A rapid evaluation method and system for fish habitat suitability based on deep learning

Abstract

The present invention provides a method and system for rapid evaluation of fish habitat suitability based on deep learning, which relates to the field of river ecological technology, including: using a physical mechanism model of fish habitat suitability to simulate and obtain the evaluation results of the physical mechanism model of river fish habitat suitability corresponding to each flow condition, and thereby establish a fish habitat suitability training sample set; using the sample set to train a deep learning model for predicting river fish habitat suitability; identifying flow-sensitive intervals and establishing a fish habitat suitability incremental learning training sample set, and performing incremental learning training on the deep learning model. The present invention establishes a mapping relationship between the distribution of fish habitat suitability in the river and factors such as flow rate through autonomous learning in the data samples simulated by the physical mechanism model, thereby achieving a rapid and accurate evaluation of the distribution of fish habitat suitability in the river. The present invention reduces the time and cost investment of traditional prediction technologies, and has significant practicality and broad application prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of water conservancy projects and ecological environment protection, and in particular to a method and system for rapid evaluation of fish habitat suitability based on deep learning. Background Art

[0002] In the field of ecological and environmental science and water resources management, calculating the suitability of fish habitats under a large number of different downstream flows and different annual water flow conditions is an important basic work in the work of comparing and selecting watershed hydropower ecological dispatching and operation plans, and studying fish habitat suitability restoration and management plans.

[0003] Currently, commonly used methods for evaluating fish habitat adaptability rely on constructing hydrodynamic models to calculate the suitability of fish habitats under various operating conditions. This computational process is time-consuming and costly, making it difficult to rapidly compute the massive amounts of discrete data required for detailed simulations of hydropower ecological scheduling schemes in river basins. Relevant scholars have proposed methods for fitting habitat suitability curves, such as using hydrological statistical indicators such as cumulative flow frequency as a reference to simulate and predict habitat suitability under different flow rates. However, these methods have limited accuracy, making it difficult to accurately assess the distribution of fish habitat suitability within a river channel. Summary of the Invention

[0004] In response to the defects of the existing technology, the present invention provides a method and system for rapid evaluation of fish habitat suitability based on deep learning, which can effectively solve the above problems.

[0005] The technical solution adopted in the present invention is as follows:

[0006] The present invention provides a method for rapid evaluation of fish habitat suitability based on deep learning, comprising the following steps:

[0007] Step S1: Analyze the historical flow data of the target river area for many years and determine the flow condition range [Q min ,Q max ]; among them, Q min and Q max , are the lower and upper bounds of the flow rate operating range respectively;

[0008] Step S2, in the flow operating range [Q min ,Q max ], according to the flow interval ΔQ acture , K flow conditions are selected and expressed as flow condition Q k , k=1,2,...,K;

[0009] Step S3: Using the physical mechanism model of fish habitat suitability, simulate each flow condition Q k Corresponding evaluation results of the physical mechanism model of river fish habitat suitabilityacture,k ;

[0010] Step S4: Evaluation results of the physical mechanism model of the river fish habitat suitability acture,k , construct the fish habitat suitability training sample set D = {(Q k ,Result acture,k )};

[0011] Step S5, using the fish habitat suitability training sample set D = {(Q k ,Result acture,k )}, training the pre-established river fish habitat suitability prediction deep learning model to obtain a river fish habitat suitability prediction deep learning model that has been trained once;

[0012] Step S6, in the flow operating range [Q min ,Q max ], according to the flow interval ΔQ acture Flow interval Re-encrypt and select to obtain L flow conditions, expressed as flow condition Q l , l=1,2,...,L;

[0013] The deep learning model for predicting river fish habitat suitability, which was trained once, was used to simulate the Q of each flow condition. l Corresponding evaluation results of the deep learning model for river fish habitat suitability prediction,l Result of deep learning model evaluation on the suitability of river fish habitat prediction,l Analyze and identify flow-sensitive intervals that are sensitive to the suitability of river fish habitats in, and are the lower and upper bounds of the flow-sensitive interval respectively;

[0014] Step S7: Establishing fish habitat suitability incremental learning training sample set D * :

[0015] In the identified traffic sensitive area In the incremental learning interval ΔQ prediction , select K * K flow conditions are randomly sampled from the K flow conditions selected in step S2. ** flow conditions; among them, ΔQ prediction <ΔQ acture ;

[0016] K * Flow conditions and K **The flow condition combinations are used to form an incremental learning flow condition set;

[0017] Each flow condition in the incremental learning flow condition set is input into the fish habitat suitability physical mechanism model, and the river fish habitat suitability evaluation results corresponding to each flow condition are simulated and then combined to form the fish habitat suitability incremental learning training sample set D. * ;

[0018] Step S8, using the fish habitat suitability incremental learning training sample set D * , performing a second incremental learning training on the deep learning model for predicting the suitability of river fish habitats that has been trained once, to obtain a trained deep learning model for predicting the suitability of river fish habitats;

[0019] Step S9: Use the trained deep learning model for predicting river fish habitat suitability to predict the river fish habitat suitability of the target river area under corresponding flow conditions.

[0020] Preferably, in step S3, the fish habitat suitability physical mechanism model includes a two-dimensional hydrodynamic model, a fish individual movement update model, a fish school two-dimensional distribution density calculation model, a river fish habitat suitability calculation model, and a river fish habitat suitability area calculation model.

[0021] Preferably, step S3 is specifically as follows:

[0022] Step S3.1: Use a two-dimensional hydrodynamic model to simulate each flow condition Q k Corresponding distribution of hydrodynamic environmental factors;

[0023] Step S3.2, taking into account the distribution of the hydrodynamic environmental factors, the interaction between fish schools and the random disturbance term, the fish individual motion update model is adopted to simulate the position vector of each fish individual i in the fish school at the next time t+1 Thus, the movement behavior of each fish individual i is simulated;

[0024] Step S3.3, when the fish school position distribution reaches a relatively stable state, the fish school two-dimensional distribution density calculation model is used to obtain the fish school two-dimensional distribution density ρ(x, y);

[0025] In step S3.4, based on the two-dimensional distribution density of fish schools ρ(x,y), the river fish habitat suitability calculation model of formula (1) is used to estimate the river fish habitat suitability distribution HSI(x,y):

[0026]

[0027] Where: R represents the calculation domain range; (x, y) represents the horizontal and vertical coordinates of the calculation domain range R;

[0028] Step S3.5, convert the river fish habitat suitability distribution HSI(x,y) into discrete river fish habitat suitability value HSI acture,k,e :

[0029] Discretize the computational domain R into E computational unit grids; identify the river fish habitat suitability value HSI in each computational unit grid e according to the river fish habitat suitability distribution HSI(x,y) acture,k,e ;e=1,2,...,E;

[0030] Step S3.6: Use the river fish habitat suitability area calculation model of formula (2) to estimate the flow condition Q k Corresponding river fish habitat suitability area WUA acture,k :

[0031]

[0032] Where: A is the area of ​​each computational unit grid e;

[0033] Step S3.7, thus obtaining each flow condition Q k Corresponding evaluation results of the physical mechanism model of river fish habitat suitability acture,k ={HSI acture,k,e ,WUA acture,k}.

[0034] Preferably, a two-dimensional hydrodynamic model is used to simulate each flow condition Q k When the corresponding hydrodynamic environmental factors are distributed, the required basic data include the digital elevation model of the target river area and the river roughness distribution;

[0035] Each flow condition Q k The corresponding distribution of hydrodynamic environmental factors includes flow velocity, water depth and riverbed distribution.

[0036] Preferably, step S3.2 is specifically as follows:

[0037] Step S3.2.1, the fish individual motion update model includes a fish individual velocity update model and a fish individual position update model;

[0038] In each flow condition Q k , using the fish individual speed update model shown in formula (3), the speed vector of each fish individual i at time t is simulated

[0039]

[0040] in: is the velocity vector of fish individual i at time t-1; is the change component of the velocity vector of fish individual i at time t-1 caused by the environment; is the change component of the velocity vector of fish individual i at time t-1 caused by the interaction between fish schools; is the velocity vector change component of fish individual i at time t-1 caused by the random disturbance term; Δt is the time interval;

[0041] Calculated by formula (4):

[0042]

[0043] Where: is the ambient velocity drift term; τ is the drift sensitivity coefficient; is the position vector of fish individual i at time t-1; is the environmental flow velocity vector at the location of fish individual i at time t-1;

[0044] is the swimming speed change term caused by the fish's rheotaxis; χ is the rheotaxis sensitivity coefficient; v pref Preferred swimming speed for fish against current; is the modulus of the ambient flow velocity vector at the location of fish individual i at time t-1; is the unit vector in the direction of the adverse ambient flow velocity; is the velocity vector of fish individual i at time t-1;

[0045] is the change in the swimming speed of fish caused by the individual fish's preference for environmental factors; γ is the overall weight coefficient of environmental factor preference; M is the number of categories of hydrodynamic environmental factors; is the value of the hydrodynamic environment factor m at the location of fish individual i at time t-1, m = 1, 2, ..., M; is a function for calculating the suitability of the hydrodynamic environmental factor m at the location of fish individual i at time t-1; ω m is the weight coefficient of the hydrodynamic environment factor m;

[0046] is the comprehensive environmental factor suitability gradient at the location of fish individual i at time t-1, that is, the swimming preference direction of fish individual i at time t-1;

[0047] Calculated by formula (5):

[0048]

[0049] Where: A>0; B>0; q>p; n is the total number of fish individuals in the school;

[0050] represents the distance vector from fish individual j to fish individual i at time t-1; j = 1, 2, ..., n, i = 1, 2, ..., n, and j ≠ i; The position vector of fish individual j at time t-1; represents the modulus of the distance vector from fish individual j to fish individual i at time t-1; represents the long-distance attraction term between fish individual j and fish individual i at time t-1; represents the short-range repulsion term between fish individual j and fish individual i at time t-1; A is the attraction intensity coefficient; B is the repulsion intensity coefficient;

[0051] The power exponents p and q control the attraction and repulsion attenuation distances respectively. A smaller p and a larger q indicate that the attraction effect is smaller and the repulsion effect is larger at close distances. The attraction effect decays more slowly with distance than the repulsion effect. is the unit vector between fish individual j and fish individual i, describing the direction of attraction and repulsion;

[0052] Calculated by formula (6):

[0053]

[0054] Where: u(0,v pref ) is the fish’s preferred swimming speed v for random speeds between 0 and pref Random sampling between u(0,2π) and u(0,2π) is random sampling of the swimming direction within 360°;

[0055] Step S3.2.2, in each flow condition Q k , using the fish individual position update model shown in formula (7), the position vector of each fish individual i at the next time t+1 is simulated

[0056]

[0057] in: is the position vector of fish individual i at time t.

