pays more attention to key information; the feature maps subjected to the two convolution-pooling-deconvolution operations and the weighted processing by the channel attention mechanism are normalized to ensure that the data is on the same scale; the convolution-pooling-ReLU activation function introduction operation is further used to extract and enhance the features; finally, the fully connected layer is used to integrate global features and output a specific numerical value as the optimal NDWI index threshold.

[0014] Further, the convolutional layer uses a 3x3 convolution kernel, the pooling layer uses a 2x2 pooling kernel, the deconvolutional layer uses a 3x3 deconvolution kernel, and the step is 1.

[0015] Based on the same inventive concept, the application also provides a convolutional neural network-based adaptive optimal water body index threshold value determination system for implementing the convolutional neural network-based adaptive optimal water body index threshold value determination method as described above, comprising:

[0016] a preprocessing module for preprocessing the acquired remote sensing image;

[0017] an operation module for selecting the green light band reflectivity data and the near-infrared band reflectivity data of the preprocessed remote sensing image to perform operation and obtain the NDWI index;

[0018] a statistical module for manually extracting water surface range data and acquiring NDWI statistical information of the NDWI index in the water surface range data; wherein the NDWI statistical information comprises NDWI value spatial distribution, NDWI index maximum value, NDWI index minimum value, NDWI index mean value and NDWI index standard deviation;

[0019] a model construction and training module for taking the green light band reflectivity data, the near-infrared band reflectivity data and the NDWI statistical information as input, the optimal NDWI index threshold value as output, constructing a convolutional neural network and performing training to obtain an optimal convolutional neural network model and determine an adaptive optimal NDWI index threshold value.

[0020] Based on the same inventive concept, the application also provides a remote sensing image water surface extraction method, comprising the following steps:

[0021] Step S1, an adaptive optimal NDWI index threshold value is determined by using the convolutional neural network-based adaptive optimal water body index threshold value determination method as described above;

[0022] Step S2, the water body range is extracted according to the adaptive optimal NDWI index threshold value.

[0023] Further, in the step S2 of extracting the water body range, the region of a certain point in the remote sensing image whose NDWI index is greater than or equal to the adaptive optimal NDWI index threshold value is the water surface and is assigned a value of 1, and the region of a certain point in the remote sensing image whose NDWI index is less than the adaptive optimal NDWI index threshold value is the non-water surface and is assigned a value of 0.

[0024] Based on the same inventive concept, the application also provides a computer storage medium, and computer readable instructions are stored on the computer storage medium, the computer readable instructions are read by one or more processors, so that the one or more processors execute the remote sensing image water surface extraction method as described above.

[0025] Based on the same inventive concept, the present application also provides an electronic device, comprising a processor and a memory, the memory being used to store a computer program, and the processor being used to execute the computer program to realize the remote sensing image water surface extraction method as described above.

[0026] The present application has the following advantages:

[0027] The present application provides a convolutional neural network-based adaptive optimal water body index threshold determination method, system, remote sensing image water surface extraction method, medium and device, which only needs to input green band reflectivity data, near-infrared band reflectivity data and NDWI statistical information, and outputs a single adaptive optimal NDWI index threshold, without outputting planar water body and non-water body information, without increasing training memory occupation, effectively reducing the consumption of computing resources and the difficulty of training. Moreover, the adaptive optimal NDWI index threshold can be applied to different river characteristics without the need for feature training for different rivers; furthermore, the adaptive optimal NDWI index threshold determination does not need to rely on personnel experience and can be automatically completed for water surface extraction. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0029] Figure 1 The flowchart of the convolutional neural network-based adaptive optimal water body index threshold determination method in embodiment 1.

[0030] Figure 2 The structure diagram of the convolutional neural network in embodiment 1.

[0031] Figure 3 The NDWI index spatial distribution diagram in embodiment 3.

[0032] Figure 4 The water body (white) and non-water body (black) spatial distribution diagram based on the adaptive optimal NDWI index threshold division in embodiment 3. DETAILED DESCRIPTION

[0033] The embodiments of the present application will be described in detail below with reference to the drawings.

