Water body extraction method, system and device based on conditional random fields and sub-pixel localization

Abstract

This invention discloses a method, system, and apparatus for water body extraction based on conditional random fields and sub-pixel localization, relating to the field of remote sensing image processing technology. The method includes: preprocessing the data to be extracted to obtain low-resolution raw data, then inputting it into the generator network of a super-resolution generative adversarial network model to obtain corresponding high-resolution image data after super-resolution reconstruction; performing image binarization and connected component processing on the high-resolution image data to remove noise and obtain potential water body regions; expanding the boundaries of the potential water body regions according to a preset threshold to generate a lake buffer super-resolution image; and segmenting the lake buffer super-resolution image into several superpixel objects to obtain the final water body extraction result. The water body extraction method for the Larsmann Hills region in Antarctica obtained by this invention can provide a deeper understanding of the stability of the Antarctic ice sheet, hydrological cycle, ecosystem health, and the impact of global climate change.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method, system and apparatus for water body extraction based on conditional random fields and sub-pixel localization. Background Technology

[0002] Due to global climate change, the Antarctic environment is undergoing drastic changes, most notably the rate of increase in Antarctic water temperature, which far exceeds the global average. This warming phenomenon is particularly pronounced in Antarctic waters, resulting in a decrease in ice sheet thickness, a rapid shrinkage of glaciers, and the continuous melting of many ice shelves. Furthermore, human activities are causing water quality deterioration in Antarctica, severely damaging the region's ecosystem and further impacting global climate change.

[0003] Large-scale and rapid changes in Antarctic waters not only disrupt the overall operation of the global water cycle but also trigger a series of problems, including changes in water distribution around the world and global sea-level rise. Changes in the Antarctic water environment and climate have also altered the movement and distribution patterns of global heat and gases, leading to increased global climate instability and consequently a higher risk of regional or global natural disasters. Therefore, changes in Antarctic waters have attracted significant international attention and become one of the core research topics.

[0004] The Antarctic region has extreme climatic conditions; extremely low temperatures and strong winds not only limit the possibility of field surveys but also place extremely high demands on the acquisition and processing of remote sensing data. Lakes on the ice sheet typically form in summer; they are relatively small in area and unevenly distributed, and their appearance and disappearance exhibit significant spatiotemporal variability, making continuous monitoring difficult. The morphology and physical properties of these lakes are also influenced by ice sheet movement, the supply and discharge of meltwater, and climate change, further complicating lake identification.

[0005] The formation and changes of lakes exhibit strong spatiotemporal dynamics, requiring remote sensing data with sufficiently high spatiotemporal resolution to accurately capture these processes. However, obtaining high spatiotemporal resolution remote sensing data is challenging due to limitations in satellite orbit design and observation time. This is particularly true in the Antarctic region, where satellite transit frequencies are low and data acquisition cycles are long, making continuous monitoring of lake changes difficult.

[0006] The Larsmann Hills, located on the East Coast of Antarctica, is a key region for studying the exchange of matter and energy between the Antarctic continent and the Southern Ocean. The persistent effects of intense ice and wind erosion in the Larsmann Hills have resulted in unique geological formations, such as honeycomb-like patterns in the exposed bedrock. Furthermore, the complex and variable local climate, coupled with significant annual temperature variations, leads to a complex, dispersed, and highly variable distribution of lakes within the Larsmann Hills. Due to its unique geological and climatic conditions, the Larsmann Hills region has become a crucial node in the Antarctic water cycle. Changes in the area of ​​Larsmann Lakes have become an important indicator for studying climate change and ecological environment changes. With the increasing availability of high-resolution remote sensing data and advancements in processing technology, long-term monitoring of lakes, rivers, and glacial meltwater in this region using remote sensing technology can provide in-depth insights into the dynamic changes of ice sheet meltwater and its impact on surrounding seas. It can also more effectively monitor the area changes of more small lakes, revealing key processes such as replenishment, flow, evaporation, and freeze-thaw cycles in the Antarctic interior. This contributes to the development of more accurate Antarctic hydrological models, thereby improving the predictive capabilities of global climate models. Remote sensing monitoring can also help us identify possible groundwater flow paths and the distribution of subglacial water, which is of great scientific value for understanding the hydrodynamic processes at the bottom of the ice sheet and the mechanism of ice sheet movement. Summary of the Invention

[0007] This invention provides a method, system, and apparatus for water body extraction based on conditional random fields and sub-pixel localization. It focuses on lakes and other water bodies in the Larsmann Hills region of Antarctica as representative research subjects, investigating the extraction of water from these types of lakes characterized by small spatial scales and significant seasonal variations. This invention not only helps researchers gain a deeper understanding of Antarctic ice sheet stability, hydrological cycles, ecosystem health, and the impacts of global climate change, but also provides crucial scientific support for addressing global climate change and environmental protection. Specifically, it is achieved through the following technologies.

