Remote sensing image water body segmentation system and method based on hybrid evolution dynamic calibration

Abstract

The invention discloses a remote sensing image water body segmentation system and method based on hybrid evolution dynamic calibration, and belongs to the technical field of remote sensing image processing and artificial intelligence. The system comprises an acquisition module, a data preprocessing module, a hybrid evolution optimizer, a dynamic feature calibration segmentation network, a model training and evaluation module and a water body identification output module. The method mainly comprises the following steps: preprocessing a remote sensing image and constructing a data set; a hybrid evolutionary optimizer is utilized to integrate a genetic algorithm, ant colony optimization, particle swarm optimization and an Alpha evolutionary algorithm, and optimal hyper-parameters of the segmentation network are automatically searched through a staged dynamic weight fusion mechanism; training and iteratively optimizing a dynamic feature calibration module based on optimal hyper-parameter configuration, and adaptively enhancing key features through a spectrum and space attention sub-module; and outputting a water body segmentation result by using the trained model. The method can effectively deal with a water body segmentation problem in a complex scene, and effectively improves the precision and robustness of water body segmentation.

Description

TECHNICAL FIELD

[0001] The present application relates to a remote sensing image water body segmentation system and method based on hybrid evolutionary dynamic calibration, belonging to the technical field of remote sensing image processing and artificial intelligence, especially suitable for remote sensing image water body segmentation based on hybrid evolutionary dynamic calibration. BACKGROUND

[0002] Accurate identification of water bodies in satellite images has crucial application value for environmental monitoring, water resource management and disaster emergency response. The development history of this technical field clearly reflects the evolution trajectory from relying on physical prior knowledge to data-driven models. Early research methods were mainly based on threshold models driven by physical prior knowledge, such as using the absorption characteristics of water bodies in the near-infrared band or constructing various enhanced water body indices (such as MNDWI, AWEI) for rapid segmentation. Although these methods are computationally efficient, their accuracy is heavily dependent on the empirical threshold set by humans, and are easily disturbed by environmental noise such as terrain shadows, cloud cover and spectral confusion with wetlands, resulting in insufficient robustness. Subsequently, machine learning methods such as support vector machines (SVM) and random forests (RF) improved this situation to some extent through the data-driven paradigm, which could handle non-linear classification problems and fuse multi-spectral features. However, these methods still face the challenge of complex feature engineering, and their generalization ability is limited in the case of limited sample size, and they have not fundamentally broken away from the dependence on human intervention. In recent years, deep learning methods, especially fully convolutional neural networks (FCN), U-Net and their variants, have achieved a generational leap in water body identification technology through autonomous feature learning and end-to-end training mode.

[0003] Although these models have shown excellent performance, their successful application in practice still faces several interrelated, yet unresolved core bottlenecks. First, model performance is highly sensitive to the setting of hyperparameters (such as learning rate, batch size), and existing grid search or manual tuning methods are not only inefficient, time-consuming and labor-intensive, but also difficult to find the global optimal solution, becoming a key obstacle to the upper limit of model performance. Second, when faced with complex remote sensing scenes, such as the presence of building shadows, thick cloud interference, and water bodies with a huge scale span (from small streams to large lakes), traditional network structures lack the ability to adaptively calibrate spectral and spatial features, resulting in poor capture of boundary details and poor adaptability to multi-scale targets. Finally, the serious imbalance between the number of water body and background pixels in remote sensing images makes it difficult for a single, fixed loss function (such as cross-entropy loss) to achieve an optimal balance between pixel classification confidence and segmentation region geometric integrity, often leading to missed detection of small target water bodies or broken boundary predictions. SUMMARY

[0004] The present application aims to overcome the above prior art deficiencies, and provide a remote sensing image water body segmentation system and method based on hybrid evolutionary dynamic calibration, to realize full-automatic super parameter optimization, accurate dynamic feature calibration and adaptive loss balancing recognition method, and thus improve the water body segmentation accuracy, robustness and efficiency in complex and variable remote sensing scenes.

[0005] To achieve the above object, the present application provides the following technical solutions:

[0006] The remote sensing image water body segmentation system based on hybrid evolutionary dynamic calibration combines Figure 1 characterized in that it is composed of a collection module, a data preprocessing module, a hybrid evolutionary optimizer, a dynamic feature calibration segmentation network, a model training and evaluation module, and a water body recognition output module; the collection module is connected to the input of the data preprocessing module to realize the collection of multispectral satellite remote sensing images; the data preprocessing module preprocesses the multispectral satellite remote sensing images, and its output is connected to the input of the dynamic feature calibration segmentation network; the hybrid evolutionary optimizer, the dynamic feature calibration segmentation network, and the model training and evaluation module are connected in sequence to form a closed loop, realizing the training of network super parameters; the water body recognition output module is a display terminal, its input is connected to the output of the dynamic feature calibration segmentation network, and the image water body segmentation result is output and displayed.

