A method and system for rapidly evaluating the suitability of fish habitats based on deep learning.

Abstract

This invention provides a method and system for rapidly evaluating the suitability of fish habitats for various environments based on deep learning, which belongs to the field of river ecology technology. [Solution] The method includes the steps of: conducting a simulation using a physical mechanism model of fish habitat suitability, obtaining evaluation results of the physical mechanism model of fish habitat suitability in the river channel corresponding to various flow rate conditions, and constructing a sample set for learning fish habitat suitability; training a pre-constructed deep learning model for predicting fish habitat suitability in the river channel using the sample set for learning fish habitat suitability; and identifying the flow rate influence sensitivity range and constructing a sample set for incremental learning fish habitat suitability.

Description

[Technical Field]

[0001] This invention relates to the technical fields of hydraulic engineering and ecological environment conservation, and more particularly to a method and system for rapidly evaluating the suitability of fish habitats based on deep learning. [Background technology]

[0002] In the fields of ecological environmental science and water resource management, the process of calculating the suitability of fish habitats for various downstream discharge rates and different hydrological cycles is an extremely important fundamental step when conducting comparative studies of ecologically conscious water flow control operations for hydroelectric power generation in watersheds, or when researching restoration and maintenance schemes to improve the suitability of fish habitats.

[0003] Currently widely used methods for evaluating the suitability of fish habitats require the construction of hydraulic models and sequential calculation of the suitability of fish habitats under various operational conditions. This results in long computation times and high costs, making it difficult to meet the requirements for rapidly processing large amounts of discrete data in sophisticated simulations of ecosystem-conscious water flow control operations for hydroelectric power generation in river basins. To address this challenge, researchers have proposed habitat suitability curve fitting methods, which use hydrological statistical indicators such as cumulative flow frequency to simulate and predict habitat suitability under different flow conditions. However, these methods have limitations in accuracy, making it difficult to accurately evaluate the spatial distribution of fish habitat suitability within river channels. [Overview of the Initiative] [Problems that the invention aims to solve]

[0004] Focusing on the shortcomings inherent in conventional technology, the present invention provides a method and system for rapidly evaluating the suitability of fish habitats based on deep learning, which can effectively solve the above-mentioned problems. [Means for solving the problem]

[0005] This invention employs the following technical means.

[0006] The present invention provides a method for rapidly evaluating the suitability of fish habitats based on deep learning, which consists of the following steps: Step S1: Analyze the flow history data of the target river channel area over multiple years to determine the range [Q min , Q max of the flow conditions, where Q min and Q max respectively represent the upper and lower limit values of the range of the flow conditions. Step S2: Select K flow conditions according to the flow interval △Q min , Q max within the range [Q acture , and represent them as Q k (k = 1, 2,..., K). Step S3: Perform simulations using the physical mechanism model of fish habitat suitability to obtain the evaluation results Result k of the physical mechanism model of fish habitat suitability in the river channel corresponding to various flow conditions Q acture,K . Step S4: Based on the evaluation results Result acture,K of the physical mechanism model of fish habitat suitability in the river channel, construct a learning sample set D = {(Q k, Result acture,k )} for fish habitat suitability. Step S5: Use the learning sample set D = {(Q k, Result acture,k )} for fish habitat suitability to train a pre-constructed deep learning model for predicting fish habitat suitability in the river channel, and generate a once-trained deep learning model for predicting fish habitat suitability in the river channel. Step S6: In the range [Qmin, Qmax] of the flow conditions, select L flow conditions again at high density according to a flow interval △Q acture smaller than the flow interval △Q * prediction , and represent them as Q1 (l = 1, 2,..., L). Simulations were conducted using a pre-trained deep learning model for predicting the suitability of fish habitats within river channels, and the results of evaluating the deep learning model regarding the suitability of fish habitats within river channels corresponding to various flow rate conditions Q1 were obtained. prediction,l The results obtained are an evaluation of a deep learning model regarding the suitability of fish habitats within river channels. prediction,l We analyzed the flow rate sensitivity range interval that shows a significant impact on the suitability of the fish habitat within the river channel. sensitive =[Q * min, Q * A step of identifying max, Q * min and Q * max These indicate the upper and lower limits of the flow rate influence sensitivity range, respectively. Step S7: Sample set D for incremental learning of fish habitat suitability * This is a step in building, Identified flow rate effect sensitivity range interval sensitive =[Q * min, Q * In max, the incremental learning interval is △Q prediction According to K * Select the flow rate conditions for type K, and from the flow rate conditions for type K selected in step S2, K ** The flow rate conditions for each species were randomly sampled, and furthermore, △Q prediction <△Q acture And, K * Species flow rate conditions and K ** By combining the flow rate conditions of different species, a set of flow rate conditions for incremental learning is constructed. Various flow rate conditions from the set of flow rate conditions for incremental learning are input into the physical mechanism model of fish habitat suitability, and simulations are performed to obtain fish habitat suitability evaluation results in the river channel corresponding to various flow rate conditions. These are then combined to create a sample set D for incremental learning of fish habitat suitability. ** Constitute, Step S8: Sample set D for incremental learning of fish habitat suitability *The process involves using a first-stage trained deep learning model for predicting the suitability of fish habitats within river channels to perform a second-stage incremental learning to generate a retrained deep learning model for predicting the suitability of fish habitats within river channels. Step S9: A step in which the suitability of the fish habitat in the river channel under each flow rate condition in the target river channel area is predicted using a retrained deep learning model for predicting the suitability of the fish habitat in the river channel.

[0007] Preferably, in step S3, the physical mechanism model for fish habitat suitability includes a two-dimensional hydraulic model, a model for the renewal of individual fish movements, a two-dimensional distribution density calculation model for fish schools, a model for calculating fish habitat suitability within the river channel, and a model for calculating the area of ​​fish habitat suitability within the river channel.

[0008] Preferably, step S3 is configured as follows: Step S3.1: Conduct simulations using a two-dimensional hydraulic model, and under various flow rate conditions Q k A step to obtain the distribution of hydrological environmental factors corresponding to the Step S3.2: Taking into account the distribution of the hydraulic environmental factors, the interactions between fish schools, and the random perturbation term, the position vector X of each individual fish i in the fish school at the next time step t+1 is calculated using the individual fish movement update model. t+1 i,k By simulating this, the movement behavior of each individual fish i is simulated. Step S3.3: When the positional distribution of the fish school reaches a relatively stable state, the step of calculating the two-dimensional distribution density p(x,y) of the fish school using the two-dimensional distribution density calculation model for the fish school. Step S3.4: Based on the two-dimensional distribution density p(x,y) of the fish school, estimate the fish habitat suitability distribution HSI(x,y) within the river channel using the fish habitat suitability calculation model in equation (1).

[0009]

number

[0010] In the formula, R represents the computational domain, and (x,y) represents the coordinates of the x and y axes of the computational domain R.

[0011] Step S3.5: The HSI(x,y) distribution of suitable fish habitats within the river channel is discretized into the HSI values ​​of suitable fish habitats within the river channel. acture,k,e A step to convert, The computational domain R is discretized into E computational cell grids, and the HSI (x,y) distribution of fish habitat suitability within the river channel is used to determine the HSI (HSI) value of fish habitat suitability within the river channel for each computational cell grid e. acture,k,e Identify (e=1,2,…,E), Step S3.6: Using the model for calculating the suitable area of ​​fish habitat within the river channel according to equation (2), the flow rate condition Q k The WUA (Wide Area of ​​Area) corresponding to the suitable fish habitat area within the river channel. acture,k A step to estimate.

[0012]

number

[0013] In the formula, A represents the area of ​​each computational cell grid e.

[0014] Step S3.7: The above process determines various flow rate conditions Q k Results of the evaluation of the physical mechanism model for the suitability of the fish habitat environment within the river channel. acture,k ={HSI acture,k,e ,WUA acture,k Steps to obtain}.

[0015] Preferably, simulations are performed using a two-dimensional hydraulic model, and various flow rate conditions Q are examined. k To obtain the distribution of hydrological environmental factors corresponding to this, basic data including a digital elevation model of the target river channel area and the distribution of the river channel roughness coefficient are required. The various flow rate conditions Q mentioned above k The distribution of corresponding hydraulic environmental factors includes flow velocity, water depth, and channel sediment distribution.

[0016] Preferably, step S3.2 is configured as follows: Step S3.2.1: The motion update model for the fish individual includes a speed update model for the fish individual and a position update model for the fish individual. Various flow conditions Q k In this simulation, the velocity update model for individual fish shown in equation (3) was used to obtain the velocity vector V of each individual fish i at time t. t i,k The step of calculating.

[0017]

number

[0018] TIFF0007849830000005.tif51166

[0019] TIFF0007849830000006.tif17166

[0020]

number

[0021] TIFF0007849830000008.tif173166

[0022] TIFF0007849830000009.tif18166

[0023]

number

[0024] TIFF0007849830000011.tif125166

[0025] TIFF0007849830000012.tif17166

[0026]

number

[0027] In the formula, u(0,v pref ) is the swimming speed v based on the fish's preference for reverse currents, from 0. pref The velocity is randomly sampled within the range up to , and u(0,2π) represents the value of the swimming direction randomly sampled within a 360° range.

[0028] Step S3.2.2: Various flow rate conditions Q k Below, a simulation was performed using the fish individual position update model shown in equation (7), and the position vector X of each fish individual i at the next time step t+1 was obtained. t+1 i,k Steps to obtain.