[0034] Embodiment 1

[0035] In this embodiment, a convolutional neural network-based adaptive optimal water body index threshold determination method is provided, and the flowchart is as shown in FIG. 1. Figure 1The adaptive optimal water body index threshold determination method based on the convolutional neural network includes the following steps:

[0036] Step S1, a multi-spectral sensor carried by a UAV collects spectral data, i.e., remote sensing images, and the acquired remote sensing images are preprocessed;

[0037] Step S2, green band reflectivity data and near-infrared band reflectivity data of the preprocessed remote sensing images are selected for operation to obtain an NDWI index; the NDWI index is obtained through the following formula: ;

[0038] In the formula, NDWI represents the NDWI index;

[0039] GRE represents green band reflectivity data of the remote sensing images;

[0040] NIR represents near-infrared band reflectivity data of the remote sensing images;

[0041] Step S3, water surface range data are manually extracted, and NDWI statistical information of the NDWI index in the water surface range data is obtained; the NDWI statistical information includes NDWI value spatial distribution, an NDWI index maximum value, an NDWI index minimum value, an NDWI index mean value, and an NDWI index standard deviation;

[0042] The NDWI statistical information is obtained through the following formula: ; ; ; ;

[0043] In the formula, NDWI is the NDWI index;

[0044] is the NDWI index maximum value;

[0045] is the NDWI index minimum value;

[0046] is the NDWI index mean value;

[0047] is the NDWI value of the ith point;

[0048] n is the total amount of NDWI index samples;

[0049] is the NDWI index standard deviation;

[0050] Step S4, taking the green waveband reflectivity data, the near-infrared waveband reflectivity data and the NDWI statistical information as inputs, and taking the optimal NDWI index threshold value as output, constructing and training a convolutional neural network to obtain an optimal convolutional neural network model, and determining an adaptive optimal NDWI index threshold value.

[0051] The above scheme has the beneficial effect that in the embodiment, the optimal NDWI index threshold value can be automatically and efficiently determined by combining the convolutional neural network and the NDWI index statistical information, and compared with the traditional manual threshold setting method, the precision and adaptability of water body extraction are significantly improved, and the method is suitable for water body recognition tasks in different types of remote sensing images and complex environments; at the same time, only the green waveband reflectivity data, the near-infrared waveband reflectivity data and the NDWI statistical information need to be input, and the output is a single adaptive optimal NDWI index threshold value, without the need to output the water body and non-water body information in a planar form, without the need to increase the training memory occupation, and effectively reducing the calculation resource consumption and training difficulty. Moreover, the adaptive optimal NDWI index threshold value can be applied to different river characteristics under different working conditions, without the need to train the characteristics for different rivers; furthermore, the adaptive optimal NDWI index threshold value determination does not need to rely on personnel experience.

[0052] Further, the preprocessing in step S1 includes band registration, orthophoto mosaic and radiation calibration of the remote sensing image, which can effectively correct the geometric distortion and radiation difference in the remote sensing image, improve the image quality, and provide a high-quality data basis for subsequent NDWI index calculation and model training, thereby further improving the accuracy of the optimal threshold value determination.

[0053] Further, the manual extraction of the water surface range data in step S3 is performed by manual annotation or an interactive tool, and the real water surface range data is extracted in the form of a vector polygon or a raster mask, ensuring the authenticity and accuracy of the training data. In the embodiment, the water surface range data is provided in the form of a vector polygon or a raster mask, which can accurately reflect the boundary and shape of the water surface, provide high-quality annotation information for model training, and further improve the learning ability and generalization performance of the model on the water surface features.

[0054] Further, as shown in FIG. 6, the remote sensing image 601 is a remote sensing image of a river in a certain region, and the remote sensing image 601 is divided into a plurality of patches 602. Figure 2As shown in the figure, the convolutional neural network in step S4 includes an input layer, two normalization layers, five convolutional layers, five pooling layers, four deconvolutional layers, two ReLU activation function layers, a channel attention map, a fully connected layer, and an output layer; the green band reflectance data, the near-infrared band reflectance data, and the NDWI statistical information are input into the input layer as input data, the input data is normalized by the normalization layers to ensure that the data is on the same scale and facilitate network learning; the normalized data is subjected to two convolution-pooling-deconvolution operations to realize feature extraction and recovery, then a ReLU activation function is used to introduce nonlinearity, the channel attention mechanism is used for weighted processing to highlight important feature channels and enhance the network's attention to key information; the feature maps subjected to the two convolution-pooling-deconvolution operations and the weighted processing by the channel attention mechanism are normalized to ensure that the data is on the same scale; the convolution-pooling-ReLU activation function introduction operation is used to further extract and enhance the features; finally, the global features are integrated by the fully connected layer, and a specific numerical value is output as the optimal NDWI index threshold value. In this embodiment, the normalization layers, the convolution-pooling-deconvolution operations, the ReLU activation function, the channel attention mechanism, and the fully connected layer work together to automatically extract and enhance the features related to water bodies, highlight important feature channels, effectively improve the recognition and differentiation capabilities of the model for water body features, and finally output a more accurate optimal NDWI index threshold value, further improving the accuracy and reliability of water body extraction.