[0008] In a first aspect, the present invention provides a water body extraction method based on conditional random fields and sub-pixel localization, the steps of which include: preprocessing the water body image data to be extracted to obtain low-resolution raw data;

[0009] A super-resolution generative adversarial network model is constructed. The low-resolution raw data is input into the generator network of the super-resolution generative adversarial network model for processing, and then scored and judged by the discriminator network to obtain the corresponding high-resolution image data after super-resolution reconstruction.

[0010] The high-resolution image data is subjected to image binarization processing to extract potential water bodies and remove noise to obtain potential water bodies; the boundaries of the potential water bodies are expanded according to a preset threshold to generate a super-resolution image of the water body buffer zone.

[0011] The super-resolution image of the water buffer zone is segmented into several superpixel objects. A conditional random field model is constructed. Based on the conditional random field model and the minimum energy function loss criterion, the superpixel objects are segmented into water bodies and non-water bodies to obtain the final water body extraction result.

[0012] Furthermore, the method for preprocessing the water body image data to be extracted includes: performing atmospheric correction on the water body image data to be extracted, converting the zenith reflectance of the L1C data into the L2A surface reflectance, and obtaining the low-resolution raw data.

[0013] Furthermore, the super-resolution generative adversarial network model includes a generator network, a discriminator network, and a loss function; the generator network includes residual block units and subphase convolutional units.

[0014] The training method for the super-resolution generative adversarial network model includes:

[0015] The low-resolution raw data is input into the generator network, where it undergoes 9×9 two-dimensional convolution processing and the first activation function processing to complete the initial feature extraction.

[0016] The image data after initial feature extraction is input into the residual unit for several 2D convolution processes with 3×3 kernels, normalization processes, and first activation function processes;

[0017] The image data after initial feature extraction skips the processing of the residual block unit and is directly added to the image data after the processing of the residual block unit for the first element-wise addition;

[0018] The image data obtained after the first element-wise addition is subjected to two-dimensional convolution with a 3×3 kernel and normalized to obtain convolutional image data.

[0019] The image data after initial feature extraction skips the residual block unit process and is added to the convolutional image data for a second element-wise addition;

[0020] The image data that has undergone the second element-level addition process is input into the sub-pixel convolution unit, and then subjected to 3×3 two-dimensional convolution, pixel mapping, activation function processing, and 9×9 two-dimensional convolution to obtain a high-resolution image.

[0021] The generated high-resolution image is scored and judged using a discriminator network to confirm its authenticity.

[0022] Furthermore, the method for performing image binarization and connected component processing on the high-resolution image data to remove noise and obtain the potential water body region includes:

[0023] Select a scene and use the high-resolution image data generated by the super-resolution generative adversarial network model to calculate the normalized water index. Mark the area with a normalized water index < 0.05 as water body and the area with a normalized water index ≥ 0.05 as non-water body to obtain a binarized image.

[0024] The binarized image is processed using a connected component processing method to calculate the area of ​​each water body connected component; connected components with an area smaller than the noise area threshold are marked as noise and filtered out; the remaining connected components are marked as potential lake areas.

[0025] Furthermore, for each water body connectivity region in the potential area of ​​the lake, the lake boundary is expanded according to a preset threshold to generate a super-resolution image of the lake buffer.

[0026] Furthermore, the method for obtaining the final water extraction result using the super-resolution image of the lake buffer zone includes:

[0027] The super-resolution image of the lake buffer is segmented into different superpixel objects;

[0028] A conditional random field model is constructed, wherein the unit potential energy function in the conditional random field model is composed of the band reflectivity of the optical data and its corresponding GLCM features, and the potential energy is composed of the cross-correlation function of different superpixel objects.

[0029] Based on the conditional random field model and the minimum energy function loss criterion, the superpixel object is segmented into live water and non-water regions to obtain the final water extraction result.