[0029]

number

[0030] In the formula, X t i,k This shows the position vector of fish individual i at time t.

[0031] Preferably, step S3.3: When the positional distribution of the fish school reaches a relatively stable state, the step of obtaining the two-dimensional distribution density p(x,y) of the fish school is, specifically, The two-dimensional distribution density p(x,y) of the fish school is calculated using equation (8).

[0032]

number

[0033] In the formula, n is the total number of fish individuals in the school, σ is the smoothing parameter, and (x,y) represents the coordinates of the x and y axes of the computational domain R. i ,y i The horizontal and vertical coordinates of individual fish i when the positional distribution of the fish school reaches a relatively stable state represent the position vector of individual fish i.

[0034] Preferably, the deep learning model for predicting the suitability of the fish habitat in the river channel comprises a data input and preprocessing module, a convolution encoding and attention mechanism module, a decoding and sampling module, and a habitat suitability prediction output module. The data input and preprocessing module receives topographic data and flow condition data of the target river channel area, performs preprocessing, the topographic data is two-dimensional grid data, the preprocessing step includes extracting high-dimensional features from the flow condition data using a multi-layer fully connected network, and expanding the flow condition data into a two-dimensional matrix matching the size of the topographic data, the processed topographic data, flow condition data and environmental data are joined in the channel dimension to form integrated input data, The convolutional encoding and attention mechanism module extracts features from the input data integrated via a convolutional neural network, obtains feature vectors, simultaneously applies weighting to these feature vectors using the attention mechanism to generate an attention feature map, and applies this to the output of the convolutional layer to enhance the representation of important features. The decoding and sampling module performs upsampling on the attention feature map processed by the attention mechanism using a deconvolutional network to obtain the upsampled feature map, which is then restored to the same resolution as the input terrain data. The habitat suitability prediction output module obtains evaluation results of a deep learning model regarding the suitability of fish habitats within river channels by processing the feature map after upsampling. The deep learning model for predicting the suitability of fish habitats within the river channel is configured to optimize the prediction accuracy of the deep learning model for predicting the suitability of fish habitats within the river channel using a custom loss function, and the total loss function is defined by equation (9).

[0035]

number

[0036] In the formula, α, β, and γ represent the weighting coefficients, L total L is the total loss function. habitat-RMSE L is the mean square root error loss function. habitat-CE L is the cross-entropy loss function. area-logMAE This shows the logarithmic absolute error loss function.

[0037] These are calculated using equations (10), (11), and (12), respectively.

[0038]

number

[0039]

number

[0040]

number

[0041] In the ceremony, HSI prediction,k,e This is the flow rate condition Q k At that time, the deep learning model for predicting the suitability of the fish habitat in the river channel outputs the fish habitat suitability value in the computational cell grid e, and HSI acture,k,e Flow rate condition Q k At that time, the physical mechanism model for fish habitat suitability outputs the fish habitat suitability value in the river channel in the computational cell grid e, and H acture,k,e Flow rate condition Q k At that time, the HSI value of fish habitat suitability in the river channel in the computational cell grid e output by the physical mechanism model of fish habitat suitability is acture,k,e The result of binarizing the data is shown below.

[0042] TIFF0007849830000020.tif20166

[0043] WUA prediction,k This is the flow rate condition Q kWhen the deep learning model for predicting the suitability of the fish habitat environment in the river channel outputs the suitable area of the fish habitat environment in the river channel, WUA acture,k is the flow condition Q k When, it represents the suitable area of the fish habitat environment in the river channel output by the physical mechanism model of the fish habitat environment suitability.

[0044] Preferably, by the following method, the flow influence sensitivity range [Q * min, Q * max] that shows a significant impact on the suitability of the fish habitat environment in the river channel is identified, Step S6.1: The flow influence sensitivity range Interval WUA =[Q WUA,low , Q WUA,high in which the suitable area of the fish habitat environment in the river channel changes is identified, and Q WUA,low and Q WUA,high respectively represent the upper limit value and the lower limit value of the flow influence sensitivity range in which the suitable area of the fish habitat environment in the river channel changes. Step S6.1.1: Perform simulations using the pre-trained deep learning model for predicting the suitability of the fish habitat environment in the river channel, and obtain the evaluation result Result prediction,l of the deep learning model regarding the suitability of the fish habitat environment in the river channel corresponding to various flow conditions Q1. The evaluation result includes the suitable area of the fish habitat environment in the river channel WUA prediction,l and the suitability value HSI prediction,l,e of the fish habitat environment in the river channel in each calculation cell grid e. TIFF0007849830000021.tif26166

[0045]

Number

[0046] TIFF0007849830000023.tif54166

[0047] Step S6.1.3: At the flow condition Q1 (l = 1, 2,..., L), for any two adjacent flow conditions Q α and Q βFor the local difference quotient Diff WUA,α-β relating to the suitable area of the fish habitat environment in the river channel, it is calculated using Equation (14).

[0048]

Equation

[0049] In the formula, WUA prediction,α and WUA prediction,β respectively represent the suitable area of the fish habitat environment in the river channel output by the deep learning model for predicting the suitability of the fish habitat environment in the river channel under the flow conditions Q α and Q β respectively.

[0050] TIFF0007849830000025.tif83166

[0051]

Equation

[0052] TIFF0007849830000027.tif46166

[0053] Step S6.2.2: At the flow conditions Q1 (l = 1, 2,..., L), for any two adjacent flow conditions Q α and Q β the local difference quotient Diff HSI,α-β relating to the distribution of the suitable area of the fish habitat environment in the river channel is calculated using Equation (16).

[0054]

Equation

[0055] In the formula, RHSI prediction,α and RHSI prediction,β respectively represent the flow conditions Q α and Q βAt that time, the distribution area of ​​fish habitat suitability within the river channel is shown, statistically obtained based on the fish habitat suitability values ​​within the river channel output by the deep learning model for predicting fish habitat suitability within the river channel.

[0056] TIFF0007849830000029.tif77166

[0057]

number

[0058] In the formula, Interval sensitive This range was identified as a flow rate sensitivity range that shows a significant impact on the suitability of the fish habitat within the river channel.

[0059] The present invention also provides a system for realizing the rapid evaluation method of fish habitat suitability based on deep learning, the system comprising: a first flow rate condition determination unit; a primary learning unit for a physical mechanism model of fish habitat suitability; a unit for acquiring a sample set for learning fish habitat suitability; a deep learning model for predicting fish habitat suitability in a river channel; a second flow rate condition determination unit; a flow rate influence sensitivity range; a unit for acquiring a sample set for incremental learning of fish habitat suitability; a second-stage incremental learning unit; The first flow rate condition determination unit analyzes multi-year flow rate history data for the target river channel area to determine the range of flow rate conditions [Qmin, Qmax], where Q min and Q max These indicate the upper and lower limits of the flow rate condition range, and within the flow rate condition range [Qmin, Qmax], the flow rate interval △Q acture Accordingly, select flow rate conditions of type K, and these are set to flow rate condition Q k Expressed as (k=1,2,…,K), The primary learning unit of the physical mechanism model for fish habitat suitability is the learning sample set D={(Q k , Result acture,kThis method uses )} to train a pre-built deep learning model for predicting the suitability of fish habitats within river channels, thereby generating a first-stage trained deep learning model for predicting the suitability of fish habitats within river channels. The unit for acquiring sample sets for learning fish habitat suitability evaluates the physical mechanism model results of fish habitat suitability within the river channel. acture,k Based on this, the sample set D for learning fish habitat suitability is {(Q k , Result acture,k This is for constructing )}, A deep learning model for predicting the suitability of fish habitats within river channels uses a sample set D={(Q)} for learning fish habitat suitability. k , Result acture,k This is a method for generating a pre-trained deep learning model for predicting the suitability of fish habitats within river channels by using )} for training. The second flow rate condition determination unit determines the flow rate interval △Q within the range of the flow rate conditions [Qmin, Qmax]. acture Smaller flow interval △Q * prediction Accordingly, the purpose is to select L-type flow rate conditions again at high density and express them as flow rate conditions Q1 (l=1,2,…,L), The flow rate influence sensitivity range discrimination unit performs simulations using a pre-trained deep learning model for predicting the suitability of fish habitats within river channels, and evaluates the deep learning model's assessment of fish habitat suitability within river channels corresponding to various flow rate conditions Q1. prediction,l The results obtained are an evaluation of a deep learning model regarding the suitability of fish habitats within river channels. prediction,l We analyzed the flow rate sensitivity range interval that shows a significant impact on the suitability of the fish habitat within the river channel. sensitive =[Q * min, Q * This is for identifying max, Q * min and Q * max These indicate the upper and lower limits of the flow rate influence sensitivity range, respectively. The sample acquisition unit for incremental learning of fish habitat suitability identifies the flow rate effect sensitivity range interval.sensitive =[Q * min, Q * In max, the incremental learning interval is △Q prediction According to K * Select the flow rate conditions for type K, and from the flow rate conditions for type K ** The flow rate conditions for each species were randomly sampled, and furthermore, △Q prediction >△Q acture And, K * Species flow rate conditions and K ** By combining the flow rate conditions of different species, a set of flow rate conditions for incremental learning is constructed. Various flow rate conditions from the set of flow rate conditions for incremental learning are input into the physical mechanism model of fish habitat suitability, and simulations are performed to obtain fish habitat suitability evaluation results in the river channel corresponding to various flow rate conditions. These are then combined to create a sample set D for incremental learning of fish habitat suitability. * It is intended to constitute, The second-stage incremental learning unit is sample set D for incremental learning of fish habitat suitability. * This method involves performing a second stage of incremental learning on a first-stage trained deep learning model for predicting the suitability of fish habitats within a river channel, generating a retrained deep learning model for predicting the suitability of fish habitats within a river channel, and then using this retrained deep learning model to predict the suitability of fish habitats within the river channel under various flow rate conditions in the target river channel area. [Effects of the Invention]