[0055] Further, the convolutional layer uses a 3x3 convolution kernel, the pooling layer uses a 2x2 pooling kernel, and the deconvolutional layer uses a 3x3 deconvolution kernel, and the step length is 1. In this embodiment, the size and step length of the convolution kernel, the pooling kernel, and the deconvolution kernel are specified, the key parameter settings in the network structure are unified, and the stability and consistency of the convolutional neural network in the feature extraction and recovery process are ensured. This consistent parameter configuration helps to improve the efficiency and stability of model training, and is also conducive to the migration and application of the model on different data sets, further improving the practicality and universality of the optimal threshold determination method.

[0056] Embodiment 2

[0057] This embodiment provides a convolutional neural network-based adaptive optimal water body index threshold determination system to implement the convolutional neural network-based adaptive optimal water body index threshold determination method as in Embodiment 1, which includes:

[0058] A preprocessing module, configured to preprocess the acquired remote sensing image;

[0059] A calculation module, wherein the calculation module selects the pre-processed green band reflectance data and the near infrared band reflectance data of the remote sensing image to perform calculations to obtain the NDWI index;

[0060] A statistical module, the statistical module is used to manually extract water surface range data and obtain NDWI statistical information of the NDWI index within the water surface range data; wherein the NDWI statistical information includes the spatial distribution of NDWI values, the maximum NDWI index value, the minimum NDWI index value, the mean NDWI index value, and the standard deviation of the NDWI index value;

[0061] A model construction and training module, which takes the green light band reflectance data, the near-infrared band reflectance data and the NDWI statistical information as input and the optimal NDWI index threshold as output, constructs a convolutional neural network and trains it to obtain an optimal convolutional neural network model and determine an adaptive optimal NDWI index threshold.

[0062] This embodiment achieves an automated process from remote sensing image acquisition to optimal threshold determination through the collaborative work of the preprocessing module, the computation module, the statistics module, and the model building and training module. This systematic approach improves the efficiency and operability of the entire water index threshold determination process, facilitating rapid deployment and application in practical applications. It can be widely applied to fields such as water resource monitoring, environmental protection, and disaster assessment.

[0063] Example 3

[0064] This embodiment also provides a remote sensing image water surface extraction method, including the following steps:

[0065] Step S1, determining an adaptive optimal NDWI index threshold using the adaptive optimal water index threshold determination method based on a convolutional neural network as in Example 1;

[0066] Step S2: extracting the water body range according to the adaptive optimal NDWI index threshold.

[0067] The beneficial effect of the above scheme is: in this embodiment, the optimal water body index threshold determination method is applied to remote sensing image water surface extraction. By determining the adaptive optimal NDWI index threshold, the water body range can be extracted more accurately. Compared with the traditional extraction method of manually revising the threshold, it can significantly improve the accuracy and adaptability of water surface extraction, effectively reduce misjudgment and missed judgment, and provide more reliable water body distribution information for water resources management, ecological environment monitoring, etc., which has important practical application value.

[0068] For example, the spatial distribution of NDWI values ​​of a remote sensing image is as follows: Figure 3As shown, the adaptive optimal NDWI index threshold obtained in the foregoing manner is T, and the spatial distribution map of water bodies (white) and non-water bodies (black) divided based on the adaptive optimal NDWI index threshold T is as shown in FIG. 3. Figure 4

[0069] Further, in the step S2 of extracting the water body range, the region in which the NDWI index of a point in the remote sensing image is greater than or equal to the adaptive optimal NDWI index threshold is a water surface and is assigned a value of 1, and the region in which the NDWI index of a point in the remote sensing image is less than the adaptive optimal NDWI index threshold is a non-water surface and is assigned a value of 0. The present embodiment clearly defines the assignment rule of the water surface and the non-water surface, that is, the region in which the NDWI index is greater than or equal to the optimal threshold is a water surface and is assigned a value of 1, and the region in which the NDWI index is less than the optimal threshold is a non-water surface and is assigned a value of 0. This clear assignment manner can quickly and accurately generate a binary map of the water body distribution, facilitating subsequent analysis and application. This clear assignment rule improves the readability and usability of the water surface extraction result, and provides direct data support for water body area calculation, water body change monitoring, and the like.

[0070] Based on the same inventive concept, the present embodiment further provides a computer storage medium, and computer readable instructions are stored on the computer storage medium. When the computer readable instructions are read by one or more processors, the one or more processors execute the remote sensing image water surface extraction method as described above.