[0058] Preferably, in step S3.3, when the fish position distribution reaches a relatively stable state, the two-dimensional distribution density ρ(x, y) of the fish is obtained, specifically:

[0059] Using formula (8), we can get the two-dimensional distribution density of fish school ρ(x,y):

[0060]

[0061] Where: n is the total number of fish individuals in the fish school; σ is the smoothing parameter; (x, y) represents the horizontal and vertical coordinates of the calculation domain range R; (x i ,y i ) represents the horizontal and vertical coordinates of fish individual i when the position distribution of the fish school reaches a relatively stable state, and is the position vector of fish individual i.

[0062] Preferably, the deep learning model for predicting river fish habitat suitability includes a data input and preprocessing module, a convolutional coding and attention mechanism module, a decoding and sampling module, and a habitat suitability prediction output module:

[0063] Data input and preprocessing module: Receives and preprocesses terrain data and flow condition data of the target river area; wherein the terrain data is two-dimensional raster data; the preprocessing step includes extracting high-dimensional features from the flow condition data through a multi-layer fully connected network, expanding the flow condition data into a two-dimensional matrix that matches the size of the terrain data; and splicing the processed terrain data, flow condition data, and environmental data in the channel dimension to form fused input data;

[0064] Convolutional coding and attention mechanism module: This module extracts features from the fused input data using a convolutional neural network to obtain a feature vector. It also performs weighted processing on the feature vector using an attention mechanism to generate an attention feature map, which is applied to the output of the convolutional layer to enhance the representation of important features.

[0065] Decoding and sampling module: The attention feature map processed by the attention mechanism is upsampled through the deconvolution network to obtain the upsampled feature map and restore it to the same resolution as the input terrain data:

[0066] Habitat suitability prediction output module: By processing the upsampled feature map, the deep learning model evaluation results of river fish habitat suitability are obtained;

[0067] Wherein: the deep learning model for predicting the suitability of river fish habitats uses a custom loss function to optimize the accuracy of the prediction of the deep learning model for predicting the suitability of river fish habitats; the total loss function is defined as shown in formula (9):

[0068] L total =αL habitat-RMSE +βL habitat-CE +γL area-logMAE (9)

[0069] Where: α, β and γ are weight coefficients respectively;

[0070] L total is the total loss function; L habitat-RMSE is the root mean square error loss function, L habitat-CE is the cross entropy loss function, L area-logMAE is the logarithmic absolute error loss function, which is calculated by formula (10), formula (11) and formula (12) respectively:

[0071]

[0072] L area-logMAE =log(WUA prediction,k -WUA acture,k ) (12)

[0073] Of which: HSI prediction,k,e is the flow condition Q k When the deep learning model for predicting the suitability of river fish habitats outputs the river fish habitat suitability value in the calculation unit grid e; HSI acture,k,e is the flow condition Q k When the fish habitat suitability physical mechanism model outputs the river fish habitat suitability value in the calculation unit grid e; H acture,k,e is the flow condition Q k When the fish habitat suitability physical mechanism model outputs the river fish habitat suitability value HSI in the calculation unit grid e acture,k,e The binarization result of

[0074] WUA prediction,k is the flow condition Q k When the deep learning model for predicting river fish habitat suitability outputs the river fish habitat suitability area; WUA acture,k is the flow condition Q k When the fish habitat suitability physical mechanism model is used, the river fish habitat suitability area is output.

[0075] Preferably, the following method is used to identify the flow sensitive intervals that are sensitive to the suitability of river fish habitats:

[0076] Step S6.1: Identify the flow-sensitive intervals where the suitable area of ​​fish habitat changes in the river. WUA =[Q WUA,low ,Q WUA,high ]; Q WUA,low and Q WUA,high , are the lower and upper bounds of the flow-sensitive interval for changes in the area of ​​suitable river fish habitats, respectively;

[0077] Step S6.1.1: Use the deep learning model for predicting river fish habitat suitability that has been trained once to simulate each flow condition Q l Corresponding evaluation results of the deep learning model for river fish habitat suitability prediction,l , including the river fish habitat suitability area (WUA) prediction,l And the fish habitat suitability value HSI in each calculation unit grid e prediction,l,e ;

[0078] In step S6.1.2, use formula (13) to calculate the global average difference quotient of the suitable area of ​​river fish habitat

[0079]

[0080] Where: ε is the flow condition, and Q ε , respectively, flow conditions and flow condition ε;

[0081] and WUA prediction,ε , respectively, flow conditions and flow condition ε, the river fish habitat suitability area output by the river fish habitat suitability prediction deep learning model;

[0082] Step S6.1.3, in flow condition Q l , l=1,2,...,L, for every two adjacent flow conditions Q α and Q β , formula (14) is used to calculate the local difference quotient Diff of the suitable area of ​​river fish habitat WUA,α-β :

[0083]

[0084] Among them: WUA prediction,α and WUA prediction,β , respectively, flow condition Q α and Q β When , the river fish habitat suitability area output by the river fish habitat suitability prediction deep learning model;

[0085] Step S6.1.4, compare the local difference quotient Diff of the fish habitat suitability area of ​​each river channel in turn WUA,α-β and the global mean difference quotient of the suitable area of ​​river fish habitat The size relationship of the local difference quotient Diff WUA,α-β Greater than the global mean difference quotient The concentrated flow area is the flow sensitive interval of the change of the suitable area of ​​river fish habitat WUA =[Q WUA,low ,Q WUA,high ]; Q WUA,low and Q WUA,high , are the lower and upper bounds of the flow-sensitive interval respectively;

[0086] Step S6.2: Identify the flow-sensitive intervals where the distribution of fish habitat suitability changes in the river. HSI =[Q HSI,low ,Q HSI,high ]; Q HSI,low and Q HSI,high , are the lower and upper bounds of the flow-sensitive interval for changes in the distribution of fish habitat suitability in rivers;

[0087] In step S6.2.1, use formula (15) to calculate the global mean difference quotient of the distribution of river fish habitat suitability.

[0088]

[0089] Among them: RHSI prediction,ε and RHSI prediction,ε+ΔQp*rediction , respectively, flow condition ε and flow condition When the river fish habitat suitability value HSI in each calculation unit grid e is output according to the river fish habitat suitability prediction deep learning model, prediction,l,e , the statistically obtained suitability value HSI prediction,l,e The area of ​​the calculation unit grid that is greater than 0 is referred to as the suitable distribution area of ​​river fish habitat;

[0090] Step S6.2.2, in flow condition Q l , l=1,2,...,L, for every two adjacent flow conditions Q α and Q β , using formula (16) to calculate the local difference quotient Diff of the distribution of river fish habitat suitability HSI,α-β :

[0091]

[0092] Among them: RHSI prediction,α and RHSI prediction,β , respectively, flow condition Q α and Q β , the distribution area of ​​river fish habitat suitability obtained based on the river fish habitat suitability value output by the river fish habitat suitability prediction deep learning model;

[0093] Step S6.2.3, compare the local difference quotient Diff of the fish habitat suitability distribution in each river channel in turn HSI,α-β and the global mean difference quotient of the distribution of fish habitat suitability in the river The size relationship of the local difference quotient Diff HSI,α-β Greater than the global mean difference quotient The concentrated flow area is the flow sensitive interval where the distribution of fish habitat suitability changes in the river. HSI =[Q HSI,low ,Q HSI,high ]; Q HSI,low and Q HSI,high , are the lower and upper bounds of the flow-sensitive interval for changes in the distribution of fish habitat suitability in rivers;

[0094] Step S6.3, using formula (17), the flow sensitive interval Interval of the change in the suitable area of ​​river fish habitat obtained in step S6.1.4 is WUA =[Q WUA,low ,Q WUA,high ] and the flow sensitive interval Interval of the change in the distribution of fish habitat suitability in the river obtained in step S6.2.3 HSI =[Q HSI,low ,Q HSI,high ] and take the union of the two sets to identify the flow sensitive interval that is sensitive to the suitability of river fish habitats.

[0095]

[0096] Where: Interval sensitive The flow-sensitive intervals that are sensitive to the suitability of river fish habitats are identified.