[0060] The deep learning-based method and system provided by the present invention for rapidly evaluating the suitability of fish habitats has the following excellent advantages: Based on modules such as convolutional encoding and attention mechanisms, the present invention constructs a highly efficient deep learning model for predicting the suitability of fish habitats in river channels. By autonomously learning based on data samples simulated by a physical mechanism model, it establishes a correspondence between the distribution of suitable fish habitats in river channels and factors such as flow rate, thereby achieving rapid and highly accurate evaluation of the distribution of suitable fish habitats in river channels. Compared to conventional prediction techniques, the present invention can significantly reduce the time and cost burden and has high practicality and broad applicability in fields such as river ecosystem conservation and water resource management. [Brief explanation of the drawing]

[0061] [Figure 1] This is a flowchart of the rapid evaluation method for fish habitat suitability based on deep learning, according to the present invention. [Figure 2] This figure shows an example of matrix-formatted terrain data according to an embodiment of the present invention. [Figure 3] This figure shows a comparison of the distribution of suitable habitats for fish and the area of ​​suitable habitats for fish, as predicted by a deep learning model for predicting the suitability of fish habitats in river channels and a physical mechanism model for fish habitat suitability, according to an embodiment of the present invention. [Figure 4] This figure shows the relative error results of the suitable area for fish habitats predicted by a deep learning model for predicting the suitability of fish habitats in river channels, according to an embodiment of the present invention. [Modes for carrying out the invention]

[0062] In order to more clearly understand the technical problems, technical means, and advantageous effects that the present invention aims to solve, the present invention will be described in more detail below with reference to the accompanying drawings and embodiments. The specific embodiments described herein are merely examples to aid in understanding the present invention and do not limit it.

[0063] This invention improves the fitting performance of fish habitat suitability while reducing the computation time of the physical model by performing a simulation of the physical mechanism of habitat suitability based on intelligent fish individuals, constructing a sample dataset concerning the distribution of fish habitat suitability within river channels, constructing a deep learning model for predicting fish habitat suitability based on a deep learning mechanism, and training and testing the deep learning model using an incremental learning method.

[0064] Please refer to Figure 1. The present invention provides a method for rapidly evaluating the suitability of fish habitats based on deep learning. The method includes the following steps.

[0065] Step S1: A step in which the flow rate history data of the target river channel area over several years is analyzed and the range of flow rate conditions [Qmin, Qmax] is determined, wherein Q min and Q max These indicate the upper and lower limits of the flow rate conditions, respectively. Step S2: Within the range of the flow rate conditions [Qmin, Qmax], the flow rate interval △Q acture Accordingly, select the flow rate conditions for type K, and these are Q k Steps represented as (k=1,2,…,K), Step S3: Conduct simulations using a physical mechanism model of fish habitat suitability, and under various flow rate conditions Q. k Results of the evaluation of the physical mechanism model for the suitability of the fish habitat environment within the river channel. acture,k Steps to obtain The physical mechanism model for fish habitat suitability is a novel model according to the present invention and includes a two-dimensional hydraulic model, a model for the renewal of individual fish movements, a two-dimensional distribution density calculation model for fish schools, a model for calculating fish habitat suitability within river channels, and a model for calculating the area of ​​fish habitat suitability within river channels.

[0066] Step S3 is structured as follows: Step S3.1: Conduct simulations using a two-dimensional hydraulic model, and under various flow rate conditions Q k A step to obtain the distribution of hydrological environmental factors corresponding to the In this step, simulations are performed using a two-dimensional hydraulic model, and various flow rate conditions Q are examined. k To obtain the distribution of hydrological environmental factors corresponding to this, basic data including a digital elevation model of the target river channel area and the distribution of the river channel roughness coefficient are required. The various flow rate conditions Q mentioned above k The distribution of corresponding hydraulic environmental factors includes flow velocity, water depth, and channel sediment distribution.

[0067] Step S3.2: Taking into account the distribution of the hydraulic environmental factors, the interactions between fish schools, and the random perturbation term, the position vector X of each individual fish i in the fish school at the next time step t+1 is calculated using the individual fish movement update model. t+1 i,k By simulating this, the movement behavior of each individual fish i is simulated. Step S3.2 is structured as follows: Step S3.2.1: The motion update model for the fish individual includes a speed update model for the fish individual and a position update model for the fish individual. Various flow conditions Q k In this simulation, the velocity update model for individual fish shown in equation (3) was used to obtain the velocity vector V of each individual fish i at time t. t i,k The step of calculating.

[0068]

number

[0069] TIFF0007849830000032.tif51166

[0070] TIFF0007849830000033.tif20166

[0071]

number

[0072] TIFF0007849830000035.tif190166

[0073] TIFF0007849830000036.tif19166

[0074]

number

[0075] TIFF0007849830000038.tif125166

[0076] TIFF0007849830000039.tif17166

[0077]

number

[0078] In the formula, u(0,v pref ) is the swimming speed v based on the fish's preference for reverse currents, from 0. pref The velocity is randomly sampled within the range up to , and u(0,2π) represents the value of the swimming direction randomly sampled within a 360° range.

[0079] Step S3.2.2: Various flow rate conditions Q k Below, a simulation was performed using the fish individual position update model shown in equation (7), and the position vector X of each fish individual i at the next time step t+1 was obtained. t+1 i,k Steps to obtain.

[0080]

number

[0081] In the formula, X t i,k This shows the position vector of fish individual i at time t.

[0082] Step S3.3: When the positional distribution of the fish school reaches a relatively stable state, the step of calculating the two-dimensional distribution density p(x,y) of the fish school using the two-dimensional distribution density calculation model for the fish school. The two-dimensional distribution density p(x,y) of the fish school is calculated using equation (8).

[0083]

number

[0084] In the formula, n is the total number of fish individuals in the school, σ is the smoothing parameter, and (x,y) represents the coordinates of the x and y axes of the computational domain R. i ,y i The horizontal and vertical coordinates of individual fish i when the positional distribution of the fish school reaches a relatively stable state represent the position vector of individual fish i.

[0085] Step S3.4: Based on the two-dimensional distribution density p(x,y) of the fish school, estimate the fish habitat suitability distribution HSI(x,y) within the river channel using the fish habitat suitability calculation model in equation (1).

[0086]

number

[0087] In the formula, R represents the computational domain, and (x,y) represents the coordinates of the x and y axes of the computational domain R.

[0088] In this invention, the normalized two-dimensional distribution density of fish schools is used as the suitable distribution of fish habitats within the river channel.

[0089] Step S3.5: The continuous fish habitat suitability distribution HSI(x,y) within the river channel is converted to the discretized fish habitat suitability value HSI within the river channel. acture,k,e A step to convert, Specifically, the computational domain R is discretized into E computational cell grids, and the HSI (x,y) of the fish habitat suitability distribution within the river channel is calculated for each computational cell grid e. acture,k,e Identify (e=1,2,…,E), Step S3.6: Using the model for calculating the suitable area of ​​fish habitat within the river channel according to equation (2), the flow rate condition Q k The WUA (Wide Area of ​​Area) corresponding to the suitable fish habitat area within the river channel. acture,k A step to estimate.

[0090]

number

[0091] In the formula, A represents the area of ​​each computational cell grid e.

[0092] Step S3.7: The above process determines various flow rate conditions Q k Results of the evaluation of the physical mechanism model for the suitability of the fish habitat environment within the river channel. acture,k ={HSI acture,k,e ,WUA acture,k Steps to obtain}.

[0093] Step S4: Evaluation results of the physical mechanism model for the suitability of the fish habitat environment within the river channel. acture,k Based on this, the sample set D for learning fish habitat suitability is {(Q k , Result acture,k Steps to construct )} Step S5: Sample set D for learning fish habitat suitability = {(Q k , Result acture,k The steps include: using )} to train a pre-built deep learning model for predicting the suitability of fish habitats in river channels, thereby generating a first-stage trained deep learning model for predicting the suitability of fish habitats in river channels; In the present invention, the deep learning model for predicting the suitability of the fish habitat environment within the river channel comprises a data input and preprocessing module, a convolution encoding and attention mechanism module, a decoding and sampling module, and a habitat environment suitability prediction output module. The data input and preprocessing module receives topographic data and flow condition data of the target river channel area, performs preprocessing, the topographic data is two-dimensional grid data, the preprocessing step includes extracting high-dimensional features from the flow condition data using a multi-layer fully connected network, and expanding the flow condition data into a two-dimensional matrix matching the size of the topographic data, the processed topographic data, flow condition data and environmental data are joined in the channel dimension to form integrated input data, The convolutional encoding and attention mechanism module extracts features from the input data integrated via a convolutional neural network, obtains feature vectors, simultaneously applies weighting to these feature vectors using the attention mechanism to generate an attention feature map, and applies this to the output of the convolutional layer to enhance the representation of important features. The decoding and sampling module performs upsampling on the attention feature map processed by the attention mechanism using a deconvolutional network to obtain the upsampled feature map, which is then restored to the same resolution as the input terrain data. Specifically, the decoding is performed using mathematical formulas.