[0030] In a second aspect, the present invention provides a water body extraction system based on conditional random fields and sub-pixel localization, including an original image acquisition and preprocessing module, a super-resolution reconstruction module, and a lake water body extraction module;

[0031] The original image acquisition and preprocessing module is used to receive image data of the water body to be extracted and to preprocess the image data of the water body to be extracted to obtain low-resolution original data.

[0032] The super-resolution reconstruction module is used to construct and train a super-resolution generative adversarial network model, and outputs corresponding high-resolution image data from the input low-resolution raw data.

[0033] The lake water extraction module is used to perform image binarization and connected component processing on the high-resolution image data, remove noise, and obtain potential water areas; expand the boundaries of the potential water areas to generate a water buffer super-resolution image; segment the water buffer super-resolution image to obtain several superpixel objects; and based on the conditional random field model and the minimum energy function loss criterion, segment the superpixel objects into water bodies and non-water bodies to obtain the final water extraction result.

[0034] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, the computer program being used to cause a computer to perform the above-described water extraction method based on conditional random fields and sub-pixel localization.

[0035] In a fourth aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it is used to implement the above-described water extraction method based on conditional random fields and sub-pixel localization.

[0036] Compared with the prior art, the advantages of the present invention are:

[0037] 1. This invention utilizes a super-resolution generative adversarial network model to achieve super-resolution reconstruction of optical images and improve the ability to describe lake boundary features; on this basis, it uses a conditional random field model of water body buffer to extract lake water based on the minimum energy function loss criterion.

[0038] 2. The water extraction method of the Larsemann Hills region in Antarctica obtained by this invention can not only provide a deeper understanding of the stability of the Antarctic ice sheet, the hydrological cycle, the health of the ecosystem, and the impact of global climate change, but also provide key scientific support for addressing global climate change and environmental protection. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the overall process of the water extraction method provided in this embodiment of the invention.

[0040] Figure 2 The flowchart illustrates the water extraction method provided in this embodiment of the invention, which utilizes a generator network of a super-resolution generative adversarial network model to perform super-resolution reconstruction on low-resolution raw data, ultimately obtaining high-resolution image data.

[0041] Figure 3 The flowchart illustrates the process of using a discriminator network of a super-resolution generative adversarial network model to determine the authenticity of an input image in the water extraction method provided in this embodiment of the invention.

[0042] Figure 4The flowchart illustrates the process of obtaining the final water extraction result using the high-resolution image data in the water extraction method provided in this embodiment of the invention.

[0043] Figure 5 A comparison chart of water extraction results from different algorithms. Detailed Implementation

[0044] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] like Figure 1 As shown, the water body extraction method based on conditional random field and sub-pixel localization provided by the present invention has the following general steps: preprocessing the water body image data to be extracted to obtain low-resolution raw data;

[0046] A super-resolution generative adversarial network model is constructed. The low-resolution raw data is input into the generator network of the super-resolution generative adversarial network model for processing, and then scored and judged by the discriminator network to obtain the corresponding high-resolution image data after super-resolution reconstruction.

[0047] The high-resolution image data is subjected to image binarization processing to extract potential water bodies and remove noise to obtain potential water bodies; the boundaries of the potential water bodies are expanded according to a preset threshold to generate a super-resolution image of the water body buffer zone.

[0048] The super-resolution image of the water body buffer is segmented into several superpixel objects. A conditional random field model is constructed. Based on the conditional random field model and the minimum energy function loss criterion, the superpixel objects are segmented into water bodies and non-water bodies to obtain the final water body extraction result.

[0049] Example

[0050] Sentinel-2, also known as Sentinel 2, is a high-resolution multispectral imaging satellite carrying a multispectral imager (MSI) for land monitoring. It provides images of vegetation, soil and water cover, inland waterways, and coastal areas, and can also be used for emergency rescue services. This example uses lakes in the Larsmann Hills of Antarctica as an example, based on Sentinel-2 optical imagery, to extract information about these lakes. Sentinel-2A was launched by the European Space Agency in 2015, and Sentinel-2B in 2016; both are available for free download (https: / / dataspace.copernicus.eu / ).