[0097] The present invention also provides a system for implementing the method for rapid evaluation of fish habitat suitability based on deep learning, comprising:

[0098] The first flow condition determination unit is used to analyze the historical flow data of the target river area for many years and determine the flow condition range [Q min ,Q max ]; among them, Q min and Q max , respectively the lower and upper bounds of the flow operating range; in the flow operating range [Q min ,Q max ], according to the flow interval ΔQ acture , K flow conditions are selected and expressed as flow condition Q k , k=1,2,...,K;

[0099] The fish habitat suitability physical mechanism model one-time training unit is used to use the fish habitat suitability training sample set D = {(Q k ,Result acture,k )}, training the pre-established river fish habitat suitability prediction deep learning model to obtain a river fish habitat suitability prediction deep learning model that has been trained once;

[0100] The fish habitat suitability training sample set acquisition unit is used to evaluate the results of the physical mechanism model of the river fish habitat suitability. acture,k , construct the fish habitat suitability training sample set D = {(Q k ,Result acture,k )};

[0101] A deep learning model for predicting river fish habitat suitability is used to train a fish habitat suitability sample set D = {(Q k ,Result acture,k )} to train and obtain a deep learning model for predicting river fish habitat suitability after one training;

[0102] The second flow condition determination unit is used to determine the flow condition range [Q min ,Q max ], according to the flow interval ΔQ acture Flow interval Re-encrypt and select to obtain L flow conditions, expressed as flow condition Q l , l=1,2,...,L;

[0103] The flow-sensitive interval identification unit is used to simulate the Q of each flow condition using a deep learning model for predicting the suitability of river fish habitats after a single training. l Corresponding evaluation results of the deep learning model for river fish habitat suitability prediction,l Result of deep learning model evaluation on the suitability of river fish habitat prediction,l Analyze and identify flow-sensitive intervals that are sensitive to the suitability of river fish habitats in, and are the lower and upper bounds of the flow-sensitive interval respectively;

[0104] The fish habitat suitability incremental learning training sample set acquisition unit is used to identify the flow sensitive area In the incremental learning interval ΔQ prediction , select K * K flow conditions are randomly sampled from K flow conditions.** flow conditions; among them, ΔQ prediction <ΔQ acture ;

[0105] K * Flow conditions and K ** The flow conditions are combined to form an incremental learning flow condition set; each flow condition in the incremental learning flow condition set is input into the fish habitat suitability physical mechanism model, and the river fish habitat suitability evaluation results corresponding to each flow condition are simulated to form the fish habitat suitability incremental learning training sample set D. * ;

[0106] Secondary incremental learning training unit, used to incrementally learn the training sample set D using fish habitat suitability * , a second incremental learning training is performed on the deep learning model for predicting the suitability of river fish habitats that has been trained once to obtain a trained deep learning model for predicting the suitability of river fish habitats; the trained deep learning model for predicting the suitability of river fish habitats in the target river area under the corresponding flow conditions is used to predict the suitability of river fish habitats.

[0107] The present invention provides a method and system for rapid evaluation of fish habitat suitability based on deep learning, which has the following advantages:

[0108] This paper constructs an efficient deep learning model for predicting river fish habitat suitability based on modules such as convolutional coding and an attention mechanism. By autonomously learning from data samples simulated by a physical mechanism model, it establishes a mapping relationship between the distribution of fish habitat suitability within a river and factors such as flow rate, enabling rapid and accurate assessment of the distribution of fish habitat suitability within a river. This method reduces the time and cost investment of traditional prediction technologies and is applicable to fields such as river ecological protection and water resource management, demonstrating significant practicality and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0109] Figure 1 A flowchart of a method for rapid evaluation of fish habitat suitability based on deep learning provided by the present invention;

[0110] Figure 2 An example diagram of matrixed terrain data provided by an embodiment of the present invention;

[0111] Figure 3 A comparison chart of the distribution of fish habitat suitability and the area of ​​fish habitat suitability predicted by the deep learning model for predicting river fish habitat suitability and the physical mechanism model for fish habitat suitability provided by an embodiment of the present invention;

[0112] Figure 4This is a graph of the relative error results of the fish habitat suitability area predicted by the deep learning model for predicting river fish habitat suitability provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0113] In order to make the technical problems, technical solutions and beneficial effects solved by the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0114] The present invention simulates the physical mechanism of habitat suitability based on intelligent fish individuals and creates a sample dataset of habitat suitability distribution of river fish; constructs a deep learning model for predicting fish habitat suitability based on a deep learning mechanism; and uses an incremental learning method to train and test the deep learning model, thereby reducing the calculation time of the physical model and further improving the performance of the deep learning model in fitting fish habitat suitability.

[0115] See Figure 1 The present invention provides a method for rapid evaluation of fish habitat suitability based on deep learning, comprising the following steps:

[0116] Step S1: Analyze the historical flow data of the target river area for many years and determine the flow condition range [Q min ,Q max ]; among them, Q min and Q max , are the lower and upper bounds of the flow rate operating range respectively;

[0117] Step S2, in the flow operating range [Q min ,Q max ], according to the flow interval ΔQ acture , K flow conditions are selected and expressed as flow condition Q k , k=1,2,...,K;

[0118] Step S3: Using the physical mechanism model of fish habitat suitability, simulate each flow condition Q k Corresponding evaluation results of the physical mechanism model of river fish habitat suitability acture,k ;

[0119] The physical mechanism model of fish habitat suitability is an innovative model of the present invention, which includes a two-dimensional hydrodynamic model, a fish individual movement and renewal model, a two-dimensional fish distribution density calculation model, a river fish habitat suitability calculation model, and a river fish habitat suitability area calculation model.

[0120] Step S3 is specifically as follows:

[0121] Step S3.1: Use a two-dimensional hydrodynamic model to simulate each flow condition Q k Corresponding distribution of hydrodynamic environmental factors;

[0122] In this step, a two-dimensional hydrodynamic model is used to simulate each flow condition Q k When the corresponding hydrodynamic environmental factors are distributed, the required basic data include the digital elevation model of the target river area and the river roughness distribution;

[0123] Each flow condition Q k The corresponding distribution of hydrodynamic environmental factors includes flow velocity, water depth and riverbed distribution.

[0124] Step S3.2, taking into account the distribution of the hydrodynamic environmental factors, the interaction between fish schools and the random disturbance term, the fish individual motion update model is adopted to simulate the position vector of each fish individual i in the fish school at the next time t+1 Thus, the movement behavior of each fish individual i is simulated;

[0125] Step S3.2 is specifically as follows:

[0126] Step S3.2.1, the fish individual motion update model includes a fish individual velocity update model and a fish individual position update model;

[0127] In each flow condition Q k , using the fish individual speed update model shown in formula (3), the speed vector of each fish individual i at time t is simulated

[0128]

[0129] in: is the velocity vector of fish individual i at time t-1; is the change component of the velocity vector of fish individual i at time t-1 caused by the environment; is the change component of the velocity vector of fish individual i at time t-1 caused by the interaction between fish schools; is the velocity vector change component of fish individual i at time t-1 caused by the random disturbance term; Δt is the time interval;

[0130] Calculated by formula (4):

[0131]

[0132] Where: is the ambient velocity drift term; τ is the drift sensitivity coefficient; is the position vector of fish individual i at time t-1; is the environmental flow velocity vector at the location of fish individual i at time t-1;

[0133] is the swimming speed change term caused by the fish's rheotaxis; χ is the rheotaxis sensitivity coefficient; v pref Preferred swimming speed for fish against current; is the modulus of the ambient flow velocity vector at the location of fish individual i at time t-1; is the unit vector in the direction of the adverse ambient flow velocity; is the velocity vector of fish individual i at time t-1;

[0134] is the change in the swimming speed of fish caused by the individual fish's preference for environmental factors; γ is the overall weight coefficient of environmental factor preference; M is the number of categories of hydrodynamic environmental factors, for example, including flow velocity, water depth and riverbed distribution; is the value of the hydrodynamic environment factor m at the location of fish individual i at time t-1, m = 1, 2, ..., M; is a function for calculating the suitability of the hydrodynamic environmental factor m at the location of fish individual i at time t-1. During the calculation, the velocity suitability curve, water depth suitability curve, and riverbed suitability curve can be used to convert the flow velocity, water depth, and riverbed suitability distribution into the corresponding flow velocity suitability distribution, water depth suitability distribution, and riverbed suitability distribution, respectively; ω m is the weight coefficient of the hydrodynamic environment factor m;

[0135] is the comprehensive environmental factor suitability gradient at the location of fish individual i at time t-1, that is, the swimming preference direction of fish individual i at time t-1;

[0136] Calculated by formula (5):

[0137]

[0138] Where: A>0; B>0; q>p; n is the total number of fish individuals in the school;

[0139] represents the distance vector from fish individual j to fish individual i at time t-1; j = 1, 2, ..., n, i = 1, 2, ..., n, and j ≠ i; The position vector of fish individual j at time t-1; represents the modulus of the distance vector from fish individual j to fish individual i at time t-1; represents the long-distance attraction term between fish individual j and fish individual i at time t-1; represents the short-range repulsion term between fish individual j and fish individual i at time t-1; A is the attraction intensity coefficient; B is the repulsion intensity coefficient;

[0140] The power exponents p and q control the attraction and repulsion attenuation distances respectively. A smaller p and a larger q indicate that the attraction effect is smaller and the repulsion effect is larger at close distances. The attraction effect decays more slowly with distance than the repulsion effect. is the unit vector between fish individual j and fish individual i, describing the direction of attraction and repulsion;

[0141] Calculated by formula (6):

[0142]

[0143] Where: u(0,v pref ) is the fish’s preferred swimming speed v for random speeds between 0 and pref Random sampling between u(0,2π) and u(0,2π) is random sampling of the swimming direction within 360°;

[0144] Step S3.2.2, in each flow condition Q k , using the fish individual position update model shown in formula (7), the position vector of each fish individual i at the next time t+1 is simulated

[0145]

[0146] in: is the position vector of fish individual i at time t.