[0094]

number

[0095] In the formula, F up This shows a deconvolutional network, σ deconv is the activation function, and here we adopt ReLU, F input is the input feature map to the deconvolutional network, K' is the size of the convolutional kernel, S' is the stride, P' is the padding, b deconv This represents the bias term.

[0096] The habitat suitability prediction output module obtains evaluation results of a deep learning model regarding the suitability of fish habitats within river channels by processing the feature map after upsampling. The deep learning model for predicting the suitability of fish habitats within the river channel is configured to optimize the prediction accuracy of the deep learning model for predicting the suitability of fish habitats within the river channel using a custom loss function, and the total loss function is defined by equation (9).

[0097]

number

[0098] In the formula, α, β, and γ represent the weighting coefficients, respectively, and the total loss function is L. habitat-RMSE , L habitat-CE and L area-log MAE It is a weighted sum of, L total This shows the total loss function. L habitat-RMSE This is the mean square root error loss function, which evaluates the continuous difference between the fish habitat suitability value in the river channel output by the deep learning model for predicting fish habitat suitability in the river channel and the fish habitat suitability value in the river channel output by the physical mechanism model for fish habitat suitability. L habitat-CE This is the cross-entropy loss function, which evaluates the classification results of the suitable distribution of fish habitats. L area-log MAE This is a logarithmic absolute error loss function that evaluates the continuous difference between the area of ​​suitable fish habitats in the river channel output by the deep learning model for predicting fish habitat suitability within the river channel and the area of ​​suitable fish habitats within the river channel output by the physical mechanism model for fish habitat suitability.

[0099] L habitat-RMSE , L habitat-CE and L area-log MAE These are calculated using equations (10), (11), and (12), respectively.

[0100]

number

[0101]

number

[0102]

number

[0103] In the ceremony, HSI prediction,k,e This is the flow rate condition Q k At that time, the deep learning model for predicting the suitability of the fish habitat in the river channel outputs the fish habitat suitability value in the computational cell grid e, and HSI acture,k,e Flow rate condition Q k At that time, the physical mechanism model for fish habitat suitability outputs the fish habitat suitability value in the river channel in the computational cell grid e, and H acture,k,e Flow rate condition Q k At that time, the HSI value of fish habitat suitability in the river channel in the computational cell grid e output by the physical mechanism model of fish habitat suitability is acture,k,e The result of binarizing the data is shown below.

[0104] TIFF0007849830000050.tif20166

[0105] WUA prediction,k This is the flow rate condition Q k At that time, the deep learning model for predicting the suitability of the fish habitat in the river channel outputs the area of ​​suitable fish habitat in the river channel, WUA acture,k Flow rate condition Q k At that time, the area of ​​the fish habitat suitable for fish within the river channel, as output by the physical mechanism model for fish habitat suitability, is shown.

[0106] In practical applications, Bayesian optimization techniques can be used to optimize the model's hyperparameters and minimize the total loss function on the validation dataset.

[0107] Step S6: Within the range of the flow rate conditions [Qmin, Qmax], the flow rate interval △Q acture Smaller flow interval △Q * prediction Accordingly, we again selected L-type flow rate conditions at high density and expressed them as Q1 (l=1,2,…,L), Simulations were conducted using a pre-trained deep learning model for predicting the suitability of fish habitats within river channels, and the results of evaluating the deep learning model regarding the suitability of fish habitats within river channels corresponding to various flow rate conditions Q1 were obtained. prediction,l The results obtained are an evaluation of a deep learning model regarding the suitability of fish habitats within river channels. prediction,l We analyzed the flow rate sensitivity range interval that shows a significant impact on the suitability of the fish habitat within the river channel. sensitive =[Q * min, Q * A step of identifying max, Q * min and Q * max These indicate the upper and lower limits of the flow rate influence sensitivity range, respectively. Specifically, the following method was used to determine the flow rate sensitivity range [Q] that shows a significant impact on the suitability of the fish habitat within the river channel. * min, Q * Identify max], Step S6.1: Interval of the flow rate-sensitive range in which the suitable area of ​​fish habitat within the river channel changes. WUA =[Q WUA,low ,Q WUA,high Identify ] and Q WUA,low and Q WUA,high These values ​​represent the upper and lower limits of the flow rate sensitivity range, where the suitable area for fish habitats within the river channel changes. Step S6.1.1: Conduct a simulation using a pre-trained deep learning model for predicting the suitability of fish habitats in river channels, and evaluate the deep learning model's assessment of the suitability of fish habitats in river channels for various flow rate conditions Q1. prediction,l The evaluation results include the WUA (Whole Usable Area) of suitable fish habitats within the river channel. prediction,l and the HSI (Hyper-Saturated Index) of the fish habitat suitability value within the river channel in each calculation cell grid e. prediction,l,e It includes, TIFF0007849830000051.tif26166

[0108]

number

[0109] TIFF0007849830000053.tif54166

[0110] Step S6.1.3: In flow rate condition Q1 (l=1,2,…,L), any two adjacent flow rate conditions Q α and Q β In contrast, using equation (14), the local difference quotient Diff relating to the area suitable for fish habitats within the river channel is used. WUA,α-β Calculate.

[0111]

number

[0112] During the ceremony, WUA prediction,α and WUA prediction,β These are the flow rate conditions Q, respectively. α and Q β At that time, the deep learning model for predicting the suitability of the fish habitat within the river channel outputs the area of ​​the suitable fish habitat within the river channel.

[0113] TIFF0007849830000055.tif96166

[0114]

number

[0115] TIFF0007849830000057.tif45166

[0116] Step S6.2.2: In flow rate condition Q1 (l=1,2,…,L), any two adjacent flow rate conditions Q α and Q βFor this, the local difference quotient Diff HSI,α-β regarding the suitability distribution of the fish habitat environment in the river channel is calculated using Equation (16).

[0117]

Number

[0118] TIFF0007849830000059.tif26166

[0119] TIFF0007849830000060.tif77166

[0120]

Number

[0121] In the formula, Interval sensitive is identified as the flow rate influence sensitivity range that significantly affects the suitability of the fish habitat environment in the river channel.

[0122] Step S7: Step of constructing the sample set D * for incremental learning of the fish habitat environment suitability: For the identified flow rate influence sensitivity range Interval sensitive =[Q * min, Q * max], according to the incremental learning interval △Q prediction select K * species of flow rate conditions, randomly sample K ** species of flow rate conditions from the K species of flow rate conditions selected in Step S2 or K prediction <△Q acture where K * species of flow rate conditions and K ** species of flow rate conditions are combined to form a set of flow rate conditions for incremental learning, Various flow rate conditions from the set of flow rate conditions for incremental learning are input into the physical mechanism model of fish habitat suitability, and simulations are performed to obtain fish habitat suitability evaluation results in the river channel corresponding to various flow rate conditions. These are then combined to create a sample set D for incremental learning of fish habitat suitability. * Constitute, Step S8: Sample set D for incremental learning of fish habitat suitability * The process involves using a first-stage trained deep learning model for predicting the suitability of fish habitats within river channels to perform a second-stage incremental learning to generate a retrained deep learning model for predicting the suitability of fish habitats within river channels. Step S9: A step in which the suitability of the fish habitat in the river channel under each flow rate condition in the target river channel area is predicted using a retrained deep learning model for predicting the suitability of the fish habitat in the river channel.

[0123] Specifically, when a user inputs river channel flow rate, a deep learning model for predicting the suitability of fish habitats within the river channel performs calculations based on that input, predicting the area of ​​suitable fish habitats within the river channel corresponding to that flow rate, as well as the fish habitat suitability values ​​within the river channel for each calculation cell grid. The model's output results can be directly used for decision-making related to river channel management and protection, and the rapid dissemination and application of information is possible, for example, by displaying prediction result maps via geographic information systems or other visualization methods.