[0007] Further, the hybrid evolutionary optimizer integrates the genetic algorithm (GA), ant colony optimization (ACO), particle swarm optimization (PSO), and Alpha evolutionary algorithm (AE) in parallel, providing the dynamic feature calibration segmentation network with optimal super parameter configuration.

[0008] Further, the dynamic feature calibration segmentation network takes U-Net as the basic architecture, and is composed of a layer encoder, a layer decoder, one skip connection, and one dynamic feature calibration module.

[0009] Each layer of the encoder is composed of a convolution block one and a pooling layer in series, for extracting multi-level features from the input image, and its input is connected to the data preprocessing module and its output is connected to the input of the decoder;

[0010] Each layer of the decoder is composed of upsampling and a convolution block two in series, realizing image water body segmentation extraction;

[0011] wherein the upsampling is transposed convolution, and the convolution block one and the convolution block two are composed of convolution layers and ReLU layers in series;

[0012] Each of the jump connections is connected with the output of the first convolutional block in the corresponding level in the encoder path and the up-sampled output in the corresponding previous level in the decoder path respectively, and the concatenated result is used as the input of the dynamic feature calibration module in the corresponding level;

[0013] Each of the dynamic feature calibration modules is composed of a spectral attention sub-module and a spatial attention sub-module connected in series with a dynamic weight generation sub-module, and the output thereof is connected with the second convolutional block in the decoder;

[0014] The spectral attention sub-module is composed of a global average pooling layer connected in series with a multi-layer perceptron and then connected in series with a nonlinear activation function, uses one-dimensional convolution to capture the spectral dependence between channels, and adaptively enhances the band features sensitive to water bodies;

[0015] The spatial attention sub-module is composed of a global average pooling layer connected in parallel with a maximum pooling layer, then connected in series with a direction perception convolution and then connected in series with a nonlinear activation function, uses the direction perception convolution to extract the spatial context information in the horizontal and vertical directions, and strengthens the continuity of the water body boundary and linear features;

[0016] The dynamic weight generation sub-module is a lightweight network that dynamically generates spatially adaptive attention strength coefficients and feature retention coefficients according to the complexity of the input features, and realizes accurate feature calibration according to the regional requirements;

[0017] The second convolutional block is connected with a water body recognition output module, and outputs the water body segmentation result of the image.

[0018] Further, for the feature map output by the first convolutional block in any layer of the encoder , the output of the spectral attention sub-module is , wherein is a global average pooling operation, is a multi-layer perceptron, is a nonlinear activation function;

[0019] The output of the spatial attention sub-module is ; wherein, is a maximum pooling operation; is a direction perception convolution operation, specifically, a horizontal direction convolution kernel and a vertical direction convolution kernel are used to extract context information in different directions respectively, then the output results of the two directions are spliced, and then fused through a standard convolution.

[0020] The output of the dynamic weight generation sub-module is ; wherein, is an element-wise multiplication; ; The operation is composed of a global average pooling and two parallel The convolutional layers are connected in series, and the output satisfies The constraints.

[0021] Furthermore, the model training and evaluation module consists of a model training unit and a model evaluation unit connected in series. The model training unit uses the optimal hyperparameters obtained by the hybrid evolutionary optimizer to configure the dynamic feature calibration segmentation network and trains it using the training set. During training, a linearly weighted multi-task loss function is used, and its weight coefficients can be dynamically adjusted through an evolutionary strategy. During the training process, the model evaluation unit periodically uses the validation set to evaluate the model performance. If the evaluation result does not reach a preset threshold, the hybrid evolutionary optimizer is triggered to re-search for hyperparameters and retrain the model, forming a closed-loop optimization.

[0022] Furthermore, the model training unit employs a linearly weighted multi-task loss function that combines cross-entropy loss and Dice loss.

[0023] A water segmentation method for remote sensing images based on hybrid evolutionary dynamic calibration, combined with Figure 4 It includes the following steps:

[0024] S1: Data Acquisition and Preprocessing: Real-time acquisition of multispectral satellite remote sensing images, followed by normalization, data augmentation, and mask cropping preprocessing to obtain standard-sized multispectral satellite remote sensing patches;

[0025] S2: Dataset Construction: Historical data of some standard-sized multispectral satellite remote sensing tiles were manually labeled and divided into training set and validation set;

[0026] S3: Model Construction and Initialization: Build a dynamic feature calibration segmentation network, initialize the hybrid evolutionary optimizer, and set the initial population and hyperparameter search space for each evolutionary algorithm;

[0027] S4: Hybrid Evolutionary Hyperparameter Optimization: Start the hybrid evolutionary optimizer, execute the cooperative search process, and output the optimal hyperparameter combination under the current number of updates;

[0028] S5: Model Training: The optimal hyperparameter combination obtained in step S4 is configured into the dynamic feature calibration segmentation network. The model training unit uses a dynamically weighted multi-task loss function and the training set to train the weights of the dynamic feature calibration segmentation network.