[0147] Step S3.3, when the fish school position distribution reaches a relatively stable state, the fish school two-dimensional distribution density calculation model is used to obtain the fish school two-dimensional distribution density ρ(x, y);

[0148] Using formula (8), we can get the two-dimensional distribution density of fish school ρ(x,y):

[0149]

[0150] Where: n is the total number of fish individuals in the fish school; σ is the smoothing parameter; (x, y) represents the horizontal and vertical coordinates of the calculation domain range R; (x i ,y i ) represents the horizontal and vertical coordinates of fish individual i when the position distribution of the fish school reaches a relatively stable state, and is the position vector of fish individual i.

[0151] In step S3.4, based on the two-dimensional distribution density of fish schools ρ(x,y), the river fish habitat suitability calculation model of formula (1) is used to estimate the river fish habitat suitability distribution HSI(x,y):

[0152]

[0153] Wherein: R represents the calculation domain range; (x, y) represents the horizontal and vertical coordinates of the calculation domain range R; in the present invention, the normalized two-dimensional distribution density of fish schools is used as the distribution of river fish habitat suitability.

[0154] Step S3.5, convert the continuous river fish habitat suitability distribution HSI(x,y) into discrete river fish habitat suitability value HSI acture,k,e :

[0155] Specifically, the computational domain R is discretized into E computational unit grids; according to the river fish habitat suitability distribution HSI (x, y), the river fish habitat suitability value HSI in each computational unit grid e is identified. acture,k,e ;e=1,2,...,E;

[0156] Step S3.6: Use the river fish habitat suitability area calculation model of formula (2) to estimate the flow condition Q k Corresponding river fish habitat suitability area WUA acture,k :

[0157]

[0158] Where: A is the area of ​​each computational unit grid e;

[0159] Step S3.7, thus obtaining each flow condition Q k Corresponding evaluation results of the physical mechanism model of river fish habitat suitability acture,k ={HSI acture,k,e ,WUA acture,k}.

[0160] Step S4: Evaluation results of the physical mechanism model of the river fish habitat suitability acture,k , construct the fish habitat suitability training sample set D = {(Q k ,Result acture,k )};

[0161] Step S5, using the fish habitat suitability training sample set D = {(Q k ,Result acture,k)}, training the pre-established river fish habitat suitability prediction deep learning model to obtain a river fish habitat suitability prediction deep learning model that has been trained once;

[0162] In the present invention, the deep learning model for predicting river fish habitat suitability includes a data input and preprocessing module, a convolutional coding and attention mechanism module, a decoding and sampling module, and a habitat suitability prediction output module:

[0163] Data input and preprocessing module: Receives and preprocesses terrain data and flow condition data of the target river area; wherein the terrain data is two-dimensional raster data; the preprocessing step includes extracting high-dimensional features from the flow condition data through a multi-layer fully connected network, expanding the flow condition data into a two-dimensional matrix that matches the size of the terrain data; and splicing the processed terrain data, flow condition data, and environmental data in the channel dimension to form fused input data;

[0164] Convolutional coding and attention mechanism module: This module extracts features from the fused input data using a convolutional neural network to obtain a feature vector. It also performs weighted processing on the feature vector using an attention mechanism to generate an attention feature map, which is applied to the output of the convolutional layer to enhance the representation of important features.

[0165] Decoding and sampling module: The attention feature map processed by the attention mechanism is upsampled through the deconvolution network to obtain the upsampled feature map and restore it to the same resolution as the input terrain data:

[0166] Specifically, the formula is used for decoding:

[0167] F up =σ deconv (ConvTranspose2D(F input ;K′,S′,P′)+b deconv )

[0168] Where: F up is the output of the deconvolution network; σ deconv is the activation function, ReLU is used here; F input is the feature map of the input deconvolution network; K′ is the convolution kernel size; S′ is the step size; P′ is the padding; b deconv is the bias term.

[0169] Habitat suitability prediction output module: By processing the upsampled feature map, the deep learning model evaluation results of river fish habitat suitability are obtained;

[0170] Wherein: the deep learning model for predicting the suitability of river fish habitats uses a custom loss function to optimize the accuracy of the prediction of the deep learning model for predicting the suitability of river fish habitats; the total loss function is defined as shown in formula (9):

[0171] L total =αL habitat-RMSE +βL habitat-CE +γL area-logMAE (9)

[0172] Where: α, β and γ are weight coefficients respectively; the total loss function is L habitat-RMSE 、L habitat-CE and L area-logMAE The weighted sum of

[0173] L total is the total loss function;

[0174] L habitat-RMSE is the root mean square error loss function, which measures the continuous difference between the river fish habitat suitability values ​​output by the deep learning model for predicting river fish habitat suitability and the river fish habitat suitability values ​​output by the physical mechanism model for fish habitat suitability;

[0175] L habitat-CE is the cross entropy loss function, which measures the classification results of fish habitat suitability distribution;

[0176] L area-logMAE is a logarithmic absolute error loss function that measures the continuous difference between the suitable area of ​​river fish habitats output by the deep learning model for predicting river fish habitat suitability and the suitable area of ​​river fish habitats output by the physical mechanism model of fish habitat suitability;

[0177] L habitat-RMSE 、L habitat-CE and L area-logMAE Calculated by formula (10), formula (11) and formula (12) respectively:

[0178]

[0179] L area-logMAE =log(WUA prediction,k -WUA acture,k ) (12)

[0180] Of which: HSI prediction,k,e is the flow condition Q k When the deep learning model for predicting the suitability of river fish habitats outputs the river fish habitat suitability value in the calculation unit grid e; HSI acture,k,e is the flow condition Q kWhen the fish habitat suitability physical mechanism model outputs the river fish habitat suitability value in the calculation unit grid e; H acture,k,e is the flow condition Q k When the fish habitat suitability physical mechanism model outputs the river fish habitat suitability value HSI in the calculation unit grid e acture,k,e The binarization result of

[0181] WUA prediction,k is the flow condition Q k When the deep learning model for predicting river fish habitat suitability outputs the river fish habitat suitability area; WUA acture,k is the flow condition Q k When the fish habitat suitability physical mechanism model is used, the river fish habitat suitability area is output.

[0182] In practical applications, Bayesian optimization methods can be used to optimize model hyperparameters and minimize the total loss function on the validation set.

[0183] Step S6, in the flow operating range [Q min ,Q max ], according to the flow interval ΔQ acture Flow interval Re-encrypt and select to obtain L flow conditions, expressed as flow condition Q l , l=1,2,...,L;

[0184] The deep learning model for predicting river fish habitat suitability, which was trained once, was used to simulate the Q of each flow condition. l Corresponding evaluation results of the deep learning model for river fish habitat suitability prediction,l Result of deep learning model evaluation on the suitability of river fish habitat prediction,l Analyze and identify flow-sensitive intervals that are sensitive to the suitability of river fish habitats in, and are the lower and upper bounds of the flow-sensitive interval respectively;

[0185] Specifically, the following method is used to identify the flow-sensitive intervals that are sensitive to the suitability of river fish habitats:

[0186] Step S6.1: Identify the flow-sensitive intervals where the suitable area of ​​fish habitat changes in the river. WUA =[Q WUA,low ,Q WUA,high ]; Q WUA,low and Q WUA,high, are the lower and upper bounds of the flow-sensitive interval for changes in the area of ​​suitable river fish habitats, respectively;

[0187] Step S6.1.1: Use the deep learning model for predicting river fish habitat suitability that has been trained once to simulate each flow condition Q l Corresponding evaluation results of the deep learning model for river fish habitat suitability prediction,l , including the river fish habitat suitability area (WUA) prediction,l And the fish habitat suitability value HSI in each calculation unit grid e prediction,l,e ;

[0188] In step S6.1.2, use formula (13) to calculate the global average difference quotient of the suitable area of ​​river fish habitat

[0189]

[0190] Where: ε is the flow condition, and Q ε , respectively, flow conditions and flow condition ε;

[0191] and WUA prediction,ε , respectively, flow conditions and flow condition ε, the river fish habitat suitability area output by the river fish habitat suitability prediction deep learning model;

[0192] Step S6.1.3, in flow condition Q l , l=1,2,...,L, for every two adjacent flow conditions Q α and Q β , formula (14) is used to calculate the local difference quotient Diff of the suitable area of ​​river fish habitat WUA,α-β :

[0193]

[0194] Among them: WUA prediction,α and WUA prediction,β , respectively, flow condition Q α and Q β When , the river fish habitat suitability area output by the river fish habitat suitability prediction deep learning model;

[0195] Step S6.1.4, compare the local difference quotient Diff of the fish habitat suitability area of ​​each river channel in turn WUA,α-β and the global mean difference quotient of the suitable area of ​​river fish habitat The size relationship of the local difference quotient Diff WUA,α-β Greater than the global mean difference quotient The concentrated flow area is the flow sensitive interval of the change of the suitable area of ​​river fish habitat WUA =[Q WUA,low ,Q WUA,high ]; Q WUA,low and Q WUA,high , are the lower and upper bounds of the flow-sensitive interval respectively;

[0196] Step S6.2: Identify the flow-sensitive intervals where the distribution of fish habitat suitability changes in the river. HSI =[Q HSI,low ,Q HSI,high ]; Q HSI,low and Q HSI,high , are the lower and upper bounds of the flow-sensitive interval for changes in the distribution of fish habitat suitability in rivers;

[0197] In step S6.2.1, use formula (15) to calculate the global mean difference quotient of the distribution of river fish habitat suitability.