[0124] The present invention also provides a system for realizing a rapid evaluation method for fish habitat suitability based on deep learning, the system comprising: a first flow rate condition determination unit; a primary learning unit for a physical mechanism model of fish habitat suitability; a unit for acquiring a sample set for learning fish habitat suitability; a deep learning model for predicting fish habitat suitability in a river channel; a second flow rate condition determination unit; a flow rate influence sensitivity range; a unit for acquiring a sample set for incremental learning of fish habitat suitability; a second-stage incremental learning unit; The first flow rate condition determination unit analyzes multi-year flow rate history data for the target river channel area to determine the range of flow rate conditions [Qmin, Qmax], where Q min and Q max These indicate the upper and lower limits of the flow rate condition range, and within the flow rate condition range [Qmin, Qmax], the flow rate interval △Q acture Accordingly, select flow rate conditions of type K, and these are set to flow rate condition Q k Expressed as (k=1,2,…,K), The primary learning unit of the physical mechanism model for fish habitat suitability is the learning sample set D={(Q k , Result acture,k This method uses )} to train a pre-built deep learning model for predicting the suitability of fish habitats within river channels, thereby generating a first-stage trained deep learning model for predicting the suitability of fish habitats within river channels. The unit for acquiring sample sets for learning fish habitat suitability evaluates the physical mechanism model results of fish habitat suitability within the river channel. acture,k Based on this, the sample set D for learning fish habitat suitability is {(Q k , Result acture,k This is for constructing )}, A deep learning model for predicting the suitability of fish habitats within river channels uses a sample set D={(Q)} for learning fish habitat suitability. k , Result acture,k This is a method for generating a pre-trained deep learning model for predicting the suitability of fish habitats within river channels by using )} for training. The second flow rate condition determination unit determines the flow rate interval △Q within the range of the flow rate conditions [Qmin, Qmax]. acture Smaller flow interval △Q * prediction Accordingly, the purpose is to select L-type flow rate conditions again at high density and express them as flow rate conditions Q1 (l=1,2,…,L), The flow rate influence sensitivity range discrimination unit performs simulations using a pre-trained deep learning model for predicting the suitability of fish habitats within river channels, and evaluates the deep learning model's assessment of fish habitat suitability within river channels corresponding to various flow rate conditions Q1. prediction,l The results obtained are an evaluation of a deep learning model regarding the suitability of fish habitats within river channels. prediction,l We analyzed the flow rate sensitivity range interval that shows a significant impact on the suitability of the fish habitat within the river channel. sensitive =[Q * min, Q * This is for identifying max, Q * min and Q * max These indicate the upper and lower limits of the flow rate influence sensitivity range, respectively. The sample acquisition unit for incremental learning of fish habitat suitability identifies the flow rate effect sensitivity range interval. sensitive =[Q * min, Q * In max, the incremental learning interval is △Q prediction According to K * Select the flow rate conditions for type K, and from the flow rate conditions for type K ** The flow rate conditions for each species were randomly sampled, and furthermore, △Q prediction <△Q acture And, K * Species flow rate conditions and K ** By combining the flow rate conditions of different species, a set of flow rate conditions for incremental learning is constructed. Various flow rate conditions from the set of flow rate conditions for incremental learning are input into the physical mechanism model of fish habitat suitability, and simulations are performed to obtain fish habitat suitability evaluation results in the river channel corresponding to various flow rate conditions. These are then combined to create a sample set D for incremental learning of fish habitat suitability. * It is intended to constitute, The second-stage incremental learning unit is sample set D for incremental learning of fish habitat suitability. *This method involves performing a second stage of incremental learning on a first-stage trained deep learning model for predicting the suitability of fish habitats within a river channel, generating a retrained deep learning model for predicting the suitability of fish habitats within a river channel, and then using this retrained deep learning model to predict the suitability of fish habitats within the river channel under various flow rate conditions in the target river channel area.

[0125] The method and system provided by the present invention for rapidly evaluating the suitability of fish habitats based on deep learning has the following features: (1) The present invention makes it possible to accurately calculate the distribution of suitable fish habitats within a river channel and the area of ​​suitable fish habitats within a river channel by comprehensively considering the distribution of hydrological environmental factors in the target river channel area, inter-school interactions and random perturbation terms, etc., and constructing a precise physical mechanism model of the suitability of fish habitats within the river channel. However, the time required for calculation is relatively long. (2) The present invention constructs a set of training samples corresponding to various flow rate conditions based on the distribution and area of ​​suitable fish habitats within a river channel, which are precisely calculated using a physical mechanism model of suitable fish habitats within a river channel. This set of training samples is then used to train a deep learning model for predicting suitable fish habitats within a river channel. Because the accuracy of the sample set is high, the accuracy of the deep learning model for predicting suitable fish habitats within a river channel can be improved. Furthermore, by performing incremental learning training on a deep learning model for predicting fish habitat suitability within a river channel through the identification of the flow rate influence sensitivity range, it is possible to obtain a deep learning model for predicting fish habitat suitability within a river channel that is close to the predictive accuracy of the physical mechanism model for fish habitat suitability within a river channel. By using this deep learning model for predicting fish habitat suitability within a river channel, it is possible to achieve highly accurate predictions of fish habitat suitability within a target river channel area, while also compensating for the problem that the time required to calculate the physical mechanism model for fish habitat suitability within a river channel is relatively long. The deep learning model for predicting fish habitat suitability within a river channel of the present invention is capable of efficiently and accurately predicting fish habitat suitability within a target river channel area.

[0126] Therefore, the present invention constructs a highly efficient deep learning model for predicting the suitability of fish habitats in river channels based on modules such as convolutional encoding and attention mechanisms. By autonomously learning based on data samples simulated by a physical mechanism model, it establishes a correspondence between the distribution of suitable fish habitats in river channels and factors such as flow rate, thereby achieving rapid and highly accurate evaluation of the distribution of suitable fish habitats in river channels. Compared to conventional prediction techniques, the present invention significantly reduces the time and cost burden and has high practicality and broad applicability in fields such as river ecosystem conservation and water resource management.

[0127] To better understand the technical means described above, exemplary embodiments of the present invention will be described in detail below with reference to the attached drawings. However, it should be understood that these are merely exemplary embodiments of the present invention, and the present invention is not limited to the embodiments described herein, but can be realized in various forms. These embodiments are provided to enable those skilled in the art to understand the present invention more clearly and fully.

[0128] 1. Target river channel area In this embodiment, a certain river channel section is used as an example of the target river channel area. The total length of the target river channel area is 2.7 km. This section has a complex topography, with many shallow areas and oxbow lakes. The bottom sediment consists of gravel, pebbles, and muddy sand, and there is no vegetation along the riverbanks. The target river channel area is suitable for the spawning of native fish species, can meet the conditions for embryonic development, and is an important fish habitat in the region.

[0129] 2. Physical Simulation Samples In order to provide sufficient training samples for a deep learning model for predicting the suitability of fish habitats within river channels, this embodiment uses a physical mechanism model of fish habitat suitability to perform hydraulic simulations and fish habitat suitability simulations.

[0130] The physical mechanism model for fish habitat suitability includes a two-dimensional hydraulic model, a model for the renewal of individual fish movements, a model for calculating the distribution density of fish schools, a model for calculating the suitability of fish habitats within river channels, and a model for calculating the area of ​​suitable fish habitats within river channels. The model for the renewal of individual fish movements, the model for calculating the distribution density of fish schools, the model for calculating the suitability of fish habitats within river channels, and the model for calculating the area of ​​suitable fish habitats within river channels constitute the fish habitat suitability model.

[0131] The finite volume method is employed to solve the model, and the procedure mainly includes initialization, flux calculation, time evolution, source term processing, and iterative solution. In this embodiment, the computational domain R is divided into an unstructured triangular grid, and the explicit Euler method is used for time discreteness, with a time step of 0.1 seconds and a total simulation time of 24 hours. The Roe approximation Riemann solution method is used for flux calculation, and the hydraulic field within the river channel is calculated through boundary condition definition and source term processing.

[0132] The fish habitat suitability model employs a behavioral simulation calculation method based on individual fish. This method constructs a function that describes the movement of individual fish during the breeding season and combines it with the calculation results of a two-dimensional hydraulic model to obtain the suitable distribution of flow velocity and water depth within the target river channel area. In the computational domain R, spawning grounds are identified based on the distribution after individual fish move freely according to the set behavior.

[0133] Specifically, the statistical results of hydrological data over multiple years are used as a method for selecting typical flow rate conditions. Frequency analysis is performed on flow rate data within long-term hydrological data, and flood peak flow rates are calculated for different return periods. The range of typical flow rate conditions for the calculation is then selected according to the requirements. In this example, the spawning season of fish was the main research target, and daily hydrological data for the past 50 years in the target river channel area of ​​the example were statistically analyzed. As a result, the flow rate at a cumulative frequency of 10% was 230 m³. 3 The rate is / s, and the maximum flow rate is 3320m 3 The value was / s. In this example, the flow rate condition range [Qmin, Qmax] was 200~3500m 3 Let / s be the flow interval ΔQ actureto 15 m 3 was determined to be / s, and 330 typical flow conditions were obtained, that is, K was 330.

[0134] In this example, the terrain data was obtained from laser ranging and field surveys using an ADCP. The flow rate data is derived from the data of hydrological observation stations within the river section. The distribution of hydraulic elements at the main cross-section was measured using an acoustic Doppler velocity profiler (ADCP). The number of cross-sections was 10, the measurement accuracy was ±0.01 meters, and the data units were meters and meters per second. The substrate data was determined by aerial surveys using an unmanned aircraft and on-site sampling. The fish preference, that is, the correlation coefficient, was comprehensively determined by combining fish ecological research and calibration of model parameters based on field data. This example targets the spawning period of fish including the growth, movement, and spawning processes of adult fish. As a result of calibrating the model parameters based on field data, the flow velocity range suitable for fish spawning is 0.15 m / s to 1.2 m / s, the minimum sensitive flow velocity is 0.1 m / s, the maximum swimming force is 1.5 m / s, the spawning preference flow velocity range is 0.4 to 0.9 m / s, the water depth range suitable for fish spawning is 0.2 m to 2.5 m, the spawning preference water depth range is 0.6 to 1.5 m, the threshold of the viable body length is 12 cm, and the spawning preference water temperature is 9 to 13 °C. After the calculation was completed, based on the two-dimensional distribution density p(x,y) of the fish population, the suitability distribution HSI(x,y) of the fish spawning habitat environment was estimated.

[0135] Furthermore, as a normalization process for the calculation results, the calculation results based on the terrain grid, substrate grid, and physical mechanism model of fish habitat suitability were normalized and interpolated to obtain normalized matrix data. In this example, the data was normalized by the linear interpolation method to obtain 256×256 matrix data.