[0029] S6: Iterative optimization: The model evaluation unit periodically uses the validation set to evaluate the performance of the dynamic feature calibration segmentation network based on the number of training rounds. If the segmentation accuracy does not reach the preset threshold, it returns to step S4 and triggers the hybrid evolutionary optimizer to re-search for hyperparameters until the maximum number of updates is reached.

[0030] S7: water body segmentation and result output: input the standard size multispectral satellite remote sensing image block to be recognized into the trained dynamic feature calibration segmentation network for forward propagation, output a binary segmentation image with the same size as the input, and the water body recognition output module displays the binary segmentation image to identify the water body area and complete the water body segmentation task.

[0031] Further, the step S4 is specifically:

[0032] S401: In each iteration, each algorithm generates a set of hyperparameter candidate solutions in parallel based on its inherent mechanism; wherein the inherent mechanism of GA is selection, crossover and mutation; the inherent mechanism of ACO is pheromone guided path selection; the inherent mechanism of PSO is individual and group experience update; and the inherent mechanism of AE is disturbance and selection of elite solution space;

[0033] S402: According to the ratio of the current iteration number to the total maximum iteration number, the iteration process is divided into three stages of early, middle and late to adjust the weight of each algorithm in joint prediction; wherein the weight of GA is increased in the early stage to strengthen global exploration; the weights of ACO and PSO are balanced in the middle stage to form dual driving of path optimization and gradient driving; and the weight of AE is improved in the later stage for fine search of solution space;

[0034] S403: Each set of hyperparameter candidate solutions is assigned a selected probability according to the weight of each algorithm in joint prediction, and a set of hyperparameters is randomly selected according to the probability

[0035] S404: Each set of hyperparameters is input into the dynamic feature calibration segmentation network, and the F1-Score is calculated on the validation set as the fitness value to guide the evolution direction;

[0036] S405: When the maximum iteration number or the fitness converges, the optimal hyperparameter combination under the current update number is output.

[0037] Further, the evaluation indexes of the model evaluation unit include: pixel accuracy (Pixel Accuracy, PA), F1-Score, average intersection over union (Intersection over Union, IoU), mean pixel accuracy (Mean Pixel Accuracy, MPA) and class pixel accuracy (Class Pixel Accuracy, CPA) based on confusion matrix.

[0038] An electronic device, comprising at least one processor; and a memory connected in communication with the at least one processor; wherein,

[0039] ​The memory stores a computer program executed by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the remote sensing image water body segmentation method based on hybrid evolution dynamic calibration.

[0040] Finally, the application also discloses a computer readable storage medium storing computer instructions for enabling a processor to implement the remote sensing image water body segmentation method based on hybrid evolution dynamic calibration when executed.

[0041] The application has the advantages that the application provides a remote sensing image water body segmentation system and method based on hybrid evolution dynamic calibration, proposes a synergistic framework of "evolution optimization + dynamic network + adaptive loss", realizes full automation of super parameter optimization through a hybrid evolution optimizer, gets rid of the dependence on artificial experience, and simultaneously uses a dynamic feature calibration segmentation network to adaptively enhance key spectral and spatial features, effectively overcomes shadow, noise interference and class imbalance problems, and has good network robustness. The method realizes super parameter automatic optimization and feature dynamic calibration, can effectively cope with water body segmentation problems in complex scenes, and significantly improves segmentation precision of water body boundaries and multi-scale targets. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to make the purpose, technical scheme of the application clearer, the application is provided with the following drawings for description:

[0043] Figure 1 Fig. 1 is a structural diagram of the remote sensing image water body segmentation system based on hybrid evolution dynamic calibration in the application; wherein, solid lines are data transmission directions, and dashed lines are training feedback directions;

[0044] Figure 2 Fig. 2 is a dynamic feature calibration segmentation network architecture diagram in embodiment 1 of the application;

[0045] Figure 3 Fig. 3 is a dynamic feature calibration module architecture diagram in embodiment 1 of the application;

[0046] Figure 4 Fig. 4 is a flowchart of the remote sensing image water body segmentation method based on hybrid evolution dynamic calibration in the application;

[0047] Figure 5 Fig. 5 is a hybrid evolution optimization flowchart in embodiment 2 of the application;

[0048] Figure 6 Fig. 6 is a river basin multispectral satellite remote sensing image in embodiment 2 of the application;

[0049] Figure 7 Fig. 7 is a river basin water body real segmentation annotation image in embodiment 2 of the application;

[0050] Figure 8 This is a diagram showing the Unet segmentation results in the comparative experiment of Embodiment 2 of the present invention;

[0051] Figure 9 This is a diagram showing the results of dynamically calibrated UNet water segmentation in the comparative experiment of Embodiment 2 of the present invention;

[0052] Figure 10 This is a diagram showing the water body segmentation results of the method of the present invention in the comparative experiment of Embodiment 2 of the present invention;

[0053] Figure 11 This is a schematic diagram of the electronic device in Embodiment 3 of the present invention. Detailed Implementation

[0054] To make the objectives and technical solutions of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0055] Example 1: For example Figure 6 The aforementioned multispectral satellite remote sensing image of a certain river basin, combined with Figure 7 The ground truth water body segmentation results provide technical support for accurate water body segmentation, environmental monitoring, water resource management, and disaster emergency response. This invention proposes a "remote sensing image water body segmentation system based on hybrid evolutionary dynamic calibration".