[0051] Step 1: Data Preprocessing

[0052] Sentinel-2 optical imagery contains two types of data: L1C and L2A. The L1C data is not atmospherically corrected. To eliminate the influence of the atmosphere on lake extraction, the zenith reflectance of the L1C data needs to be converted to L2A surface reflectance (i.e., atmospheric correction is performed) to obtain preprocessed Sentinel-2 data, namely the original L2A surface reflectance (10m resolution).

[0053] Optionally, this process is implemented using Sen2cor software.

[0054] Step 2: Use a super-resolution generative adversarial network model to perform super-resolution processing on the preprocessed low-resolution impact data of Sentinel-2 to obtain high-resolution image data.

[0055] The main steps include: constructing a super-resolution generative adversarial network model (specifically, a generator network, a discriminator network, and a loss function); inputting Sentinel-2 data (raw L2A surface reflectance) preprocessed in step one as the training and testing sets; using the generator network, discriminator network, and loss function to train and test the super-resolution generative adversarial network model to obtain the optimal super-resolution generative adversarial network model.

[0056] In practical applications, the same method is used to input the preprocessed Sentinel-2 data to be tested into the model for super-resolution processing, and finally generate the corresponding high-resolution image data.

[0057] 1. Construct a super-resolution generative adversarial network model

[0058] The super-resolution generative adversarial network model consists of a generator network, a discriminator network, and a loss function.

[0059] (1) Generator Network

[0060] The generator network consists of a first input unit, residual blocks, sub-pixel convolution units, and a first output unit. There are multiple residual blocks.

[0061] Furthermore, the generator network employs a deep convolutional neural network with 16 residual block units. Specific network architecture parameters of the generator network include kernel size, stride, and number of feature maps.

[0062] The first input unit includes a two-dimensional convolutional layer (9×9 kernel, 64 feature maps, and a stride of 1) and a first activation function.

[0063] Each residual block contains: a 2D convolutional layer (3×3 kernel, 64 feature maps, stride 1), a batch normalization (BN) layer, a first activation function, and a skip connection. Subpixel convolutional units consist of a 2D convolutional layer, pixel maps, and a first activation function. Subpixel convolutional units are used for upsampling; the 2D convolutional layer has a 3×3 kernel, and the number of output feature maps is related to the upsampling factor.

[0064] Generally, parametric ReLU is used as the first activation function for residual block units and sub-pixel convolutional units to improve the nonlinear mapping capability of super-resolution generative adversarial network models.

[0065] (2) Discriminator Network

[0066] The discriminator network consists of a second input unit, a convolutional layer, a fully connected unit, and a second output unit. The discriminator network's function is to extract and analyze features from super-resolution data at different scales using a convolutional neural network, score the generated super-resolution reconstructed data, and determine whether the output image meets the required high-resolution standard.

[0067] ① Second input unit: includes input of real high-resolution images (HR, acquired through drones or aerial platforms) and high-resolution images (SR) generated by the generator network.

[0068] ② Convolutional layers: The discriminator network extracts image features through a series of convolutional layers. Each convolutional layer is usually followed by an activation function and a Batch Normalization (BN) layer.

[0069] The discriminator network consists of 8 convolutional layers with a kernel size of 3×3. The number of feature maps in each convolutional layer gradually increases with the network depth; initially, there are 64 feature maps, which then increase exponentially to 512 feature maps. To reduce the spatial resolution of the input image, stride convolution is used to reduce the image size each time the number of feature maps is doubled.

[0070] The second activation function is Leaky ReLU, which is used after each of the above convolutional layers to ensure that negative inputs also have non-zero gradients, thus enhancing feature extraction capabilities.

[0071] ③ Fully Connected Unit: After all convolutions and downsampling are completed, the feature map is flattened into a one-dimensional vector and passed to the fully connected unit. The fully connected unit consists of two fully connected layers.

[0072] For example, the fully connected layer Dense (1024): contains 1024 neurons followed by a Leaky ReLU activation function. Dense (1024) maps the high-dimensional features upsampled by the Pixel Shuffler back to the output high-resolution image.

[0073] Fully connected layer Dense (1): Output layer, containing 1 neuron, used to distinguish whether the input image is a real image or a generated image. The last layer is usually a convolution operation, without an activation function (such as ReLU), directly predicting the final pixel value.

[0074] ④ Second output unit: The last layer uses the third activation function (Sigmoid activation function) to process and make a judgment based on the output result.

