A method and system for spatial prediction of monthly-scale AIS fishing effort based on decoupled spatio-temporal deep learning

Abstract

This invention proposes a method and system for spatial prediction of monthly AIS fishing effort based on decoupled spatiotemporal deep learning. The method includes establishing a prediction network to predict monthly AIS fishing effort. The prediction network uses a spatiotemporal long short-term memory network to extract the temporal variation and spatial distribution characteristics of historical marine environmental variables and AIS fishing effort. Then, it models the generation of potential fishing hotspots and the persistence and diffusion of historical fishing areas, respectively, and makes predictions based on the established models. The two prediction results are then fused at the pixel level according to the contextual information of different spatial locations. Finally, a ReLU function is used to apply non-negative constraints to obtain the predicted results of future monthly AIS fishing effort.

Description

Technical Field

[0001] This invention proposes a method and system for spatial prediction of monthly AIS fishing effort based on decoupled spatiotemporal deep learning, which involves the interdisciplinary field of marine fisheries remote sensing and spatiotemporal deep learning. Background Technology

[0002] Predicting the spatial distribution of marine fishing activities is a crucial technical challenge in dynamic fisheries management, marine spatial planning, and fishing pressure assessment. With the widespread application of Automatic Identification Systems (AIS), identifying fishing vessel operational status based on AIS trajectories and further extracting gridded fishing durations has become an important technical approach for estimating and characterizing AIS-derived apparent fishing effort. Global Fishing Watch has released gridded apparent fishing effort data based on AIS, providing daily and monthly fishing hour information. Kroodsma et al. have used AIS data to characterize the global marine fishing activity footprint, demonstrating that AIS data can be used for large-scale, long-term fisheries activity reconstruction. Unlike traditional catch or resource abundance indicators, AIS apparent fishing effort reflects the spatial distribution pattern of fishing vessel operations in the ocean, providing a spatial reference for identifying high-intensity fishing areas, assessing fishing pressure, and deploying regulatory measures.

[0003] However, predicting apparent fishing effort using AIS at a monthly scale remains challenging. The spatial distribution of fishing vessel operations is influenced not only by marine environmental factors such as sea surface temperature, chlorophyll concentration, ocean currents, salinity, sea level, and dissolved oxygen, but also by the fleet's historical operational experience, fishing ground inertia, and spatial dwelling behavior. Therefore, AIS fishing effort typically exhibits strong spatial clustering, steep hotspot boundaries, and a large amount of zero background sea area. Ordinary continuous-field prediction models are prone to problems such as over-smoothing of hotspots, false positives in background areas, and accumulation of rolling prediction errors when processing this type of data. In recent years, deep learning methods have been increasingly used for fisheries activity prediction and complex spatiotemporal field prediction. For example, Shi et al. proposed the HyFish model, which integrates hydrological factors for predicting the spatial distribution of fishing effort, demonstrating the important role of environmental variables in modeling the spatial distribution of fisheries activities. In the field of general spatiotemporal prediction, ConvLSTM models spatial features and temporal dependencies simultaneously through a convolutional gating structure; PredRNN further introduces spatiotemporal long short-term memory units, enhancing the ability to model long-term spatiotemporal dependencies. These methods provide a fundamental technical approach for marine fisheries activity prediction. Existing technologies typically splice together marine environmental variables and historical fishing activities and input them into a unified backbone network, rarely making explicit distinctions between the two types of processes: "environment-driven new hotspot generation" and "stay-and-diffusion caused by historical operational inertia." At the same time, there is insufficient attention paid to the spatial stability of monthly AIS fishing efforts under rolling multi-month forecasting conditions. Summary of the Invention

[0004] In view of this, in order to fill the gaps and deficiencies in the existing technology, this invention proposes a monthly-scale AIS fishing effort spatial prediction method and system based on decoupled spatiotemporal deep learning, in order to solve the problems that have emerged in the background technology.

[0005] The prediction network of this invention employs a low incremental complexity design. To ensure that the model improves the prediction ability of fishing hotspots while avoiding a significant increase in the number of parameters and computational cost, the network structure reuses shared spatiotemporal coding features as much as possible. Lightweight convolutional operations, such as 1×1 convolution, 3×3 convolution, and dilated convolution, are introduced into each prediction module. Simultaneously, the number of feature channels is strictly controlled in the attention-guided residual predictor, spatial persistence module, and context-aware router to effectively control model complexity. This invention increases only a limited number of parameters and multiply-accumulate operations, enabling improvements in the prediction performance of monthly AIS fishing effort and hotspot areas with relatively low additional computational overhead.