[0198]

[0199] Among them: RHSI prediction,ε and They are flow condition ε and flow condition When the river fish habitat suitability value HSI in each calculation unit grid e is output according to the river fish habitat suitability prediction deep learning model, prediction,l,e , the statistically obtained suitability value HSI prediction,l,e The area of ​​the calculation unit grid that is greater than 0 is referred to as the suitable distribution area of ​​river fish habitat;

[0200] Step S6.2.2, in flow condition Q l , l=1,2,...,L, for every two adjacent flow conditions Q α and Q β , using formula (16) to calculate the local difference quotient Diff of the distribution of river fish habitat suitability HSI,α-β :

[0201]

[0202] Among them: RHSI prediction,α and RHSI prediction,β , respectively, flow condition Q α and Q β, the distribution area of ​​river fish habitat suitability obtained based on the river fish habitat suitability value output by the river fish habitat suitability prediction deep learning model;

[0203] Step S6.2.3, compare the local difference quotient Diff of the fish habitat suitability distribution in each river channel in turn HSI,α-β and the global mean difference quotient of the distribution of fish habitat suitability in the river The size relationship of the local difference quotient Diff HSI,α-β Greater than the global mean difference quotient The concentrated flow area is the flow sensitive interval where the distribution of fish habitat suitability changes in the river. HSI =[Q HSI,low ,Q HSI,high ]; Q HSI,low and Q HSI,high , are the lower and upper bounds of the flow-sensitive interval for changes in the distribution of fish habitat suitability in rivers;

[0204] Step S6.3, using formula (17), the flow sensitive interval Interval of the change in the suitable area of ​​river fish habitat obtained in step S6.1.4 is WUA =[Q WUA,low ,Q WUA,high ] and the flow sensitive interval Interval of the change in the distribution of fish habitat suitability in the river obtained in step S6.2.3 HSI =[Q HSI,low ,Q HSI,high ] and take the union of the two sets to identify the flow sensitive interval that is sensitive to the suitability of river fish habitats.

[0205]

[0206] Where: Interval sensitive The flow-sensitive intervals that are sensitive to the suitability of river fish habitats are identified.

[0207] Step S7: Establishing fish habitat suitability incremental learning training sample set D * :

[0208] In the identified traffic sensitive area In the incremental learning interval ΔQ prediction , select K * K flow conditions are randomly sampled from the K flow conditions selected in step S2. ** flow conditions; among them, ΔQ prediction <ΔQ acture ;

[0209] K* Flow conditions and K ** The flow condition combinations are used to form an incremental learning flow condition set;

[0210] Each flow condition in the incremental learning flow condition set is input into the fish habitat suitability physical mechanism model, and the river fish habitat suitability evaluation results corresponding to each flow condition are simulated and then combined to form the fish habitat suitability incremental learning training sample set D. * ;

[0211] Step S8, using the fish habitat suitability incremental learning training sample set D * , performing a second incremental learning training on the deep learning model for predicting the suitability of river fish habitats that has been trained once, to obtain a trained deep learning model for predicting the suitability of river fish habitats;

[0212] Step S9: Use the trained deep learning model for predicting river fish habitat suitability to predict the river fish habitat suitability of the target river area under corresponding flow conditions.

[0213] Specifically, the user inputs the river flow, and the deep learning model for predicting river fish habitat suitability calculates and predicts the river fish habitat suitability area under the flow and the river fish habitat suitability value in each calculation unit grid e based on the input. The model output results can be directly used for river management and protection decisions, such as displaying the prediction result map through a geographic information system or other visualization methods to achieve rapid transmission and application of information.

[0214] The present invention also provides a system for rapid evaluation of fish habitat suitability based on deep learning, comprising:

[0215] The first flow condition determination unit is used to analyze the historical flow data of the target river area for many years and determine the flow condition range [Q min ,Q max ]; among them, Q min and Q max , respectively the lower and upper bounds of the flow operating range; in the flow operating range [Q min ,Q max ], according to the flow interval ΔQ acture , K flow conditions are selected and expressed as flow condition Q k , k=1,2,...,K;

[0216] The fish habitat suitability physical mechanism model one-time training unit is used to use the fish habitat suitability training sample set D = {(Q k ,Result acture,k)}, training the pre-established river fish habitat suitability prediction deep learning model to obtain a river fish habitat suitability prediction deep learning model that has been trained once;

[0217] The fish habitat suitability training sample set acquisition unit is used to evaluate the results of the physical mechanism model of the river fish habitat suitability. acture,k , construct the fish habitat suitability training sample set D = {(Q k ,Result acture,k )};

[0218] A deep learning model for predicting river fish habitat suitability is used to train a fish habitat suitability sample set D = {(Q k ,Result acture,k )} to train and obtain a deep learning model for predicting river fish habitat suitability after one training;

[0219] The second flow condition determination unit is used to determine the flow condition range [Q min ,Q max ], according to the flow interval ΔQ acture Flow interval Re-encrypt and select to obtain L flow conditions, expressed as flow condition Q l , l=1,2,...,L;

[0220] The flow-sensitive interval identification unit is used to simulate the Q of each flow condition using a deep learning model for predicting the suitability of river fish habitats after a single training. l Corresponding evaluation results of the deep learning model for river fish habitat suitability prediction,l Result of deep learning model evaluation on the suitability of river fish habitat prediction,l Analyze and identify flow-sensitive intervals that are sensitive to the suitability of river fish habitats in, and are the lower and upper bounds of the flow-sensitive interval respectively;

[0221] The fish habitat suitability incremental learning training sample set acquisition unit is used to identify the flow sensitive area In the incremental learning interval ΔQ prediction , select K * K flow conditions are randomly sampled from K flow conditions. ** flow conditions; among them, ΔQ prediction <ΔQ acture ;

[0222] K *Flow conditions and K ** The flow conditions are combined to form an incremental learning flow condition set; each flow condition in the incremental learning flow condition set is input into the fish habitat suitability physical mechanism model, and the river fish habitat suitability evaluation results corresponding to each flow condition are simulated to form the fish habitat suitability incremental learning training sample set D. * ;

[0223] Secondary incremental learning training unit, used to incrementally learn the training sample set D using fish habitat suitability * , a second incremental learning training is performed on the deep learning model for predicting the suitability of river fish habitats that has been trained once to obtain a trained deep learning model for predicting the suitability of river fish habitats; the trained deep learning model for predicting the suitability of river fish habitats in the target river area under the corresponding flow conditions is used to predict the suitability of river fish habitats.

[0224] The present invention provides a method and system for rapid evaluation of fish habitat suitability based on deep learning, which has the following characteristics:

[0225] (1) The present invention comprehensively considers the distribution of hydrodynamic environmental factors in the target river area, the interaction between fish schools, and random disturbance terms, and establishes an accurate physical mechanism model of river fish habitat suitability. It can accurately calculate the distribution and area of ​​river fish habitat suitability, but the calculation is time-consuming.

[0226] (2) The present invention establishes a training sample set for various flow conditions based on the distribution of river fish habitat suitability and the river fish habitat suitability area accurately calculated according to the physical mechanism model of river fish habitat suitability. The training sample set is used to train the deep learning model for predicting river fish habitat suitability. Since the sample set has high accuracy, the accuracy of the deep learning model for predicting river fish habitat suitability can be improved.

[0227] Furthermore, by identifying flow-sensitive intervals, the deep learning model for predicting river fish habitat suitability is incrementally trained, thereby obtaining a deep learning model for predicting river fish habitat suitability with a prediction accuracy close to that of the physical mechanism model for river fish habitat suitability. By adopting this deep learning model for predicting river fish habitat suitability, it is possible to not only achieve high-precision prediction of river fish habitat suitability for the target river channel area, but also make up for the problem of long calculation time of the physical mechanism model for river fish habitat suitability. The deep learning model for predicting river fish habitat suitability of the present invention can efficiently and accurately predict the river fish habitat suitability for the target river channel area.

[0228] Therefore, the present invention constructs an efficient deep learning model for predicting river fish habitat suitability based on modules such as convolutional coding and an attention mechanism. By autonomously learning from data samples simulated by a physical mechanism model, it establishes a mapping relationship between the distribution of fish habitat suitability within a river and factors such as flow rate, enabling rapid and accurate assessment of the distribution of fish habitat suitability in a river. This method reduces the time and cost investment of traditional prediction technologies and is applicable to fields such as river ecological protection and water resource management, demonstrating significant practicality and broad application prospects.

[0229] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in detail below with reference to the accompanying drawings. These are merely exemplary embodiments of the present invention. However, it should be understood that the present invention can be implemented in various forms and is not limited to the embodiments described herein. These embodiments are intended to enable those skilled in the art to understand the present invention more clearly and thoroughly.

[0230] 1. Target river area

[0231] This example uses a specific river section as the target channel area. The target channel area is 2.7 km long and has complex terrain, dense shallows, backwaters, and a bottom composed of pebbles, gravel, and silt, with no vegetation along the banks. The target channel area is suitable for spawning and embryonic development of native fish, making it a key local fish habitat.