[0075] As the number of layers increases, the number of feature maps gradually increases. To avoid the problem of infinite loops during model training, the Leaky ReLU activation function is used, with a hyperparameter of 0.2.

[0076] (3) Loss function

[0077] To evaluate the super-resolution reconstruction performance in the network, the loss function consists of three parts: adversarial loss, perceptual loss, and content loss. The adversarial loss describes the error between the high-resolution image data generated by the super-resolution generative adversarial network model and the real high-resolution data. The perceptual loss is calculated based on the Euclidean distance between the generated high-resolution image and the real image. The content loss represents the mean square error (MSE) between the generated image and the real image in pixel space.

[0078] The loss function is used to calculate the loss of the generator network and the discriminator network after each forward propagation. The forward propagation refers to the generator network receiving a low-resolution image input and generating a high-resolution image, and the discriminator network receiving the high-resolution image generated by the generator network and the real high-resolution image, and outputting the true and false probabilities.

[0079] The problem of solving the loss function can be transformed into a minimax optimization problem of an adversarial model between the generator network and the discriminator network, as shown in the following equation.

[0080] Formula I;

[0081] In the formula, For the parameters of the generator network, For the parameters of the discriminator network, Represents the distribution of real data High-resolution image samples extracted from the data. These represent super-resolution image samples generated from low-resolution images. The probability that the true high-resolution image output by the discriminator network is a true image. The probability that the super-resolution generated image output by the discriminator network is a real image. This indicates that the goal of the generator network is to minimize this term so that the generated image satisfies the discrimination condition of the discriminator network.

[0082] The core objective of Generative Adversarial Networks (GANs) is to enable the generator and discriminator networks to jointly optimize the objective function through adversarial training.

[0083] For the generator, the purpose of the loss function is to: (1) minimize the probability that the discriminator misclassifies the generated sample as false, that is, minimize the probability of image super-resolution error; (2) alternately update the objective functions of the discriminator and the generator during the iteration process until the termination condition of the iteration is met.

[0084] For the discriminator, the loss function has two purposes: (1) to maximize the probability that the discriminator correctly classifies the real sample, that is, to maximize the probability that the discriminator image is a high-resolution image after upsampling; (2) to maximize the probability that the discriminator correctly identifies the generated sample as fake, that is, to maximize the probability that the discriminator image is the original low-resolution image that has not been upsampled.

[0085] 2. The training process of the super-resolution generative adversarial network model: High-resolution image data is obtained by using the trained super-resolution generative adversarial network model.

[0086] First, a generator network is introduced to enable the super-resolution generative adversarial network model to achieve a certain generative capability without adversarial loss. Then, the parameters of the generator network are fixed, and the parameters of the discriminator network are updated to reduce the loss function. Finally, the Adam optimizer (loss function) is used for training, giving the super-resolution generative adversarial network model good convergence performance and outputting the final super-resolution reconstructed image. To update the weight parameters of the corresponding convolutional kernels, each convolutional layer in the generator participates in the parameter update of the Adam optimizer.

[0087] The Adam optimizer is primarily used to update the parameters of the generator and discriminator networks. Specifically, in each iteration, it optimizes and adjusts the network parameters based on the gradient calculated from the loss function. Correspondingly, the Adam optimizer also minimizes the loss functions of the generator and discriminator. In Adam optimization, the learning rate is 10. -4 First and second momentum parameters They are 0.9 and 0.999 respectively.

[0088] The specific training process and practical application process of the super-resolution generative adversarial network model are as follows:

[0089] (1) such as Figure 2 As shown, high-resolution image data is obtained using the output of a generator network.

[0090] ① Initial feature extraction: The preprocessed Sentinel-2 data from step one, i.e., low-resolution images (LR), is input into the generator network of the super-resolution generative adversarial network model and subjected to two-dimensional convolution (9×9 kernel, 64 feature maps, stride of 1) to extract the basic features of the low-resolution images; then, the P ReLU (Parametric ReLU) activation function is used to increase the nonlinear expressive power.

[0091] ② Residual Unit Processing: The image data after initial feature extraction is input into the residual unit for two-dimensional convolution processing. Each layer of two-dimensional convolution processing includes two-dimensional convolution (3×3 kernel, 64 feature maps, stride of 1), normalization, and P ReLU activation function processing.

[0092] ③ The image data after initial feature extraction is skipped from the residual block unit processing and directly combined with the image data after the residual block unit processing for the first element-wise sum, thereby forming residual features. This can alleviate the gradient vanishing problem in deep networks and improve training efficiency.