[0006] This invention proposes a monthly-scale AIS fishing effort spatial prediction method and system based on decoupled spatiotemporal deep learning, including the following:

[0007] According to a first aspect of the present invention, the present invention proposes a spatial prediction method for monthly-scale AIS fishing effort based on decoupled spatiotemporal deep learning. The method is characterized by establishing a prediction network to predict monthly-scale AIS fishing effort; wherein the prediction network uses a spatiotemporal long short-term memory network to extract the temporal variation characteristics and spatial distribution characteristics of historical marine environmental variables and AIS fishing effort, then models the generation of potential fishing hotspots and the persistence and diffusion of historical fishing areas, and makes predictions based on the established models; and performs pixel-level adaptive fusion of the two prediction results based on contextual information from different spatial locations. Finally, a non-negative constraint is applied using the ReLU function to obtain the predicted result of future monthly-scale AIS fishing effort.

[0008] Furthermore, the method also includes the following:

[0009] Step S1: Encode the multivariate spatiotemporal sequence of several consecutive months using a spatiotemporal long short-term memory network; provide basic spatiotemporal features for subsequent fishing effort prediction; then integrate the hidden states output by different spatiotemporal long short-term memory network layers to obtain shared fused spatiotemporal features;

[0010] Step S2: An attention-guided residual predictor is used to process the fused spatiotemporal features with channel attention and spatial attention to generate potential hotspot generation responses. At the same time, residual connections are used to preserve the spatial details in the original features and reduce the loss of hotspot boundary information during feature transformation.

[0011] Step S3: Use a multi-scale spatial persistence module to fuse the attention enhancement features with the AIS fishing effort spatial field of the last historical month to extract the persistence of historical fishing hotspots and the neighborhood diffusion trend.

[0012] Step S4: Calculate pixel-level routing weights using a context-aware router and adaptively fuse hotspot generation responses and historical persistent responses; including generating generation branch weights and persistent branch weights based on attention enhancement features and the last observed fishing effort space field.

[0013] Step S5: Apply the ReLU activation function to the non-negative constraint of the adaptive fusion result to obtain the monthly AIS fishing effort prediction result, including the spatial field of the AIS estimated fishing effort for the next month, and calculate the error.

[0014] Further, step S1 includes the following:

[0015] In step S1, multivariate spatiotemporal data is used as an input sequence and encoded in a spatiotemporal long short-term memory network in chronological order. The spatiotemporal long short-term memory network takes multivariate spatiotemporal data from several consecutive months as input. The multivariate spatiotemporal data for each month includes potential temperature, chlorophyll concentration, eastward current velocity, northward current velocity, salinity, sea surface height, dissolved oxygen, and AIS-estimated fishing hours. At the same time, the spatiotemporal long short-term memory network extracts the AIS fishing effort spatial field from the last time step of the input sequence.

[0016] In step S1, the spatiotemporal long short-term memory network transmits information between different time steps and network layers through hidden states and shared spatiotemporal memory to extract marine environmental changes, historical fishing activities and their cross-month evolution features. After encoding all historical sequences, the hidden states output by different layers of the spatiotemporal long short-term memory network at the last time step are fused to obtain shared spatiotemporal features. The shared spatiotemporal features contain both shallow spatial distribution information and deep temporal dependence information, providing basic features for the subsequent two prediction branches.

[0017] Further, step S2 includes the following:

[0018] In step S2, the attention-guided residual predictor includes a convolutional block attention module, a convolutional prediction branch, and a convolutional residual branch;

[0019] In step S2, the shared spatiotemporal features are input into the convolutional block attention module; the convolutional block attention module uses global average pooling and global max pooling to calculate channel attention and enhance the channels of heavy features; then the channel weighted features are calculated to calculate the channel average and maximum values ​​and generate spatial attention to enhance key ocean grid locations; then the spatial attention weighted result is added to the original input features by residual addition to obtain attention-enhanced features;

[0020] Furthermore, the attention-enhanced features are fed into the convolutional prediction branch and the convolutional residual branch, respectively; the convolutional prediction branch maps the multi-channel features into single-channel hotspot generation results; the residual branch directly generates single-channel responses; the hotspot generation response is obtained by adding the two branches together.

[0021] Finally, the attention-guided residual predictor outputs attention-enhanced features and hotspot generation responses.

[0022] Further, step S3 includes the following:

[0023] In step S3, the multi-scale spatial persistence module uses the AIS fishing effort spatial field of the last historical month as a spatial anchor point to extract local persistence information and neighborhood diffusion information of historical hotspots, including the following:

[0024] Attention-enhancing features and spatial anchors are concatenated along the channel dimension to form a continuous predictive context;

[0025] The spliced ​​features are fed into two parallel branches. The local persistence branch uses convolution, group normalization, and LeakyReLU to extract the local structure in the last observed hotspot and its immediate neighborhood. The neighborhood diffusion branch uses dilated convolution, group normalization, and LeakyReLU to expand the receptive field without increasing the convolution kernel size, and is used to extract information about the movement or diffusion of fishing activities to the surrounding sea areas.