[0136] Thereby, a learning sample set D = {(Q k , Result acture,k )} was constructed. Here, k = 1, 2, …, 330, and Result acture,k includes HSI acture,k,e and WUA acture,k . HSI acture,k,e is the flow condition Qk At this time, it is the fish habitat suitability value in the river channel at the computational cell grid e output by the physical mechanism model of fish habitat suitability. WUA acture,k This is the flow rate condition Q k At that time, this is the area of ​​the fish habitat suitable for fish within the river channel, as output by the physical mechanism model for fish habitat suitability.

[0137] 3. Construction and training of a deep learning model for predicting the suitability of fish habitats within river channels. The deep learning model for predicting the suitability of fish habitats within river channels in this embodiment was constructed based on the PyTorch deep learning framework, and data processing was performed using libraries such as Numpy, scipy, and Matplotlib.

[0138] In this embodiment, the deep learning model is a predictive model based on a convolutional neural network and an attention mechanism, primarily used to predict the suitability of fish habitats and the area of ​​fish habitats within river channels. The model's input includes two-dimensional topographic data, flow rate data, and environmental factor data. Multiple fully connected layers extract features from the flow rate data, expand this into a two-dimensional matrix, and then combine it with the topographic data before inputting it into the convolutional layer for processing. The model employs a three-layer convolutional and deconvolutional structure, with a convolutional layer kernel size of 3x3 and an activation function of ReLU. It uses a maximum pooling layer to sequentially downsample and extract deep features from the input data. To preserve more spatial detail, the model introduces skip connections between the convolutional and deconvolutional layers, connecting convolutional layer 1 to deconvolutional layer 2, and convolutional layer 2 to deconvolutional layer 1. The attention mechanism weights important regions in the feature map, ensuring that the model focuses on key features. During the decoding process, the original resolution of the feature map is sequentially restored through deconvolution operations, and finally, the fish habitat suitability value and fish habitat suitability area are output via two output layers.

[0139] Furthermore, the total loss function L total =αL habitat-RMSE +βL habitat-CE +γL area-logMAEThe loss value is calculated using this method.

[0140] Furthermore, the model's hyperparameters (including learning rate, batch size, and convolution kernel size) are optimized using Bayesian optimization, and the total loss function on the validation data set is analyzed. Minimize TIFF0007849830000062.tif16166. In this embodiment, the search space for the learning rate is 0.0005 to 0.001, and the search space for the batch size is 8 to 64.

[0141] 4. The sensitivity range of flow rate effects that significantly affect the suitability of fish habitats within river channels [Q * min, Q * Identification of max] In this embodiment, the identified flow rate effect sensitivity range [Q * min, Q * [max] is 400-900 m 3 It is / s.

[0142] 5. Incremental learning of the flow rate sensitivity range in a fish habitat suitability model within a river channel. Furthermore, the data set used during incremental learning training includes all newly added simulation results and sampled data extracted from the original data set. In this embodiment, 400-900m 3 5m within the range of / s 3 One flow rate condition was added for every / s, resulting in a total of 150 flow rate conditions. The data set used during incremental learning includes these 150 additional flow rate conditions and 150 flow rate conditions randomly selected from the original data set.

[0143] 6. Execute the prediction Furthermore, when a user inputs river channel flow rate, the model quickly performs calculations based on that input and predicts the fish habitat suitability value and fish habitat suitability area within the river channel corresponding to that input flow rate.

[0144] Furthermore, the output of the deep learning model for predicting the suitability of fish habitats within river channels can be directly used for decision-making regarding river channel management and protection. For example, displaying the predicted distribution map using a Geographic Information System (GIS) or other visualization methods enables rapid dissemination and application of information. In this embodiment, the results are output as 256 x 256 matrix data, and the distribution of fish habitat suitability is visualized using Matplotlib.

[0145] To more visually demonstrate the effectiveness of this embodiment, Figure 2 shows the matrix-formatted topographic data of the model input, and Figure 3 shows a comparison of the fish habitat suitability distribution and fish habitat suitability area predicted by the deep learning model for predicting fish habitat suitability within the river channel and the physical mechanism model for fish habitat suitability. Figure (a) shows the flow rate condition of 307.5 m 3 In / s, this is a distribution map of suitable fish habitats predicted by a deep learning model for predicting suitable fish habitats within a river channel. The predicted area of ​​suitable fish habitats is 134670.693 m². 2 In contrast, Figure (b) shows the same flow rate conditions at 307.5 m³. 3 In / s, the distribution map of suitable habitats for fish calculated by a physical mechanism model of fish habitat suitability is 139459.673 m². 2 That was the case.

[0146] Figure (c) shows the flow rate condition 802.5 m³. 3 In / s, this is a distribution map of suitable fish habitats predicted by a deep learning model for predicting suitable fish habitats within a river channel. The predicted area of ​​suitable fish habitats is 228355.362 m². 2 In contrast, Figure (d) shows the same flow rate condition of 802.5 m³. 3 In / s, the distribution map of suitable fish habitats was calculated by a physical mechanism model of fish habitat suitability, and the calculated area of ​​suitable fish habitats is 237727.377 m². 2 That was the case.

[0147] As is clear from Figure 3, the results predicted by the deep learning model for predicting the suitability of fish habitats in river channels according to the present invention are in close proximity to the results calculated by the physical model.

[0148] Figure 4 shows the relative error results of the fish habitat suitability area predicted by the deep learning model for predicting fish habitat suitability in river channels within the range of flow rate conditions used for model training. That is, under various flow rate conditions, if WUA1 is the fish habitat suitability area predicted by the deep learning model for predicting fish habitat suitability in river channels, and WUA2 is the fish habitat suitability area calculated by the physical mechanism model for fish habitat suitability, the relative error can be calculated using the formula (WUA1-WUA2) / WUA2, and the prediction accuracy of the deep learning model for predicting fish habitat suitability in river channels can be evaluated. As is clear from Figure 4, under various flow rate conditions, the relative prediction error of the deep learning model for predicting fish habitat suitability in river channels is small, confirming high prediction accuracy.

[0149] Therefore, the present invention constructs a deep learning model for predicting the suitability of fish habitats in river channels that is highly efficient, based on a deep learning algorithm. By learning from sample data obtained from simulations of the fish habitat suitability model using its physical mechanisms, the present invention establishes an intrinsic correspondence between the distribution of fish habitat suitability in river channels and factors such as flow rate. This makes it possible to quickly evaluate the distribution of fish habitat suitability in river channels under given flow rate conditions based on given flow rate data.

[0150] The foregoing describes only one example of a preferred embodiment of the present invention. Those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also covered within the scope of the present invention.

Claims

1. A rapid evaluation method for fish habitat suitability based on deep learning, Step S1: A step in which the flow rate history data over multiple years of the target river channel area is analyzed and the range of flow rate conditions [Qmin, Qmax] is determined, wherein Q ｍｉｎ and Q ｍａｘ These indicate the upper and lower limits of the flow rate conditions, respectively. Step S2: Within the range of the flow rate conditions [Qmin, Qmax], the flow rate interval △Q ａｃｔｕｒｅ Accordingly, select the flow rate conditions for type K, and these are Q ｋ Steps expressed as (k = 1, 2, ..., K), Step S3: Conduct simulations using a physical mechanism model of fish habitat suitability, and under various flow rate conditions Q ｋ Results of the evaluation of the physical mechanism model for the suitability of the fish habitat environment within the river channel. ａｃｔｕｒｅ，ｋ Steps to obtain Step S4: Evaluating the physical mechanism model of the fish habitat suitability in the river channel, and obtaining the evaluation result Result ａｃｔｕｒｅ，ｋ Based on this, constructing a learning sample set D = {(Q ｋ , Result ａｃｔｕｒｅ，ｋ )} Step S5: The sample set D for learning the suitability of the fish habitat to the environment = {(Q ｋ , Result ａｃｔｕｒｅ，ｋ The steps include: training a pre-built deep learning model for predicting the suitability of fish habitats in river channels using )} to generate a first-stage trained deep learning model for predicting the suitability of fish habitats in river channels; Step S6: Within the range of the flow rate conditions [Qmin, Qmax], the flow rate interval △Q ａｃｔｕｒｅ Smaller flow interval △Q ＊ ｐｒｅｄｉｃｔｉｏｎ Accordingly, select the flow rate conditions for type L again at high density, and these are Q １ Expressed as (l = 1, 2, ..., L), Simulations were conducted using the aforementioned pre-trained deep learning model for predicting the suitability of fish habitats within river channels, under various flow rate conditions Q. １ Evaluation results of a deep learning model regarding the suitability of fish habitats within river channels. ｐｒｅｄｉｃｔｉｏｎ，ｌ The results of the evaluation of the deep learning model regarding the suitability of the fish habitat environment within the river channel were obtained. ｐｒｅｄｉｃｔｉｏｎ，ｌ The analysis revealed the flow rate sensitivity range within the river channel that shows a significant impact on the suitability of the fish habitat. ｓｅｎｓｉｔｉｖｅ = [Q ＊ min, Q ＊ A step in identifying max, Q ＊ ｍｉｎ and Q ＊ ｍａｘ These indicate the upper and lower limits of the flow rate influence sensitivity range, respectively. Step S7: Sample set D for incremental learning of fish habitat suitability ＊ This is a step in building, Identified flow rate effect sensitivity range Interval ｓｅｎｓｉｔｉｖｅ = [Q ＊ min, Q ＊ In max, the incremental learning interval is △Q. ｐｒｅｄｉｃｔｉｏｎ According to K ＊ Select the flow rate conditions of type K, and from the flow rate conditions of type K selected in step S2, K ＊＊ The flow rate conditions for each species were randomly sampled, and furthermore, △Q ｐｒｅｄｉｃｔｉｏｎ < △Q ａｃｔｕｒｅ And, The aforementioned K ＊ The flow rate conditions of the species and the K ＊＊ By combining the flow rate conditions of different species, a set of flow rate conditions for incremental learning is constructed. The various flow rate conditions in the incremental learning flow rate condition set are input into the physical mechanism model of fish habitat suitability and simulations are performed to obtain fish habitat suitability evaluation results in the river channel corresponding to the various flow rate conditions, and these are combined to form the incremental learning sample set D of fish habitat suitability. ＊ Constitute, Step S8: Sample set D for incremental learning of fish habitat suitability ＊ The steps include: using the above-mentioned first-stage trained deep learning model for predicting the suitability of fish habitats in river channels, performing a second-stage incremental learning to generate a retrained deep learning model for predicting the suitability of fish habitats in river channels; and Step S9: Includes the step of predicting the suitability of the fish habitat in the river channel under each flow rate condition in the target river channel area using the retrained deep learning model for predicting the suitability of the fish habitat in the river channel. A method for rapidly evaluating the suitability of fish habitats based on deep learning, characterized by the features of this method.