[0093] ④ The residual features after the first element-level addition are subjected to two-dimensional convolution (3×3 kernel, 64 feature maps, stride of 1) and normalization.

[0094] ⑤ The image data after initial feature extraction skips the residual block unit process and the two-dimensional convolution and normalization processes in step ④, and is then subjected to a second element-wise summation with the image data processed in step ④. This process aims to improve gradient propagation efficiency and prevent gradient vanishing.

[0095] ⑥ Sub-pixel convolution processing: Image data with stable features that have undergone a second element-wise addition process is input into sub-pixel convolution units, and upsampling is performed using the Pixel Shuffle method.

[0096] Specifically, the process proceeds sequentially: 2D convolution processing (3×3 kernel, 64 feature maps, stride 1) generates 256 feature maps; pixel mapping processing, which rearranges the pixels of the feature maps to increase resolution, ultimately generating 1024 feature maps, equivalent to expanding the 256-dimensional features to 1024, which is a 2x upsampling; followed by P ReLU activation function processing. Finally, the low-resolution image is gradually restored to a high-resolution image.

[0097] ⑦ The data processed by the subphase convolution unit is then processed by the first output unit for a final two-dimensional convolution (9×9 kernel, 64 feature maps, stride of 1) to obtain the final reconstructed high-resolution image data (SR).

[0098] (2) The discriminator network is used to score the high-resolution results output by the generator network to determine the authenticity of the input image.

[0099] ① The high-resolution image (SR) generated by the generator network and the real high-resolution image (HR) are input into the second input unit and processed by two-dimensional convolution (3×3 convolution kernel, 64 feature maps, stride of 1) to extract the basic features of the low-resolution image; then the Leaky ReLU activation function is used for processing.

[0100] ② The image data processed by the second input unit is input into the convolutional layer for further processing. Each processing step includes two-dimensional convolution (3×3 kernel), normalization, and a second activation function. After all convolutions and upsampling are completed, the feature map is flattened into a one-dimensional vector.

[0101] ④ Pass the one-dimensional vector data obtained after processing by the convolutional layer to the fully connected layer.

[0102] ⑤ Input the data processed by the convolutional layer into the second output unit, use the Sigmoid activation function, and output the values ​​in the range [0,1].

[0103] The closer the output value is to 1, the more likely it is the original low-resolution image. The closer it is to 0, the more likely the model considers the output image to be a generated high-resolution image (SR).

[0104] (3) Loss function.

[0105] In the training or super-resolution processing of the above-mentioned super-resolution generative adversarial network model, a two-dimensional convolutional layer at the sub-phase scale is constructed based on the super-resolution parameter size set at input. The input low-resolution image is processed by two-dimensional convolution to extract basic features, and then processed by the combination of multiple residual block units to achieve feature enhancement. Thus, feature extraction units for high-resolution images are constructed based on low-resolution images, and finally processed by a discriminator network.

[0106] Step 3: Generation of super-resolution images of the lake mask area

[0107] Due to the freeze-thaw effect caused by temperature changes, the lakes in the Larsmann Hills region exhibit a cyclical freezing-thawing process. During freezing periods, the lake water is covered by ice and snow, reducing its contrast with surrounding features and leading to larger errors in lake area extraction. Furthermore, water reflectivity exhibits significant absorption across the visible to near-infrared bands in optical imagery. Therefore, to reduce computational load and improve the accuracy of lake area extraction, it is necessary to extract the potential locations of the lakes and generate lake buffer zones (i.e., lake mask areas).

[0108] like Figure 1 As shown, the specific steps include:

[0109] 1. Image binarization

[0110] (1) Sentinel-2 optics contains data from 13 bands from visible light to near-infrared. Select band 3 (green band) and band 8 (near-infrared band, NIR) data (i.e. high-resolution image data) of a summer Sentinel-2 image generated in step two, and calculate its normalized water index (NDWI) according to the following formula II. At this time, the potential lake area is obtained by using the OTU global threshold.

[0111] Formula II;

[0112] (2) Mark NDWI < 0.05 as water bodies; otherwise, mark them as non-water bodies, and finally obtain the binarized image.