[0056] Combination Figure 1 The system consists of an acquisition module (1), a data preprocessing module (2), a hybrid evolutionary optimizer (3), a dynamic feature calibration segmentation network (4), a model training and evaluation module (5), and a water body identification output module (6). The acquisition module (1) is connected to the input of the data preprocessing module (2) to acquire multispectral satellite remote sensing images. The data preprocessing module (2) preprocesses the multispectral satellite remote sensing images, and its output is connected to the input of the dynamic feature calibration segmentation network (4). The hybrid evolutionary optimizer (3), the dynamic feature calibration segmentation network (4), and the model training and evaluation module (5) are sequentially connected in a closed loop to train the network hyperparameters. The water body identification output module (6) is a display terminal, and its input is connected to the output of the dynamic feature calibration segmentation network (4) to output and display the water body segmentation results of the image.

[0057] Furthermore, the hybrid evolutionary optimizer (3) integrates parallel genetic algorithms, ant colony optimization, particle swarm optimization and Alpha evolutionary algorithms to provide the optimal hyperparameter configuration for the dynamic feature calibration segmentation network (4).

[0058] Furthermore, combined with Figure 2, the dynamic feature calibration segmentation network (4) takes U-Net as the basic architecture, which is composed of a 50-layer encoder (41), a layer decoder (42), 50 skip connections (43), and 50 dynamic feature calibration modules (44);

[0059] The encoder (41) is realized by a pre-trained ResNet50, and each layer of the encoder (41) is composed of a convolution block one (411) and a pooling layer (412) in series, which is used to extract multi-level features from the input image, and is connected to the data preprocessing module (2) and the input of the decoder (42);

[0060] Each layer of the decoder (42) is composed of an up-sampling (421) and a convolution block two (422) in series, which realizes the extraction of water body segmentation of the image;

[0061] Among them, the up-sampling (421) is a transposed convolution, and the convolution block one (411) and the convolution block two (422) are convolution layers and ReLU layers connected in series;

[0062] Each of the skip connections (43) is connected to the output of the convolution block one (411) in the corresponding level in the encoder (41) path and the output of the up-sampling (421) in the corresponding previous level in the decoder (42) path, respectively, and is used as the input of the dynamic feature calibration module (44) in the corresponding level after being spliced by Concat;

[0063] In combination with Figure 3 Each of the dynamic feature calibration modules (44) is composed of a spectral attention sub-module (441) and a spatial attention sub-module (442) connected in series, and a dynamic weight generation sub-module (443) connected in series, which is connected to the convolution block two (422) in the decoder (42);

[0064] Among them, the spectral attention sub-module (441) is composed of a global average pooling layer connected in series with a multi-layer perceptron and then connected in series with a nonlinear activation function, which uses one-dimensional convolution to capture the spectral dependence between channels and adaptively enhances the band features sensitive to water bodies;

[0065] The spatial attention sub-module (442) is composed of a global average pooling layer connected in parallel with a maximum pooling layer, then connected in series with a direction perception convolution and a nonlinear activation function, which uses a direction perception convolution to extract spatial context information in horizontal and vertical directions and strengthens the continuity of water body boundaries and linear features;

[0066] The dynamic weight generation sub-module (443) is a lightweight network that dynamically generates spatially adaptive attention strength coefficients and feature retention coefficients according to the complexity of the input features, and realizes precise feature calibration according to regional needs; ​

[0067] The convolutional block two (422) is connected with the water body recognition output module (6), and outputs an image water body segmentation result.

[0068] Further, for the feature map output by the convolutional block one (411) in any layer of the encoder (41) The output of the spectral attention sub-module (441) is , wherein is a global average pooling operation, is a multi-layer perception, is a nonlinear activation function;

[0069] The output of the spatial attention sub-module (442) is , wherein is a max pooling operation; is a direction perception convolution operation, specifically, a horizontal direction convolution kernel and a vertical direction convolution kernel are used to extract context information in different directions respectively, then the output results of the two directions are spliced and fused through a standard convolution.

[0070] The output of the dynamic weight generation sub-module (443) is , wherein is an element-wise multiplication; ; The operation is composed of a global average pooling and two convolutional layers in parallel with , and the output satisfies the constraint condition .

[0071] Further, the model training and evaluation module (5) is composed of a model training unit (51) and a model evaluation unit (52) in series; the model training unit (51) configures the dynamic feature calibration segmentation network (4) with the optimal hyperparameters obtained by the hybrid evolutionary optimizer (3), and trains using a training set; in the training, a linearly weighted multi-task loss function is used, and the weight coefficients can be dynamically adjusted through an evolutionary strategy; the model evaluation unit (52) regularly evaluates the model performance using a validation set during the training process, and if the evaluation result does not reach a preset threshold, triggers the hybrid evolutionary optimizer (3) to re-search for hyperparameters and train the model, forming a closed-loop optimization.