2. In step S3, the physical mechanism model for fish habitat suitability includes a two-dimensional hydraulic model, a model for the renewal of individual fish movements, a two-dimensional distribution density calculation model for fish schools, a model for calculating fish habitat suitability within the river channel, and a model for calculating the area of ​​fish habitat suitability within the river channel. A method for rapidly evaluating the suitability of a fish habitat for its habitat based on deep learning, as described in claim 1.

3. Step S3 is, Step S3.1: Conduct a simulation using the two-dimensional hydraulic model, and under various flow rate conditions Q ｋ A step to obtain the distribution of hydrological environmental factors corresponding to the Step S3.2: Taking into account the distribution of the hydraulic environmental factors, the interactions between fish schools, and the random perturbation term, the position vector X of each individual fish i in the fish school at the next time step t+1 is calculated using the individual fish movement update model. ｔ＋１ ｉ,k The step of simulating the movement behavior of each individual fish i by simulating the following: Step S3.3: When the positional distribution of the fish school reaches a relatively stable state, the two-dimensional distribution density p(x,y) of the fish school is calculated using the two-dimensional distribution density calculation model for the fish school. Step S3.4: Based on the two-dimensional distribution density p(x,y) of the fish school, estimate the fish habitat suitability distribution HSI(x,y) within the river channel using the fish habitat suitability calculation model in equation (1). [Math 1] In the formula, R represents the computational domain, and (x, y) represents the coordinates of the horizontal and vertical axes of the computational domain R. Step S3.5: The fish habitat suitability distribution HSI(x,y) within the river channel is discretized into fish habitat suitability values ​​HSI within the river channel. ａｃｔｕｒｅ，ｋ，ｅ A step to convert, The computational domain R is discretized into E computational cell grids, and the HSI value for fish habitat suitability in the river channel is calculated for each computational cell grid e based on the HSI(x,y) distribution of fish habitat suitability in the river channel. ａｃｔｕｒｅ，ｋ，ｅ Identify (e = 1, 2, ..., E), Step S3.6: Using the model for calculating the suitable area of ​​fish habitat within the river channel according to equation (2), the various flow rate conditions Q ｋ WUA (Wet Usable Area) of Fish Habitat in River Channels ａｃｔｕｒｅ，ｋ Steps to estimate, [Math 2] In the formula, A represents the area of ​​each computational cell grid e. Step S3.7: Various flow rate conditions Q ｋ Results of the evaluation of the physical mechanism model for the suitability of the fish habitat environment within the river channel. ａｃｔｕｒｅ，ｋ = {HSI ａｃｔｕｒｅ，ｋ，ｅ , WUA ａｃｔｕｒｅ，ｋ Steps to obtain} Composed of A method for rapidly evaluating the suitability of a fish habitat for its environment based on deep learning, as described in claim 2.

4. A simulation was performed using the aforementioned two-dimensional hydraulic model, and the various flow rate conditions Q were used. ｋ In obtaining the distribution of the aforementioned hydraulic environmental factors corresponding to the target river channel area, basic data including a digital elevation model of the target river channel area and the distribution of the river channel roughness coefficient are required. The various flow rate conditions Q mentioned above ｋ The distribution of the aforementioned hydraulic environmental factors includes flow velocity, water depth, and river channel sediment distribution. A method for rapidly evaluating the suitability of a fish habitat for living environments based on deep learning, as described in claim 3.

5. Step S3.2 is, Step S3.2.1: The motion update model for the fish individual includes a speed update model for the fish individual and a position update model for the fish individual. The various flow rate conditions Q mentioned above ｋ In this simulation, the velocity update model for the fish individuals shown in equation (3) was used to obtain the velocity vector V of each fish individual i at time t. ｔ ｉ,k Calculation steps [Math 3] 【number】 [Math 4] 【number】 [Math 5] 【number】 【number】 [Math 6] In the formula, u(0, v ｐｒｅｆ ) is the swimming speed v based on the fish's preference for reverse currents, starting from 0. ｐｒｅｆ The velocity is randomly sampled within the range up to , and u(0, 2π) represents the value of the swimming direction randomly sampled within a 360° range. Step S3.2.2: Various flow rate conditions Q ｋ Below, a simulation was performed using the fish individual position update model shown in equation (7), and the position vector X of each fish individual i at the next time step t+1 was obtained. t+1 ｉ,k Steps to obtain [Number 7] In the formula, X t ｉ,k This shows the position vector of fish individual i at time t. A method for rapidly evaluating the suitability of a fish habitat for living environments based on deep learning, as described in claim 3.

6. Step S3.3: When the positional distribution of the fish school reaches a relatively stable state, the step of obtaining the two-dimensional distribution density p(x,y) of the fish school is performed, specifically, A method for rapidly evaluating the suitability of a fish habitat for fish based on deep learning, characterized in that the two-dimensional distribution density p(x,y) of the fish school is calculated using equation (8), as described in claim 3. [Number 8] In the formula, n is the total number of fish individuals in the school, σ is the smoothing parameter, (x, y) represents the coordinates of the horizontal and vertical axes of the computational domain R, and (x ｉ , y ｉ The horizontal and vertical coordinates of individual fish i when the positional distribution of the fish school reaches a relatively stable state represent the position vector of individual fish i. Composed of A method for rapidly evaluating the suitability of a fish habitat for living environments based on deep learning, as described in claim 3.

7. The deep learning model for predicting the suitability of fish habitats within a river channel comprises a data input and preprocessing module, a convolution encoding and attention mechanism module, a decoding and sampling module, and a habitat suitability prediction output module. The data input and preprocessing module receives topographic data and flow condition data of the target river channel area, performs preprocessing, the topographic data is two-dimensional grid data, the preprocessing step includes extracting high-dimensional features from the flow condition data using a multi-layer fully connected network, and expanding the flow condition data into a two-dimensional matrix matching the size of the topographic data, and then combining the processed topographic data, flow condition data and environmental data in the channel dimension to form integrated input data. The aforementioned convolutional encoding and attention mechanism module extracts features from the integrated input data via a convolutional neural network, obtains feature vectors, simultaneously applies weighting to the feature vectors using the attention mechanism to generate an attention feature map, and applies this to the output of the convolutional layer to enhance the representation of important features. The decoding and sampling module performs upsampling on the attention feature map processed by the attention mechanism using a deconvolutional network to obtain an upsampled feature map, which is then restored to the same resolution as the input terrain data. The habitat suitability prediction output module obtains evaluation results of the deep learning model regarding the suitability of the fish habitat within the river channel by processing the feature map after upsampling. The deep learning model for predicting the suitability of fish habitats within the river channel is configured to optimize the prediction accuracy of the deep learning model for predicting the suitability of fish habitats within the river channel using a custom loss function, and the total loss function is defined by equation (9). [Number 9] In the formula, α, β, and γ represent the weighting coefficients, L ｔｏｔａｌ L is the total loss function. ｈａｂｉｔａｔ－ＲＭＳＥ L is the mean square root error loss function. ｈａｂｉｔａｔ－ＣＥ L is the cross-entropy loss function. ａｒｅａ－ｌｏｇＭＡＥ This shows the logarithmic absolute error loss function. These are calculated using equations (10), (11), and (12), respectively. [Number 10] [Math 11] [Math 12] In the formula, HSI ｐｒｅｄｉｃｔｉｏｎ，ｋ，ｅ This is the flow rate condition Q ｋ At that time, the deep learning model for predicting the suitability of the fish habitat in the river channel outputs the fish habitat suitability value in the computational cell grid e, and HSI ａｃｔｕｒｅ，ｋ，ｅ Flow rate condition Q ｋ At that time, the value of the fish habitat suitability in the river channel in the computational cell grid e output by the physical mechanism model of fish habitat suitability is shown, H ａｃｔｕｒｅ，ｋ，ｅ Flow rate condition Q ｋ At that time, the HSI value of fish habitat suitability in the river channel in the computational cell grid e output by the physical mechanism model of fish habitat suitability is ａｃｔｕｒｅ，ｋ，ｅ The result of binarizing the data is shown below. 【number】 WUA ｐｒｅｄｉｃｔｉｏｎ，ｋ This is the flow rate condition Q ｋ At that time, the deep learning model for predicting the suitability of the fish habitat in the river channel outputs the area of ​​suitable fish habitat in the river channel, WUA ａｃｔｕｒｅ，ｋ Flow rate condition Q ｋ At that time, the area of ​​the fish habitat suitable for fish within the river channel, as output by the physical mechanism model for fish habitat suitability, is shown. A method for rapidly evaluating the suitability of a fish habitat for living environments based on deep learning, as described in claim 3.