[0113] 2. Extraction from lake areas

[0114] For the obtained binarized image, a connected component processing method is used to calculate the area of ​​each water body's connected component; areas < 0.001 km² are considered. 2 Water bodies connected to the noise area threshold are marked as noise and filtered out; the remaining water bodies connected to the threshold are marked as potential lake regions.

[0115] 3. Generation of lake buffer zone (lake mask area)

[0116] Lakes exhibit seasonal variations, and the NDWI index can also introduce errors. Therefore, for each water body connectivity object, a specific threshold (e.g., 50m) is set to extend the lake boundaries, generating a super-resolution image of the lake buffer zone (lake mask area).

[0117] Step 4: Lake water extraction based on the Conditional Random Field (CRF) model

[0118] like Figure 4 As shown, the main process of this step includes: (1) segmenting the super-resolution image of the lake buffer into different superpixel objects; (2) constructing a conditional random field model, wherein the unit potential function is composed of the band reflectance of the optical data and its corresponding GLCM (Gray-Level Co-occurrence Matrix) features, and the potential energy is composed of the cross-correlation function of different superpixel objects; (3) based on the conditional random field model and the minimum energy function loss criterion, the Graph-Cut method is used to realize the superpixel-level segmentation and obtain the final water extraction result.

[0119] The specific steps are as follows:

[0120] 1. Using the Meanshift superpixel segmentation method, the super-resolution image of the lake buffer is segmented into different superpixel objects;

[0121] 2. The Conditional Random Field (CRF) model is constructed using the methods described in equations III to V below.

[0122] Formula III;

[0123] Formula IV;

[0124] Formula V;

[0125] in, It is the normalization coefficient; These are the weight parameters of the CRF model; The potential energy function represents the connection potential between different superpixel objects, and is constructed using a cross-correlation function (CRF for the potential energy function). Represents each superpixel The unit potential energy function; Features representing superpixels The features representing superpixel i (in this embodiment, these include spectral reflectance and various calculated grayscale histogram features); Indicates the type of water body (sea ice or seawater). These are the input features of the optical data; i and j represent superpixel units i and j, respectively; S is the set of all superpixel units; N i It is a set that does not contain superpixels i.

[0126] In addition to the spectral reflectance of each band, the features used also include: mean, standard deviation, homogeneity, contrast, correlation coefficient, and entropy, which are calculated using the following formulas VI to X respectively; the window size is set to 4, the separable distance is 4, the sliding window step size is 4, and the gray level is 64.

[0127] Formula VI

[0128] Formula VII;

[0129] Formula VIII;

[0130] Formula IX;

[0131] Formula X.

[0132] Where i and j are gray levels, k and K are the maximum values ​​of the gray levels (set to 64); Sd is the 8-neighborhood superpixel unit; μ x and μ y These are the mean values ​​along the row and column directions, respectively; σ x and σ x These are the standard deviations along the row and column directions, respectively.

[0133] 3. Water extraction based on minimum energy function.

[0134] Based on the conditional random field model and minimum energy function loss criterion constructed in the above steps, the Graph-Cut method is used to achieve superpixel segmentation, dividing the image into water bodies and non-water bodies, and obtaining the final water body extraction result.

[0135] like Figure 5 As shown, the experimental water bodies selected for this invention are lakes (Lake Progress and Lake Mochou) near the Larsmann Hills in Antarctica. Water extraction results from different algorithms are presented. Figure 4 As can be seen, the method provided by this invention achieves the best results not only in terms of the integrity of the extracted water area but also in terms of the preservation of edge details.

[0136] The above detailed embodiments describe the implementation of the present invention; however, the present invention is not limited to the specific details described in the above embodiments. Within the scope of the claims and technical concept of the present invention, various simple modifications and changes can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