[0072] Further, the model training unit (51) uses a linearly weighted multi-task loss function, which is a linear weighting of cross-entropy loss and Dice loss, i.e. .

[0073] , wherein is a classification loss, is a Dice loss, and are dynamic weight coefficients. Initial default settings =1, =1, both of which can be self-adaptively adjusted by the hybrid evolutionary optimizer (3) during the training process to balance the contribution of different loss terms to the optimization of the dynamic feature calibration segmentation network (4).

[0074] are updated in each round of hybrid evolutionary algorithm hyperparameter tuning, and are fixed and unchanged during the training period of the dynamic feature calibration segmentation network (4). Their "updates" occur in the iterations of the evolutionary algorithm, and each evolutionary iteration attempts a new set of and .

[0075] Cross-entropy loss ; wherein, is the total number of pixels, is the total number of classes, indicates whether the pixel belongs to class of the true label, is the predicted probability of the dynamic feature calibration segmentation network (4) that the pixel belongs to class , is the class weight.

[0076] Dice loss ; wherein, indicates the number of intersection pixels between the prediction and the real mask, and are the total number of pixels of the prediction and the real mask, is the smoothing coefficient.

[0077] Embodiment 2: For the scenario of embodiment 1, in combination with the system proposed in embodiment 1, the present application also proposes a "remote sensing image water body segmentation method based on hybrid evolutionary dynamic calibration".

[0078] In combination with Figure 4 , the method comprises the following steps:

[0079] S1: Data acquisition and preprocessing: the acquisition module (1) acquires multispectral satellite remote sensing images in real time, and the data preprocessing module (2) performs preprocessing operations such as normalization, data enhancement, and mask cropping to obtain standard size multispectral satellite remote sensing image blocks;

[0080] S2: Data set construction: part of the historical data of standard size multispectral satellite remote sensing image blocks are manually labeled and divided into a training set and a validation set;

[0081] S3: Model building and initialization: build the dynamic feature calibration segmentation network (4) and initialize the hybrid evolutionary optimizer (3), set the initial population and hyperparameter search space of each evolutionary algorithm;

[0082] S4: Hybrid evolutionary hyperparameter optimization: start the hybrid evolutionary optimizer (3) and perform a collaborative search process to output the optimal hyperparameter combination at the current update number;

[0083] S5: Model training: configure the optimal hyperparameter combination obtained in step S4 to the dynamic feature calibration segmentation network (4), and use the dynamic weighted multi-task loss function and the training set to train the weights of the dynamic feature calibration segmentation network (4) using the model training unit (51);

[0084] S6: Iterative optimization: the model evaluation unit (52) periodically evaluates the performance of the dynamic feature calibration segmentation network (4) using the validation set according to the training round number, and if the segmentation accuracy does not reach the preset threshold, returns to step S4 to trigger the hybrid evolutionary optimizer (3) to search for hyperparameters again until the maximum number of updates;

[0085] S7: Water body segmentation and result output: input the standard size multispectral satellite remote sensing image to be recognized into the trained dynamic feature calibration segmentation network (4) for forward propagation, output a binary segmentation map of the same size as the input, and display the binary segmentation map using the water body recognition output module (6) to identify the water body area and complete the water body segmentation task.

[0086] Further, the collection module (1) is connected with the GEE cloud platform API, automatically searches for the corresponding river basin, and downloads Landsat 8 / 9 OLI, Sentinel-2 and other multispectral satellite remote sensing images. The image bands include red, green, blue, near-infrared, and short-wave infrared.

[0087] Further, the data preprocessing module (2) performs normalization, data enhancement (random scaling, flipping, color transformation) and other operations on the multispectral satellite remote sensing image. This module also crops the image and its corresponding artificial annotation mask into standard size tiles and divides them into training, validation and test sets.

[0088] Further, step S4 is specifically:

[0089] S401: In each generation iteration, each algorithm generates a set of hyperparameter candidate solutions based on its inherent mechanism; wherein the inherent mechanism of GA is selection, crossover and mutation; the inherent mechanism of ACO is pheromone guided path selection; the inherent mechanism of PSO is individual and group experience update; and the inherent mechanism of AE is elite solution space disturbance and selection;

[0090] S402: Referring to Table 1, based on the current iteration number Total maximum number of iterations The proportion divides the iteration process into the early stage ( ), medium term ( ) and later ( The weights of each algorithm in joint prediction are adjusted in three stages: in the early stage, the weight of GA is increased to strengthen global exploration; in the middle stage, the weights of ACO and PSO are balanced to form a dual drive of path optimization and gradient-driven; and in the later stage, the weight of AE is increased for fine search of the solution space.