8. The following method determines the flow rate influence sensitivity range [Q] that shows a significant impact on the suitability of the fish habitat within the river channel. ＊ min, Q ＊ Identify max], Step S6.1: Interval of the flow rate influence sensitivity range in which the suitable area of ​​fish habitat within the river channel changes. ＷＵＡ = [Q ＷＵＡ，ｌｏｇ Q ＷＵＡ，ｈｉｇｈ Identify ] and Q ＷＵＡ，ｌｏｇ and Q ＷＵＡ，ｈｉｇｈ These values ​​represent the upper and lower limits of the flow rate sensitivity range in which the area suitable for fish habitats within the river channel changes, respectively. Step S6.1.1: A simulation is performed using the deep learning model for predicting the suitability of the fish habitat in the river channel, which has been trained in the first stage, and the various flow rate conditions Q are used. １ Evaluation results of the deep learning model regarding the suitability of the fish habitat environment within the river channel, corresponding to the above. ｐｒｅｄｉｃｔｉｏｎ，ｌ The evaluation results include the WUA (Whole Usable Area) of the fish habitat within the river channel. ｐｒｅｄｉｃｔｉｏｎ，ｌ and the HSI (Hyper-Saturated Identification Value) of the fish habitat suitability in the river channel in each of the calculation cell grids e. ｐｒｅｄｉｃｔｉｏｎ，ｌ，ｅ It includes, 【number】 [Number 13] 【number】 Step S6.1.3: Flow rate condition Q １ For (l = 1, 2,..., L), for any two adjacent flow rate conditions Q α and Q β using Equation (14), calculate the local difference quotient Diff ＷＵＡ，α－β with respect to the suitable area of the fish habitat environment in the river channel, and [Number 14] Wherein, WUA ｐｒｅｄｉｃｔｉｏｎ，α and WUA ｐｒｅｄｉｃｔｉｏｎ，β respectively represent the suitability area of the fish habitat environment in the river channel output by the deep learning model for predicting the suitability of the fish habitat environment in the river channel when the flow conditions are α and Q β . 【number】 Step S6.2: Interval of the flow rate influencing the distribution of suitable fish habitats within the river channel. ＨＳＩ = [Q ＨＳＩ，ｌｏｇ Q ＨＳＩ，ｈｉｇｈ Identify ] and Q ＨＳＩ，ｌｏｇ and Q ＨＳＩ，ｈｉｇｈ These values ​​represent the upper and lower limits of the flow rate sensitivity range in which the distribution of suitable habitats for fish within the river channel changes, respectively. 【number】 [Number 15] 【number】 Step S6.2.2: Flow rate condition Q １ In (l = 1, 2, ..., L), any two adjacent flow rate conditions Q α and Q β For this, use equation (16) to obtain the local difference quotient Diff for the suitable distribution of fish habitats within the river channel. ＨＳＩ，α－β Calculate, [Number 16] In the formula, RHSI ｐｒｅｄｉｃｔｉｏｎ，α and RHSI ｐｒｅｄｉｃｔｉｏｎ，β These are the flow rate conditions Q, respectively. α and Q β At that time, the distribution area of ​​fish habitat suitability within the river channel is shown, which is statistically obtained based on the fish habitat suitability value within the river channel output by the deep learning model for predicting fish habitat suitability within the river channel. 【number】 Step S6.3: Using equation (17), determine the flow rate effect sensitivity range in which the suitable area of ​​fish habitat within the river channel obtained in Step S6.1.4 changes. ＷＵＡ = [Q ＷＵＡ，ｌｏｗ Q ＷＵＡ，ｈｉｇｈ ] and the flow rate influence sensitivity range Interval in which the distribution of suitable fish habitats within the river channel obtained in step S6.2.3 changes. ＨＳＩ = [Q ＨＳＩ，ｌｏｗ Q ＨＳＩ，ｈｉｇｈ The union of the above is taken and identified as the flow rate influence sensitivity range [Q*min, Q*max] that shows a significant impact on the suitability of the fish habitat environment within the river channel. [Number 17] In the formula, Interval ｓｅｎｓｉｔｉｖｅ This range was identified as a flow rate sensitivity range that shows a significant impact on the suitability of the fish habitat within the aforementioned river channel. A method for rapidly evaluating the suitability of a fish habitat for living environments based on deep learning, as described in claim 3.

9. A system for realizing a method for rapidly evaluating the suitability of fish habitats based on deep learning, according to any one of claims 1 to 8, The system comprises a first flow rate condition determination unit, a primary learning unit for a physical mechanism model of fish habitat suitability, a unit for acquiring a sample set for learning fish habitat suitability, a deep learning model for predicting fish habitat suitability within a river channel, a second flow rate condition determination unit, a flow rate influence sensitivity range, a unit for acquiring a sample set for incremental learning fish habitat suitability, another unit for acquiring a sample set for incremental learning fish habitat suitability, and a second-stage incremental learning unit. The first flow rate condition determination unit analyzes flow rate history data over multiple years for the target river channel area to determine the range of flow rate conditions [Qmin, Qmax], where Q ｍｉｎ and Q ｍａｘ These indicate the upper and lower limits of the flow rate condition range, and within the flow rate condition range [Qmin, Qmax], the flow rate interval △Q ａｃｔｕｒｅ Accordingly, select flow rate conditions of type K, and these are set to flow rate condition Q ｋ Expressed as (k = 1, 2, ..., K), The primary learning unit of the physical mechanism model for fish habitat suitability is defined as the learning sample set D = {(Q)} for fish habitat suitability. ｋ , Result ａｃｔｕｒｅ，ｋ This method uses ) to train a pre-built deep learning model for predicting the suitability of fish habitats within river channels, thereby generating a first-stage trained deep learning model for predicting the suitability of fish habitats within river channels. The aforementioned sample set acquisition unit for learning the suitability of fish habitats is the result of the evaluation of the physical mechanism model of the suitability of fish habitats within the river channel. ａｃｔｕｒｅ，ｋ Based on this, the sample set D for learning the suitability of the fish habitat to the environment is { (Q ｋ , Result ａｃｔｕｒｅ，ｋ This is for constructing )} The deep learning model for predicting the suitability of the fish habitat environment within the river channel is defined as follows: D = {(Q)} ｋ , Result ａｃｔｕｒｅ，ｋ This is for generating a deep learning model for predicting the suitability of fish habitats within river channels, which has been trained once, by performing training using the above-mentioned method. The second flow rate condition determination unit determines the flow rate interval △Q within the range of the flow rate condition [Qmin, Qmax]. ａｃｔｕｒｅ Smaller flow interval △Q ＊ ｐｒｅｄｉｃｔｉｏｎ Accordingly, select the L-type flow rate conditions again at high density, and these will be set to flow rate condition Q. １ This is for expressing (l = 1, 2, ..., L), The flow rate influence sensitivity range identification unit performs simulations using the previously trained deep learning model for predicting the suitability of fish habitats within the river channel, and various flow rate conditions Q １ Evaluation results of a deep learning model regarding the suitability of fish habitats within river channels. ｐｒｅｄｉｃｔｉｏｎ，ｌ The results of the evaluation of the deep learning model regarding the suitability of the fish habitat environment within the river channel were obtained. ｐｒｅｄｉｃｔｉｏｎ，ｌ The analysis revealed the flow rate sensitivity range within the river channel that shows a significant impact on the suitability of the fish habitat. ｓｅｎｓｉｔｉｖｅ = [Q ＊ min, Q ＊ This is for identifying max, Q ＊ ｍｉｎ and Q ＊ ｍａｘ These indicate the upper and lower limits of the flow rate influence sensitivity range, respectively. The sample acquisition unit for incremental learning of fish habitat suitability is identified as the Interval flow rate effect sensitivity range. ｓｅｎｓｉｔｉｖｅ = [Q ＊ min, Q ＊ In max, the incremental learning interval is △Q. ｐｒｅｄｉｃｔｉｏｎ According to K ＊ Select the flow rate conditions for type K, and from the flow rate conditions for type K ＊＊ The flow rate conditions for each species were randomly sampled, and furthermore, △Q ｐｒｅｄｉｃｔｉｏｎ < △Q ａｃｔｕｒｅ And, The aforementioned K ＊ The flow rate conditions of the species and the K ＊＊ By combining the flow rate conditions of different species, a set of flow rate conditions for incremental learning is constructed. Various flow rate conditions from the aforementioned set of flow rate conditions for incremental learning are input into the physical mechanism model of fish habitat suitability, and simulations are performed to obtain fish habitat suitability evaluation results in the river channel corresponding to various flow rate conditions. These are then combined to form a sample set D for incremental learning of fish habitat suitability. ＊ It is intended to constitute, The second-stage incremental learning unit is the sample set D for incremental learning of fish habitat suitability. ＊ This method involves performing a second stage of incremental learning on the first-stage trained deep learning model for predicting the suitability of fish habitats in river channels, generating a retrained deep learning model for predicting the suitability of fish habitats in river channels, and then using this retrained deep learning model to predict the suitability of fish habitats in river channels under various flow rate conditions in the target river channel area. A rapid evaluation system for fish habitat suitability based on deep learning, characterized by the following:

Images