Claims

1. A water body extraction method based on conditional random fields and sub-pixel localization, characterized in that, The steps include: preprocessing the water body image data to be extracted to obtain low-resolution raw data; A super-resolution generative adversarial network model is constructed. The low-resolution raw data is input into the generator network of the super-resolution generative adversarial network model for processing, and then scored and judged by the discriminator network to obtain the corresponding high-resolution image data after super-resolution reconstruction. The high-resolution image data is subjected to image binarization processing to extract potential water bodies and remove noise to obtain potential water bodies; the boundaries of the potential water bodies are expanded according to a preset threshold to generate a super-resolution image of the water body buffer zone. The super-resolution image of the water body buffer is segmented into several superpixel objects. A conditional random field model is constructed. Based on the conditional random field model and the minimum energy function loss criterion, the superpixel objects are segmented into water bodies and non-water bodies to obtain the final water body extraction result. The super-resolution generative adversarial network model includes a generator network, a discriminator network, and a loss function; the generator network includes residual block units and subphase convolutional units. The training method for the super-resolution generative adversarial network model includes: The low-resolution raw data is input into the generator network, where it undergoes 9×9 two-dimensional convolution processing and the first activation function processing to complete the initial feature extraction. The image data after initial feature extraction is input into the residual unit for several 2D convolution processes with 3×3 kernels, normalization processes, and first activation function processes; The image data after initial feature extraction skips the processing of the residual block unit and is directly added to the image data after the processing of the residual block unit for the first element-wise addition; The image data obtained after the first element-wise addition is subjected to two-dimensional convolution with a 3×3 kernel and normalized to obtain convolutional image data. The image data after initial feature extraction skips the residual block unit process and is added to the convolutional image data for a second element-wise addition; The image data that has undergone the second element-level addition process is input into the sub-pixel convolution unit, and then subjected to 3×3 two-dimensional convolution, pixel mapping, activation function processing, and 9×9 two-dimensional convolution to obtain a high-resolution image. The generated high-resolution image is scored and judged using a discriminator network to confirm its authenticity.

2. The water body extraction method based on conditional random fields and sub-pixel localization according to claim 1, characterized in that, The method for preprocessing the water body image data to be extracted includes: performing atmospheric correction on the water body image data to be extracted, converting the zenith reflectance of the L1C data to the L2A surface reflectance, and obtaining the low-resolution raw data.

3. The water body extraction method based on conditional random fields and sub-pixel localization according to claim 1, characterized in that, A method for performing image binarization and connected component processing on the high-resolution image data to remove noise and obtain potential water bodies includes: Select a scene and use the high-resolution image data generated by the super-resolution generative adversarial network model to calculate the normalized water index. Mark the area with a normalized water index < 0.05 as water body and the area with a normalized water index ≥ 0.05 as non-water body to obtain a binarized image. The binarized image is processed using a connected component processing method to calculate the area of ​​each water body connected component; connected components with an area smaller than the noise area threshold are marked as noise and filtered out; the remaining connected components are marked as potential lake areas.

4. The water body extraction method based on conditional random fields and sub-pixel localization according to claim 3, characterized in that, For each water body connectivity region in the potential area of ​​the lake, the lake boundary is expanded according to a preset threshold to generate a super-resolution image of the lake buffer.

5. The water body extraction method based on conditional random fields and sub-pixel localization according to claim 1, characterized in that, The method for obtaining the final water body extraction result using the super-resolution image of the water body buffer zone includes: The super-resolution image of the water body buffer zone is segmented into different superpixel objects; A conditional random field model is constructed, wherein the unit potential energy function in the conditional random field model is composed of the band reflectivity of the optical data and its corresponding GLCM features, and the potential energy is composed of the cross-correlation function of different superpixel objects. Based on the conditional random field model and the minimum energy function loss criterion, the superpixel object is segmented into live water and non-water regions to obtain the final water extraction result.

6. A water body extraction system based on conditional random fields and sub-pixel localization, characterized in that, The water extraction method according to any one of claims 1-5 includes an original image acquisition and preprocessing module, a super-resolution reconstruction module, and a lake water extraction module; The original image acquisition and preprocessing module is used to receive image data of the water body to be extracted and to preprocess the image data of the water body to be extracted to obtain low-resolution original data. The super-resolution reconstruction module is used to construct and train a super-resolution generative adversarial network model, and outputs corresponding high-resolution image data from the input low-resolution raw data. The lake water extraction module is used to perform image binarization and connected component processing on the high-resolution image data, remove noise, and obtain potential water areas; expand the boundaries of the potential water areas to generate a water buffer super-resolution image; segment the water buffer super-resolution image to obtain several superpixel objects; and based on the conditional random field model and the minimum energy function loss criterion, segment the superpixel objects into water bodies and non-water bodies to obtain the final water extraction result.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is used to enable a computer to execute the water extraction method based on conditional random fields and sub-pixel localization as described in any one of claims 1-5.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it is used to implement the water extraction method based on conditional random fields and sub-pixel localization as described in any one of claims 1-5.

Images