[0071] Based on the same inventive concept, the present embodiment further provides an electronic device, and the electronic device includes a processor and a memory. The memory is configured to store a computer program, and the processor is configured to execute the computer program to implement the remote sensing image water surface extraction method as described above.​

Claims

1. A method for determining an adaptive optimal water index threshold based on a convolutional neural network, characterized in that: The following steps are involved: Step S1, preprocessing the acquired remote sensing image; Step S2, selecting the pre-processed green band reflectance data and near infrared band reflectance data of the remote sensing image to perform calculations to obtain the NDWI index; Step S3, manually extracting water surface range data and obtaining NDWI statistical information of the NDWI index in the water surface range data; wherein the NDWI statistical information includes the spatial distribution of NDWI values, the maximum NDWI index, the minimum NDWI index, the mean NDWI index, and the standard deviation of the NDWI index; Step S4, using the green light band reflectance data, the near-infrared band reflectance data, and the NDWI statistical information as input and the optimal NDWI index threshold as output, constructing a convolutional neural network and training it to obtain an optimal convolutional neural network model, and determining an adaptive optimal NDWI index threshold; The convolutional neural network in step S4 includes an input layer, two normalization layers, 5 convolution layers, 5 pooling layers, 4 deconvolution layers, 2 Relu activation function layers, a channel attention map, a fully connected layer, and an output layer; the green light band reflectance data, the near-infrared band reflectance data, and the NDWI statistical information are used as input data to import into the input layer, and the input data is normalized by the normalization layer to ensure that the data is on the same scale to facilitate network learning; the normalized data is subjected to two convolution-pooling-deconvolution operations to achieve Feature extraction and recovery, followed by the introduction of nonlinearity using the ReLu activation function, and weighted processing through the channel attention mechanism to highlight important feature channels, thereby enhancing the network's attention to key information; normalization of the feature maps obtained through two consecutive convolution-pooling-deconvolution operations and weighted processing through the channel attention mechanism to ensure that the data is at the same scale; the introduction of operations through the convolution-pooling-ReLu activation function to further extract and enhance features; and finally, the integration of global features through the fully connected layer to output a specific value as the optimal NDWI index threshold; The convolution layer uses a 3×3 convolution kernel, the pooling layer uses a 2×2 pooling kernel, and the deconvolution layer uses a 3×3 deconvolution kernel, and the step size is 1.

2. The method for determining the adaptive optimal water index threshold based on a convolutional neural network according to claim 1, wherein: The pre-processing in step S1 includes band registration, orthophoto stitching, and radiometric calibration of the remote sensing image.

3. The method for determining the adaptive optimal water index threshold based on a convolutional neural network according to claim 1, wherein: When manually extracting the water surface range data in step S3, the real water surface range data is extracted through manual annotation or interactive tools, and the real water surface range data is a vector polygon or a raster mask.

4. A system for determining an adaptive optimal water index threshold based on a convolutional neural network, for implementing the method for determining an adaptive optimal water index threshold based on a convolutional neural network according to any one of claims 1 to 3, characterized in that: include: A preprocessing module, wherein the preprocessing module is used to preprocess the acquired remote sensing image; A calculation module, wherein the calculation module selects the pre-processed green band reflectance data and the near infrared band reflectance data of the remote sensing image to perform calculations to obtain the NDWI index; A statistical module, the statistical module is used to manually extract water surface range data and obtain NDWI statistical information of the NDWI index within the water surface range data; wherein the NDWI statistical information includes the spatial distribution of NDWI values, the maximum NDWI index value, the minimum NDWI index value, the mean NDWI index value, and the standard deviation of the NDWI index value; A model construction and training module, which takes the green light band reflectance data, the near-infrared band reflectance data and the NDWI statistical information as input and the optimal NDWI index threshold as output, constructs a convolutional neural network and trains it to obtain an optimal convolutional neural network model and determine an adaptive optimal NDWI index threshold.

5. A method for extracting water surface from remote sensing images, characterized in that: The following steps are involved: Step S1, determining an adaptive optimal NDWI index threshold using the adaptive optimal water index threshold determination method based on a convolutional neural network as described in any one of claims 1 to 3; Step S2: extracting the water body range according to the adaptive optimal NDWI index threshold.

6. The method for extracting water surface from remote sensing images according to claim 5, characterized in that: When extracting the water body range in step S2, the area where the NDWI index of a point in the remote sensing image is greater than or equal to the adaptive optimal NDWI index threshold is a water surface and is assigned a value of 1, and the area where the NDWI index of a point in the remote sensing image is less than the adaptive optimal NDWI index threshold is a non-water surface and is assigned a value of 0.

7. A computer storage medium, characterized in that Computer-readable instructions are stored on the computer storage medium. When the computer-readable instructions are read by one or more processors, the one or more processors execute the remote sensing image water surface extraction method according to claim 5 or 6.

8. An electronic device, characterized in that: The electronic device includes a processor and a memory, the memory is used to store a computer program, and the processor is used to execute the computer program to implement the remote sensing image water surface extraction method as described in claim 5 or 6.

Images