[0091] Table 1. Iterative Process Stages

[0092] Evolutionary algorithm Initial stage Intermediate stage Final stage GA 0.35 0.05 0.00 ACO 0.28 0.28 0.20 PSO 0.32 0.32 0.40 AE 0.05 0.35 0.40

[0093] S403: Each set of hyperparameter candidate solutions is assigned a probability of selection based on the weights of each algorithm in the joint prediction, and selection is randomized according to the probability. Group hyperparameters;

[0094] S404: Input each set of hyperparameters into the dynamic feature calibration segmentation network (4), and calculate the F1-Score as the fitness value on the validation set to guide the evolutionary direction;

[0095] S405: When the maximum number of iterations is reached or the fitness converges, output the optimal combination of hyperparameters for the current number of updates.

[0096] Furthermore, the evaluation metrics of the model evaluation unit (52) include: pixel accuracy, F1-Score, average crossover ratio, average pixel accuracy, and category pixel accuracy based on the confusion matrix.

[0097] To better demonstrate the beneficial effects of the method of the present invention, comparative experiments and ablation experiments were conducted in this embodiment. The experimental results are as follows: Figure 8 , Figure 9 , Figure 10 As shown in the figure. First, the multispectral satellite remote sensing image of the river basin is segmented using the existing U-net network technology, resulting in the image shown in the figure. Figure 8 The results show that the network can roughly achieve water body segmentation, but it is significantly lacking in the ability to capture details of tributary water bodies; while when only the dynamic feature calibration segmentation network (4) of the present invention is used, without considering the hybrid evolutionary optimizer (3) to optimize the hyperparameters, the results are as follows. Figure 9 As shown, although it can compensate for the inability to capture details of river tributaries, its excessive segmentation causes non-water bodies to be misidentified as water bodies; only the method of this invention, which combines a hybrid evolutionary optimizer (3) and a dynamic feature calibration segmentation network (4), can achieve the desired results. Figure 10As shown, only the water body can be segmented accurately, and the ability to capture the details of the tributaries of the river is also possessed.

[0098] Embodiment 3: for the scenario of embodiment 1, Figure 11 A block diagram of an electronic device (90) with which embodiments of the application can be practiced is shown. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, tablets, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. In its most basic form, the electronic device can include one or more processors (91) and a system memory (93), including both read-only memory (ROM) (92) and random access memory (RAM) (93), coupled by an input / output (I / O) interface (95) to one another and to peripheral devices such as input and output devices.

[0099] The electronic device can also represent various forms of mobile devices such as personal digital processors, cellular telephones, smartphones, wearable devices (e.g., headsets, glasses, watches, etc.), and other similar computing devices. The components shown here, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the applications described and / or claimed in this document.

[0100] As Figure 11 As shown, the electronic device (90) includes at least one processor (91), and a memory, such as read-only memory (ROM) (92), random access memory (RAM) (93), etc., connected to the at least one processor (91) in communication. The memory stores a computer program that can be executed by the at least one processor, and the processor (91) can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) (92) or the computer program loaded from the storage unit (98) into the random access memory (RAM) (93). In the RAM 43, various programs and data required for the operation of the electronic device (90) can also be stored. The processor (91), the ROM 42, and the RAM 43 are connected to each other through a bus (94). An input / output (I / O) interface (95) is also connected to the bus (94).

[0101] A plurality of components in the electronic device (90) are connected to the I / O interface (95), including: an input unit (96), such as a keyboard, a mouse, etc.; an output unit (97), such as various types of displays, speakers, etc.; a storage unit (98), such as a magnetic disk, an optical disk, etc.; and a communication unit (99), such as a network card, a modem, a wireless communication transceiver, etc. The communication unit (99) allows the electronic device (90) to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunications networks.

[0102] The processor (91) can be various general-purpose and / or special-purpose processing components having processing and computing capabilities. Some examples of the processor (91) include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processors, controllers, microcontrollers, and the like. The processor (91) performs various methods and processes described above, such as the remote sensing image water body segmentation method based on hybrid evolutionary dynamic calibration.

[0103] In some embodiments, the remote sensing image water body segmentation method based on hybrid evolutionary dynamic calibration can be implemented as a computer program tangibly embodied in a computer readable storage medium, such as the storage unit (98). In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device (90) via the ROM (92) and / or the communication unit (99). When the computer program is loaded onto the RAM (93) and executed by the processor (91), one or more steps of the remote sensing image water body segmentation method based on hybrid evolutionary dynamic calibration described above can be performed. Alternatively, in other embodiments, the processor (91) can be configured to perform the remote sensing image water body segmentation method based on hybrid evolutionary dynamic calibration by any other appropriate means, such as by means of firmware.

[0104] Various implementations of the systems and techniques described above can be realized in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a programmable logic device (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0105] Computer programs for implementing the methods of the present application can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the computer program, when executed, enables the functions / acts specified in the flowcharts and / or block diagrams to be implemented. The computer program can be executed entirely on a machine, partially on a machine, partially on a machine as a standalone software package and partially on a remote machine or entirely on a remote machine or server.