[0232] 2. Physical simulation samples

[0233] In order to provide sufficient training samples for the deep learning model for predicting river fish habitat suitability, this embodiment uses a physical mechanism model of fish habitat suitability to perform hydrodynamic simulation and fish habitat suitability simulation.

[0234] The physical mechanism model of fish habitat suitability includes a two-dimensional hydrodynamic model, a fish individual movement and renewal model, a two-dimensional fish school distribution density calculation model, a river fish habitat suitability calculation model, and a river fish habitat suitability area calculation model. These three models comprise the fish habitat suitability model.

[0235] The model is solved using the finite volume method, which primarily includes initialization, flux calculation, time marching, source term processing, and iterative solution. In this example, the computational domain R is divided into an unstructured triangular mesh, and the time discretization is performed using the explicit Euler method with a time step of 0.1 seconds, resulting in a total simulation time of 24 hours. Flux calculations are performed using the Roe approximate Riemann method, with boundary condition definitions and source term processing to calculate the hydrodynamic field within the river channel.

[0236] The fish habitat suitability model adopts a method based on the behavioral simulation calculation of individual fish. This method establishes a function to describe the movement of individual fish during the breeding period and combines it with the calculation results of the two-dimensional hydrodynamic model to obtain the flow velocity and water depth suitability distribution in the target river area. The spawning ground is determined based on the distribution of individual fish after they move freely according to the set behavior in the calculation domain range R.

[0237] Specifically, the statistical results of multi-year hydrological data are used as a method for selecting typical flow conditions. The flow data in the long-term hydrological data are subjected to frequency analysis, and the peak flow under different recurrence periods is calculated. The typical flow condition range for calculation is selected according to the needs. In this embodiment, the fish spawning period is mainly studied. The daily hydrological data of the target river area in the embodiment for the past 50 years are statistically analyzed. The cumulative frequency 10% flow is 230m 3 / s, the maximum flow rate is 3320m 3 / s, the flow rate operating range determined in this embodiment [Q min ,Q max ] is 200-3500m 3 / s, flow interval ΔQ acture 15m 3 / s, a total of 330 typical flow conditions, that is, K is 330.

[0238] In this embodiment, the terrain data comes from field scanning by LiDAR and ADCP; the flow data comes from the data of the hydrological station in the river section; the distribution of hydrodynamic elements in key sections is measured by Acoustic Doppler Current Profiler (ADCP), the number of sections is 10, the measurement accuracy is ±0.01 meter, and the data units are meters and meters per second; the bottom sediment data is determined by drone aerial photography and field sampling, and the fish preference, i.e., the correlation coefficient, is comprehensively determined by combining fish ecology research and model parameter calibration based on field data: This embodiment targets the spawning period of fish, involving processes such as adult fish growth, movement, and spawning. The calibration shows that the flow velocity range that can support fish spawning is 0.15m / s to 1.2m / s, the minimum induced flow velocity is 0.1m / s, the maximum flow-controlling capacity is 1.5m / s, and the preferred flow velocity range for spawning is 0.4-0.9m / s; the water depth range that can support fish spawning is 0.2m to 2.5m, and the preferred spawning water depth range is 0.6-1.5m; the reproductive body length threshold is 12cm, and the preferred water temperature for breeding is 9-13℃; after the calculation is completed, the two-dimensional distribution density of fish school ρ(x,y) is used to estimate the fish spawning habitat suitability distribution HSI(x,y).

[0239] Furthermore, the calculation results are standardized by interpolating the terrain grid, the bottom grid, and the calculation results based on the physical mechanism model of fish habitat suitability to form standardized matrix data. In this embodiment, the data is standardized into 256*256 matrix data using a linear interpolation method.

[0240] The fish habitat suitability training sample set D = {(Q k ,Result acture,k )};where k=1,2,...,330;Result acture,k Including HSI acture,k,e and WUA acture,k Among them, HSI acture,k,e is the flow condition Q k When the fish habitat suitability physical mechanism model outputs the fish habitat suitability value in the calculation unit grid e; WUA acture,k is the flow condition Q k When the fish habitat suitability physical mechanism model is used, the river fish habitat suitability area is output.

[0241] 3. Construction and training of a deep learning model for predicting river fish habitat suitability

[0242] The deep learning model for predicting river fish habitat suitability described in this embodiment is built based on the Pytorch deep learning framework and uses libraries such as Numpy, scipy, and Matplotlib for data processing.

[0243] In this embodiment, the deep learning model is a prediction model based on a convolutional neural network and an attention mechanism, primarily used to predict the fish habitat suitability value and area in a river. The model's inputs include two-dimensional terrain data, flow data, and environmental factor data. The flow data is feature extracted through multiple fully connected layers, expanded into a two-dimensional matrix, and then concatenated with the terrain data before being input into the convolutional layer for processing. The model employs a three-layer convolution and deconvolution structure. The convolutional layer has a kernel size of 3×3 and a Reluctant Unit (ReLU) activation function. The maximum pooling layer is used for progressive downsampling to extract deep features of the input data. To preserve more spatial detail, the model introduces skip connections between convolution and deconvolution, connecting convolution layer 1 to deconvolution layer 2, and convolution layer 2 to deconvolution layer 1, respectively. An attention mechanism is introduced to weight important regions in the feature map, ensuring the model's focus on key features. During decoding, the original resolution of the feature map is gradually restored through deconvolution operations, ultimately outputting the fish habitat suitability value and fish habitat suitability area, respectively, through two output heads.

[0244] Furthermore, using the total loss function L total =αL habitat-RMSE +βL habitat-CE +γL area-logMAE Calculate the loss value.

[0245] Further: Use Bayesian optimization method to optimize model hyperparameters, including learning rate, batch size and convolution kernel size, to minimize the total loss function on the validation set: The learning rate search space in this example is 0.0005-0.001, and the batch size search space is 8-64.

[0246] 4. Identify flow-sensitive intervals that are sensitive to the suitability of river fish habitats

[0247] In this embodiment, the identified traffic sensitive interval 400-900m 3 / s.

[0248] 5. Incremental learning of sensitive intervals in river fish habitat suitability models

[0249] Furthermore, during incremental learning training, the dataset used includes all newly added simulation results and sampled data from the original dataset. 3 / s every 5m 3 A flow condition is added every 1s, for a total of 150 flow conditions. The data set used for incremental learning includes the 150 supplementary flow conditions and 150 flow conditions randomly selected from the original data set.

[0250] 6. Execute predictions

[0251] Furthermore, the user inputs the river flow, and the model quickly calculates and predicts the fish habitat suitability value and fish habitat suitability area in the river under the input flow.

[0252] Furthermore, the output results of the deep learning model for predicting river fish habitat suitability can be directly used in river management and protection decisions, such as displaying the predicted distribution map through a geographic information system (GIS) or other visualization methods to achieve rapid transmission and application of information. In this embodiment, the results are output as 256*256 matrix data and Matplotlib is used to visualize the distribution of fish habitat suitability.

[0253] In order to more intuitively demonstrate the implementation effect of this embodiment, Figure 2 The matrixed terrain data of the model input is shown. Figure 3 The comparison of the distribution of fish habitat suitability and the area of ​​fish habitat suitability predicted by the deep learning model for predicting river fish habitat suitability and the physical mechanism model for fish habitat suitability is shown; Figure (a) is 307.5m 3The fish habitat suitability distribution map predicted by the deep learning model for predicting river fish habitat suitability under the flow condition of 1.5 s / s. The predicted fish habitat suitability area is 134670.693 m 2 ; For comparison, Figure (b) is 307.5m 3 The fish habitat suitability distribution map calculated by the fish habitat suitability physical mechanism model under the flow condition of 1.5 s / s shows that the calculated fish habitat suitability area is 139459.673 m 2 .

[0254] Figure (c) is 802.5m 3 The fish habitat suitability distribution map predicted by the deep learning model for predicting river fish habitat suitability under the flow condition of 1.5 s / s. The predicted fish habitat suitability area is 228355.362 m 2 ; For comparison, Figure (d) is 802.5m 3 The fish habitat suitability distribution map calculated by the fish habitat suitability physical mechanism model under the flow condition of 0.13 m / s shows that the calculated fish habitat suitability area is 237727.377 m 2 .

[0255] from Figure 3 It can be proved that the results predicted by the deep learning model for predicting river fish habitat suitability of the present invention are very close to the results calculated by the physical model.

[0256] Figure 4 The relative error results of the fish habitat suitability area predicted by the deep learning model for predicting river fish habitat suitability within the training conditions are shown. That is, under each flow condition, assuming that the fish habitat suitability area predicted by the deep learning model for predicting river fish habitat suitability is WUA1, and the fish habitat suitability area calculated by the physical mechanism model for fish habitat suitability is WUA2, the relative error calculated by the formula (WUA1-WUA2) / WUA2 is used to evaluate the prediction accuracy of the deep learning model for predicting river fish habitat suitability. Figure 4 It can be seen that under various flow conditions, the prediction relative error of the deep learning model for predicting river fish habitat suitability is small and the prediction accuracy is high.

[0257] Therefore, the present invention constructs an efficient deep learning model for predicting river fish habitat suitability based on a deep learning algorithm. By learning sample data simulated by the physical mechanism of the fish habitat suitability model, a deep learning model for predicting river fish habitat suitability is constructed, and an intrinsic mapping relationship between the distribution of fish habitat suitability in the river and factors such as flow rate is established. Based on the given flow data, the distribution of river fish habitat suitability under the flow condition is quickly evaluated.