[0106] In the context of the present application, a computer-readable storage medium can be a tangible medium that can contain or store a computer program for use by or in connection with an instruction execution system, apparatus, or device. A computer-readable storage medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium can be a machine-readable signal medium. More specific examples of a machine-readable storage medium will include one or more lines of a program of instructions in a transitory signal, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0107] To provide for interaction with a user, the systems and techniques described here can be implemented on an electronic device having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the electronic device. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0108] The systems and techniques described herein can be implemented in a computing system that includes a back end component, e.g., as a data server, or that includes a middleware component, e.g., an application server, or that includes a front end component, e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described herein, or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication, e.g., a communication network. Examples of communication networks include a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0109] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system, to solve the defects of large management difficulty and weak business scalability in traditional physical host and VPS service.

[0110] Finally, it is to be noted that the above preferred embodiments are merely used to illustrate the technical solutions of the present application but not to limit the present application. Although the present application has been described in detail by the above preferred embodiments, it should be understood by those skilled in the art that various modifications can be made in form and details without departing from the scope of the present application defined by the claims.

Claims

1. A remote sensing image water body segmentation system based on hybrid evolutionary dynamic calibration, characterized in that, The application relates to a water body recognition method based on a dynamic feature calibration segmentation network, which comprises a collection module (1), a data preprocessing module (2), a hybrid evolutionary optimizer (3), a dynamic feature calibration segmentation network (4), a model training and evaluation module (5) and a water body recognition output module (6). The collection module (1) is connected with the input of the data preprocessing module (2) to realize the collection of multispectral satellite remote sensing images. The data preprocessing module (2) is used for preprocessing the multispectral satellite remote sensing images, and the output of the data preprocessing module (2) is connected with the input of the dynamic feature calibration segmentation network (4). The hybrid evolutionary optimizer (3), the dynamic feature calibration segmentation network (4) and the model training and evaluation module (5) are connected in sequence to form a closed loop, and the training of network hyperparameters is realized. The water body recognition output module (6) is a display terminal, the input of the water body recognition output module (6) is connected with the output of the dynamic feature calibration segmentation network (4), and the water body segmentation result of an image is output and displayed.

2. The remote sensing image water body segmentation system based on hybrid evolutionary dynamic calibration according to claim 1, wherein, The hybrid evolutionary optimizer (3) integrates a genetic algorithm (GA), an ant colony optimization (ACO), a particle swarm optimization (PSO) and an Alpha evolutionary algorithm (AE) in parallel, and provides optimal hyperparameter configuration for the dynamic feature calibration segmentation network (4).

3. The remote sensing image water body segmentation system based on hybrid evolutionary dynamic calibration according to claim 1, wherein, The dynamic feature calibration segmentation network (4) takes U-Net as a basic architecture, which is composed of a layer encoder (41), a layer decoder (42), a skip connection (43), a dynamic feature calibration module (44). Each layer of the encoder (41) is composed of a convolution block one (411) and a pooling layer (412) connected in series, is used for extracting multi-level features from an input image, is connected with the data preprocessing module (2), and is connected with the input of the decoder (42). Each layer of the decoder (42) is composed of an up-sampling (421) and a convolution block two (422) connected in series, and realizes image water body segmentation extraction. The up-sampling (421) is a transposed convolution, the convolution block one (411) and the convolution block two (422) are composed of a convolution layer and a ReLU layer connected in series, and each of the jump connections (43) is connected with the output of the convolution block one (411) in the corresponding level in the path of the encoder (41) and the output of the up-sampling (421) in the corresponding last level in the path of the decoder (42), and the outputs are connected with the input of the dynamic feature calibration module (44) in the corresponding level after being spliced by Concat. Each of the dynamic feature calibration modules (44) is connected with the convolution block two (422) in the decoder (42). The convolution block two (422) is connected with the water body recognition output module (6) and outputs an image water body segmentation result. Each of the dynamic feature calibration modules (44) is composed of a spectral attention submodule (441) and a spatial attention submodule (442) connected in parallel and then connected with a dynamic weight generation submodule (443).

4. The remote sensing image water body segmentation system based on hybrid evolutionary dynamic calibration according to claim 3, characterized in that, The spectral attention submodule (441) is composed of a global average pooling layer, a multi-layer perceptron and a nonlinear activation function connected in series, one-dimensional convolution is used for capturing spectral dependence between channels, and the features of wavebands sensitive to water bodies are adaptively enhanced. The spatial attention submodule (442) is composed of a global average pooling layer, a maximum pooling layer, a direction perception convolution and a nonlinear activation function connected in series, the direction perception convolution is used for extracting spatial context information in horizontal and vertical directions, and the continuity of water body boundaries and linear features is strengthened. ​ The dynamic weight generation submodule (443) is a lightweight network that dynamically generates spatially adaptive attention intensity coefficients and feature retention coefficients according to the complexity of input features, and realizes accurate feature calibration according to regional requirements.

5. The remote sensing image water body segmentation system based on hybrid evolutionary dynamic calibration according to claim 3, characterized in that, For the feature map output by the convolution block one (411) in any one layer of the encoder (41) The output of the spectrum attention sub-module (441) is Wherein is a global average pooling operation, is a multi-layer perception, is a nonlinear activation function; The output of the spatial attention sub-module (442) is ; wherein, is a max-pooling operation; is a direction-aware convolution operation, specifically, using one horizontal direction convolution kernel and one vertical direction convolution kernel to extract context information in different directions respectively, then concatenating the output results of the two directions, and then fusing through a standard convolution; The output of the dynamic weight generation submodule (443) is ; wherein, is an element-wise multiplication; ; The operation is composed of a global average pooling concatenated with two parallel convolution layers of , and the output satisfies the constraint condition of .

6. The system of claim 1, wherein, The model training and evaluation module (5) is composed of a model training unit (51) and a model evaluation unit (52) in series; the model training unit (51) configures the dynamic feature calibration segmentation network (4) with the optimal hyperparameter obtained by the hybrid evolutionary optimizer (3), and trains using the training set; a linearly weighted multi-task loss function is used in the training, and the weight coefficient is dynamically adjusted by the evolutionary strategy; the model evaluation unit (52) regularly evaluates the model performance using the validation set during the training process, and if the evaluation result does not reach the preset threshold, the hybrid evolutionary optimizer (3) is triggered to search for hyperparameters and train the model again, forming a closed-loop optimization.

7. The system of claim 1, wherein, The model training unit (51) uses a linearly weighted multi-task loss function, which is a linear weighting of cross-entropy loss and Dice loss.

8. The method for water body segmentation of remote sensing image based on hybrid evolutionary dynamic calibration, characterized in that, The method comprises the following steps: S1: data acquisition and preprocessing: real-time acquisition of multispectral satellite remote sensing images, preprocessing operations such as normalization, data enhancement and mask cropping to obtain standard size multispectral satellite remote sensing image blocks; S2: data set construction: manually label part of the historical data of standard size multispectral satellite remote sensing image blocks, and divide them into a training set and a validation set; S3: model construction and initialization: build a dynamic feature calibration segmentation network (4) and initialize a hybrid evolutionary optimizer (3), and set the initial population and hyperparameter search space of each evolutionary algorithm; S4: hybrid evolutionary hyperparameter optimization: start the hybrid evolutionary optimizer (3), perform a collaborative search process, and output the optimal hyperparameter combination at the current update time; S5: model training: configure the optimal hyperparameter combination obtained in step S4 to the dynamic feature calibration segmentation network (4), and use the dynamic weighting multi-task loss function to train the weights of the dynamic feature calibration segmentation network (4) using the training set; S6: iterative optimization: the model evaluation unit (52) regularly evaluates the performance of the dynamic feature calibration segmentation network (4) using the validation set according to the training round number, and if the segmentation accuracy does not reach the preset threshold, returns to step S4 and triggers the hybrid evolutionary optimizer (3) to search for hyperparameters again until the maximum update time; S7: water body segmentation and result output: input the standard size multispectral satellite remote sensing image block to be recognized into the trained dynamic feature calibration segmentation network (4) for forward propagation, output a binary segmentation map with the same size as the input, and display the binary segmentation map by the water body recognition output module (6) to identify the water body area and complete the water body segmentation task.

9. The method of claim 8, wherein, The step S4 is specifically: S401: In each iteration, each algorithm generates a set of hyperparameter candidate solutions in parallel based on its inherent mechanism; wherein the inherent mechanism of GA is selection, crossover and mutation; the inherent mechanism of ACO is pheromone-guided path selection; the inherent mechanism of PSO is individual and group experience update; the inherent mechanism of AE is disturbance and selection of elite solution space; S402: According to the proportion of the current iteration number and the total maximum iteration number, the iteration process is divided into three stages of early, middle and late to adjust the weight of each algorithm in joint prediction; wherein the weight of GA is increased in early stage to strengthen global exploration; the weights of ACO and PSO are balanced in middle stage to form dual driving of path optimization and gradient driving; the weight of AE is improved in late stage for fine search of solution space; S403: Each group of hyperparameter candidate solutions is assigned a selected probability according to the weight of each algorithm in the joint prediction, and is randomly selected according to the probability group hyperparameters; S404: Each set of hyperparameters is input into a dynamic feature calibration segmentation network (4), and F1-Score is calculated on the validation set as fitness value to guide the evolution direction; S405: When the maximum iteration number is reached or the fitness converges, the optimal hyperparameter combination under the current update number is output.

10. An electronic device and a computer readable storage medium, characterized by The electronic device includes at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores a computer program executed by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the remote sensing image water body segmentation method based on mixed evolution dynamic calibration according to any one of claims 8-9; the computer readable storage medium stores computer instructions for enabling the processor to execute the remote sensing image water body segmentation method based on mixed evolution dynamic calibration according to any one of claims 8-9.

Images