[0258] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for rapid evaluation of fish habitat suitability based on deep learning, characterized in that: The following steps are involved: Step S1: Analyze the historical flow data of the target river area for many years and determine the flow condition range [Q min ,Q max ]; among them, Q min and Q max , are the lower and upper bounds of the flow rate operating range respectively; Step S2, in the flow operating range [Q min ,Q max ], according to the flow interval ΔQ acture , K flow conditions are selected and expressed as flow condition Q k , k=1,2,...,K; Step S3: Using the physical mechanism model of fish habitat suitability, simulate each flow condition Q k Corresponding evaluation results of the physical mechanism model of river fish habitat suitability acture,k ; Step S4: Evaluation results of the physical mechanism model of the river fish habitat suitability acture,k , construct the fish habitat suitability training sample set D = {(Q k ,Result acture,k )}; Step S5, using the fish habitat suitability training sample set D = {(Q k ,Result acture,k )}, training the pre-established river fish habitat suitability prediction deep learning model to obtain a river fish habitat suitability prediction deep learning model that has been trained once; Step S6, in the flow operating range [Q min ,Q max ], according to the flow interval ΔQ acture Flow interval Re-encrypt and select to obtain L flow conditions, expressed as flow condition Q l , l=1,2,...,L; The deep learning model for predicting river fish habitat suitability, which was trained once, was used to simulate the Q of each flow condition. l Corresponding evaluation results of the deep learning model for river fish habitat suitability prediction,l Result of deep learning model evaluation on the suitability of river fish habitat prediction,l Analyze and identify flow-sensitive intervals that are sensitive to the suitability of river fish habitats in, and are the lower and upper bounds of the flow-sensitive interval respectively; Step S7: Establishing fish habitat suitability incremental learning training sample set D * : In the identified traffic sensitive area In the incremental learning interval ΔQ prediction , select K * K flow conditions are randomly sampled from the K flow conditions selected in step S2. ** flow conditions; among them, ΔQ prediction <ΔQ acture ; K * Flow conditions and K ** The flow condition combinations are used to form an incremental learning flow condition set; Each flow condition in the incremental learning flow condition set is input into the fish habitat suitability physical mechanism model, and the river fish habitat suitability evaluation results corresponding to each flow condition are simulated and then combined to form the fish habitat suitability incremental learning training sample set D. * ; Step S8, using the fish habitat suitability incremental learning training sample set D * , performing a second incremental learning training on the deep learning model for predicting the suitability of river fish habitats that has been trained once, to obtain a trained deep learning model for predicting the suitability of river fish habitats; Step S9: Use the trained deep learning model for predicting river fish habitat suitability to predict the river fish habitat suitability of the target river area under corresponding flow conditions.

2. The method for rapid evaluation of fish habitat suitability based on deep learning according to claim 1, characterized in that: In step S3, the fish habitat suitability physical mechanism model includes a two-dimensional hydrodynamic model, a fish individual movement update model, a fish school two-dimensional distribution density calculation model, a river fish habitat suitability calculation model, and a river fish habitat suitability area calculation model.

3. The method for rapid evaluation of fish habitat suitability based on deep learning according to claim 2, characterized in that: Step S3 is specifically as follows: Step S3.1: Use a two-dimensional hydrodynamic model to simulate each flow condition Q k Corresponding distribution of hydrodynamic environmental factors; Step S3.2, taking into account the distribution of the hydrodynamic environmental factors, the interaction between fish schools and the random disturbance term, the fish individual motion update model is adopted to simulate the position vector of each fish individual i in the fish school at the next time t+1 Thus, the movement behavior of each fish individual i is simulated; Step S3.3, when the fish school position distribution reaches a relatively stable state, the fish school two-dimensional distribution density calculation model is used to obtain the fish school two-dimensional distribution density ρ(x, y); In step S3.4, based on the two-dimensional distribution density of fish schools ρ(x,y), the river fish habitat suitability calculation model of formula (1) is used to estimate the river fish habitat suitability distribution HSI(x,y): Where: R represents the calculation domain range; (x, y) represents the horizontal and vertical coordinates of the calculation domain range R; Step S3.5, convert the river fish habitat suitability distribution HSI(x,y) into discrete river fish habitat suitability value HSI acture,k,e : Discretize the computational domain R into E computational unit grids; identify the river fish habitat suitability value HSI in each computational unit grid e according to the river fish habitat suitability distribution HSI(x,y) acture,k,e ;e=1,2,...,E; Step S3.6: Use the river fish habitat suitability area calculation model of formula (2) to estimate the flow condition Q k Corresponding river fish habitat suitability area WUA acture,k : Where: A is the area of ​​each computational unit grid e; Step S3.7, thus obtaining each flow condition Q k Corresponding evaluation results of the physical mechanism model of river fish habitat suitability acture,k ={HSI acture,k,e ,WUA acture,k }.

4. A method for rapid evaluation of fish habitat suitability based on deep learning according to claim 3, characterized in that: Using a two-dimensional hydrodynamic model, each flow condition Q is simulated. k When the corresponding hydrodynamic environmental factors are distributed, the required basic data include the digital elevation model of the target river area and the river roughness distribution; Each flow condition Q k The corresponding distribution of hydrodynamic environmental factors includes flow velocity, water depth and riverbed distribution.

5. The method for rapid evaluation of fish habitat suitability based on deep learning according to claim 3 is characterized in that: Step S3.2 is specifically as follows: Step S3.2.1, the fish individual motion update model includes a fish individual velocity update model and a fish individual position update model; In each flow condition Q k , using the fish individual speed update model shown in formula (3), the speed vector of each fish individual i at time t is simulated in: is the velocity vector of fish individual i at time t-1; is the change component of the velocity vector of fish individual i at time t-1 caused by the environment; is the change component of the velocity vector of fish individual i at time t-1 caused by the interaction between fish schools; is the velocity vector change component of fish individual i at time t-1 caused by the random disturbance term; Δt is the time interval; Calculated by formula (4): Where: is the ambient velocity drift term; τ is the drift sensitivity coefficient; is the position vector of fish individual i at time t-1; is the environmental flow velocity vector at the location of fish individual i at time t-1; is the swimming speed change term caused by the fish's rheotaxis; χ is the rheotaxis sensitivity coefficient; v pref Preferred swimming speed for fish against current; is the modulus of the ambient flow velocity vector at the location of fish individual i at time t-1; is the unit vector in the direction of the adverse ambient flow velocity; is the velocity vector of fish individual i at time t-1; is the change in the swimming speed of fish caused by the individual fish's preference for environmental factors; γ is the overall weight coefficient of environmental factor preference; M is the number of categories of hydrodynamic environmental factors; is the value of the hydrodynamic environment factor m at the location of fish individual i at time t-1, m = 1, 2, ..., M; is a function for calculating the suitability of the hydrodynamic environmental factor m at the location of fish individual i at time t-1; ω m is the weight coefficient of the hydrodynamic environment factor m; is the comprehensive environmental factor suitability gradient at the location of fish individual i at time t-1, that is, the swimming preference direction of fish individual i at time t-1; Calculated by formula (5): Where: A>0; B>0; q>p; n is the total number of fish individuals in the school; represents the distance vector from fish individual j to fish individual i at time t-1; j = 1, 2, ..., n, i = 1, 2, ..., n, and j ≠ i; The position vector of fish individual j at time t-1; represents the modulus of the distance vector from fish individual j to fish individual i at time t-1; represents the long-distance attraction term between fish individual j and fish individual i at time t-1; represents the short-range repulsion term between fish individual j and fish individual i at time t-1; A is the attraction intensity coefficient; B is the repulsion intensity coefficient; The power exponents p and q control the attraction and repulsion attenuation distances respectively. A smaller p and a larger q indicate that the attraction effect is smaller and the repulsion effect is larger at close distances. The attraction effect decays more slowly with distance than the repulsion effect. is the unit vector between fish individual j and fish individual i, describing the direction of attraction and repulsion; Calculated by formula (6): Where: u(0,v pref ) is the fish’s preferred swimming speed v for random speeds between 0 and pref Random sampling between u(0,2π) and u(0,2π) is random sampling of the swimming direction within 360°; Step S3.2.2, in each flow condition Q k , using the fish individual position update model shown in formula (7), the position vector of each fish individual i at the next time t+1 is simulated in: is the position vector of fish individual i at time t.

6. The method for rapid evaluation of fish habitat suitability based on deep learning according to claim 3, characterized in that: Step S3.3: When the fish school position distribution reaches a relatively stable state, the two-dimensional distribution density of the fish school ρ(x, y) is obtained, specifically: Using formula (8), we can get the two-dimensional distribution density of fish school ρ(x,y): Where: n is the total number of fish individuals in the fish school; σ is the smoothing parameter; (x, y) represents the horizontal and vertical coordinates of the calculation domain range R; (x i ,y i ) represents the horizontal and vertical coordinates of fish individual i when the position distribution of the fish school reaches a relatively stable state, and is the position vector of fish individual i.

7. The method for rapid evaluation of fish habitat suitability based on deep learning according to claim 3, characterized in that: The deep learning model for predicting river fish habitat suitability includes a data input and preprocessing module, a convolutional coding and attention mechanism module, a decoding and sampling module, and a habitat suitability prediction output module: Data input and preprocessing module: Receives and preprocesses terrain data and flow condition data of the target river area; wherein the terrain data is two-dimensional raster data; the preprocessing step includes extracting high-dimensional features from the flow condition data through a multi-layer fully connected network, expanding the flow condition data into a two-dimensional matrix that matches the size of the terrain data; and splicing the processed terrain data, flow condition data, and environmental data in the channel dimension to form fused input data; Convolutional coding and attention mechanism module: This module extracts features from the fused input data using a convolutional neural network to obtain a feature vector. It also performs weighted processing on the feature vector using an attention mechanism to generate an attention feature map, which is applied to the output of the convolutional layer to enhance the representation of important features. Decoding and sampling module: The attention feature map processed by the attention mechanism is upsampled through the deconvolution network to obtain the upsampled feature map and restore it to the same resolution as the input terrain data: Habitat suitability prediction output module: By processing the upsampled feature map, the deep learning model evaluation results of river fish habitat suitability are obtained; Wherein: the deep learning model for predicting the suitability of river fish habitats uses a custom loss function to optimize the accuracy of the prediction of the deep learning model for predicting the suitability of river fish habitats; the total loss function is defined as shown in formula (9): L total =αL habitat-RMSE +βL habitat-CE +γL area-logMAE (9) Where: α, β and γ are weight coefficients respectively; L total is the total loss function; L habitat-RMSE is the root mean square error loss function, L habitat-CE is the cross entropy loss function, L area-logMAE is the logarithmic absolute error loss function, which is calculated by formula (10), formula (11) and formula (12) respectively: L area-logMAE =log(WUA prediction,k -WUA acture,k ) (12) Of which: HSI prediction,k,e is the flow condition Q k When the deep learning model for predicting the suitability of river fish habitats outputs the river fish habitat suitability value in the calculation unit grid e; HSI acture,k,e is the flow condition Q k When the fish habitat suitability physical mechanism model outputs the river fish habitat suitability value in the calculation unit grid e; H acture,k,e is the flow condition Q k When the fish habitat suitability physical mechanism model outputs the river fish habitat suitability value HSI in the calculation unit grid e acture,k,e The binarization result of WUA prediction,k is the flow condition Q k When the deep learning model for predicting river fish habitat suitability outputs the river fish habitat suitability area; WUA acture,k is the flow condition Q k When the fish habitat suitability physical mechanism model is used, the river fish habitat suitability area is output.

8. The method for rapid evaluation of fish habitat suitability based on deep learning according to claim 3 is characterized in that: The following method was used to identify the flow-sensitive intervals that are sensitive to the suitability of river fish habitats: Step S6.1: Identify the flow-sensitive intervals where the suitable area of ​​fish habitat changes in the river. WUA =[Q WUA,low ,Q WUA,high ]; Q WUA,low and Q WUA,high , are the lower and upper bounds of the flow-sensitive interval for changes in the area of ​​suitable river fish habitats, respectively; Step S6.1.1: Use the deep learning model for predicting river fish habitat suitability that has been trained once to simulate each flow condition Q l Corresponding evaluation results of the deep learning model for river fish habitat suitability prediction,l , including the river fish habitat suitability area (WUA) prediction,l And the fish habitat suitability value HSI in each calculation unit grid e prediction,l,e ; In step S6.1.2, use formula (13) to calculate the global average difference quotient of the suitable area of ​​river fish habitat Where: ε is the flow condition, and Q ε , respectively, flow conditions and flow condition ε; and WUA prediction,ε , respectively, flow conditions and flow condition ε, the river fish habitat suitability area output by the river fish habitat suitability prediction deep learning model; Step S6.1.3, in flow condition Q l , l=1,2,...,L, for every two adjacent flow conditions Q α and Q β , formula (14) is used to calculate the local difference quotient Diff of the suitable area of ​​river fish habitat WUA,α-β : Among them: WUA prediction,α and WUA prediction,β , respectively, flow condition Q α and Q β When , the river fish habitat suitability area output by the river fish habitat suitability prediction deep learning model; Step S6.1.4, compare the local difference quotient Diff of the fish habitat suitability area of ​​each river channel in turn WUA,α-β and the global mean difference quotient of the suitable area of ​​river fish habitat The size relationship of the local difference quotient Diff WUA,α-β Greater than the global mean difference quotient The concentrated flow area is the flow sensitive interval of the change of the suitable area of ​​river fish habitat WUA =[Q WUA,low ,Q WUA,high ]; Q WUA,low and Q WUA,high , are the lower and upper bounds of the flow-sensitive interval respectively; Step S6.2: Identify the flow-sensitive intervals where the distribution of fish habitat suitability changes in the river. HSI =[Q HSI,low ,Q HSI,high ]; Q HSI,low and Q HSI,high , are the lower and upper bounds of the flow-sensitive interval for changes in the distribution of fish habitat suitability in rivers; In step S6.2.1, use formula (15) to calculate the global mean difference quotient of the distribution of river fish habitat suitability. Among them: RHSI prediction,ε and They are flow condition ε and flow condition When the river fish habitat suitability value HSI in each calculation unit grid e is output according to the river fish habitat suitability prediction deep learning model, prediction,l,e , the statistically obtained suitability value HSI prediction,l,e The area of ​​the calculation unit grid that is greater than 0 is referred to as the suitable distribution area of ​​river fish habitat; Step S6.2.2, in flow condition Q l , l=1,2,...,L, for every two adjacent flow conditions Q α and Q β , using formula (16) to calculate the local difference quotient Diff of the distribution of river fish habitat suitability HSI,α-β : Among them: RHSI prediction,α and RHSI prediction,β , respectively, flow condition Q α and Q β , the distribution area of ​​river fish habitat suitability obtained based on the river fish habitat suitability value output by the river fish habitat suitability prediction deep learning model; Step S6.2.3, compare the local difference quotient Diff of the fish habitat suitability distribution in each river channel in turn HSI,α-β and the global mean difference quotient of the distribution of fish habitat suitability in the river The size relationship of the local difference quotient Diff HSI,α-β Greater than the global mean difference quotient The concentrated flow area is the flow sensitive interval where the distribution of fish habitat suitability changes in the river. HSI =[Q HSI,low ,Q HSI,high ]; Q HSI,low and Q HSI,high , are the lower and upper bounds of the flow-sensitive interval for changes in the distribution of fish habitat suitability in rivers; Step S6.3, using formula (17), the flow sensitive interval Interval of the change in the suitable area of ​​river fish habitat obtained in step S6.1.4 is WUA =[Q WUA,low ,Q WUA,high ] and the flow sensitive interval Interval of the change in the distribution of fish habitat suitability in the river obtained in step S6.2.3 HSI =[Q HSI,low ,Q HSI,high ] and take the union of the two sets to identify the flow sensitive interval that is sensitive to the suitability of river fish habitats. Where: Interval sensitive The flow-sensitive intervals that are sensitive to the suitability of river fish habitats are identified.

9. A system for implementing the method for rapid evaluation of fish habitat suitability based on deep learning according to any one of claims 1 to 8, characterized in that: include: The first flow condition determination unit is used to analyze the historical flow data of the target river area for many years and determine the flow condition range [Q min ,Q max ]; among them, Q min and Q max , respectively the lower and upper bounds of the flow operating range; in the flow operating range [Q min ,Q max ], according to the flow interval ΔQ acture , K flow conditions are selected and expressed as flow condition Q k , k=1,2,...,K; The fish habitat suitability physical mechanism model one-time training unit is used to use the fish habitat suitability training sample set D = {(Q k ,Result acture,k )}, training the pre-established river fish habitat suitability prediction deep learning model to obtain a river fish habitat suitability prediction deep learning model that has been trained once; The fish habitat suitability training sample set acquisition unit is used to evaluate the results of the physical mechanism model of the river fish habitat suitability. acture,k , construct the fish habitat suitability training sample set D = {(Q k ,Result acture,k )}; A deep learning model for predicting river fish habitat suitability is used to train a fish habitat suitability sample set D = {(Q k ,Result acture,k )} to train and obtain a deep learning model for predicting river fish habitat suitability after one training; The second flow condition determination unit is used to determine the flow condition range [Q min ,Q max ], according to the flow interval ΔQ acture Flow interval Re-encrypt and select to obtain L flow conditions, expressed as flow condition Q l , l=1,2,...,L; The flow-sensitive interval identification unit is used to simulate the Q of each flow condition using a deep learning model for predicting the suitability of river fish habitats after a single training. l Corresponding evaluation results of the deep learning model for river fish habitat suitability prediction,l Result of deep learning model evaluation on the suitability of river fish habitat prediction,l Analyze and identify flow-sensitive intervals that are sensitive to the suitability of river fish habitats in, and are the lower and upper bounds of the flow-sensitive interval respectively; The fish habitat suitability incremental learning training sample set acquisition unit is used to identify the flow sensitive area In the incremental learning interval ΔQ prediction , select K * K flow conditions are randomly sampled from K flow conditions. ** flow conditions; among them, ΔQ prediction <ΔQ acture ; K * Flow conditions and K ** The flow conditions are combined to form an incremental learning flow condition set; each flow condition in the incremental learning flow condition set is input into the fish habitat suitability physical mechanism model, and the river fish habitat suitability evaluation results corresponding to each flow condition are simulated to form the fish habitat suitability incremental learning training sample set D. * ; Secondary incremental learning training unit, used to incrementally learn the training sample set D using fish habitat suitability * , a second incremental learning training is performed on the deep learning model for predicting the suitability of river fish habitats that has been trained once to obtain a trained deep learning model for predicting the suitability of river fish habitats; the trained deep learning model for predicting the suitability of river fish habitats in the target river area under the corresponding flow conditions is used to predict the suitability of river fish habitats.

Images