[0026] Then, the outputs of the two branches are concatenated along the channel dimension and a single-channel spatial variation is generated by convolution. Subsequently, this variation is added to the residual of the last observed fishing effort field.

[0027] Further, step S4 includes the following:

[0028] In step S4, the context-aware router calculates pixel-level fusion weights for the hotspot generation response and historical persistent response based on the attention enhancement features and the last observed fishing effort field; including the following:

[0029] First, attention-enhancing features and spatial anchors are concatenated along the channel dimension to form a routing context;

[0030] Then the context-aware router outputs two channels, and then generates pixel-level weights between 0 and 1 using the Sigmoid function; the first channel is the weight for generating branches, and the second channel is the weight for continuing branches.

[0031] Then, the two weights are multiplied element-wise with the prediction results of the corresponding branches and added together to complete the weighted fusion.

[0032] Further, step S5 includes the following:

[0033] In step S5, after the context-aware router completes the weighted fusion, it returns the result to the main network. The main network then applies non-negativity constraints through the ReLU function to obtain the monthly-scale AIS fishing effort prediction results, which include the spatial field of the AIS fishing effort for the next month. During the model training phase, a unified ocean mask is used to exclude land grids and filled grids. The error between the prediction results and the actual observations is calculated within the effective ocean grid, and the mask mean square error loss is used to optimize the prediction network.

[0034] According to a second aspect of the present invention, the present invention proposes a monthly-scale AIS fishing effort spatial prediction system based on decoupled spatiotemporal deep learning, for performing a monthly-scale AIS fishing effort spatial prediction method based on decoupled spatiotemporal deep learning as described in any one of the present invention, characterized in that it includes the following:

[0035] The spatiotemporal long short-term memory network module is used to encode multivariate spatiotemporal sequences over several consecutive months, providing basic spatiotemporal features for subsequent fishing effort prediction. Then, the hidden states output by different spatiotemporal long short-term memory network layers are integrated to obtain shared fused spatiotemporal features.

[0036] The attention-guided residual predictor module is used to perform channel attention and spatial attention processing on the fused spatiotemporal features to generate potential hotspot generation responses. At the same time, residual connections are used to preserve spatial details in the original features and reduce the loss of hotspot boundary information during feature transformation.

[0037] A multi-scale spatial persistence module is used to fuse attention-enhanced features with the spatial field of AIS fishing effort in the last historical month to extract the persistence of historical fishing hotspots and neighborhood diffusion trends.

[0038] The context-aware router module is used to calculate pixel-level routing weights and adaptively fuse hotspot generation responses and historical persistent responses; including generating generation branch weights and persistent branch weights based on attention enhancement features and the last observed fishing effort space field.

[0039] The constraint module uses the ReLU activation function to apply non-negative constraints to the fusion results of adaptive fusion, obtaining monthly AIS fishing effort prediction results, including the spatial field of AIS estimated fishing effort for the next month. The error between the prediction results and the actual observations is optimized by using the land-sea mask constraint loss, which only calculates the prediction error within the effective ocean grid.

[0040] According to a third aspect of the present invention, the present invention proposes a monthly-scale AIS fishing effort spatial prediction system based on decoupled spatiotemporal deep learning, comprising an electronic device, wherein the electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements a monthly-scale AIS fishing effort spatial prediction method based on decoupled spatiotemporal deep learning as described in any one of the present invention.

[0041] According to a fourth aspect of the present invention, the present invention proposes a monthly-scale AIS fishing effort spatial prediction system based on decoupled spatiotemporal deep learning, comprising a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements a monthly-scale AIS fishing effort spatial prediction method based on decoupled spatiotemporal deep learning as described in any one of the present invention.

[0042] The present invention has the following advantages:

[0043] The innovation of this invention lies in the design of a decoupled spatiotemporal deep learning network for generation, residence, and routing of apparent fishing effort spatial prediction on a monthly scale using AIS (Aquatic Information System). This network models the environment-driven new hotspot generation process and the residence-diffusion process dominated by historical operational inertia through an attention-guided residual generation module, a multi-scale spatial residence module, and a context-aware routing module, respectively, and performs pixel-level adaptive fusion of the two types of predicted responses. Unlike existing methods that directly concatenate marine environmental variables and historical fishing activities into a unified spatiotemporal prediction backbone, this invention addresses the characteristics of zero expansion of AIS fishing effort, strong spatial clustering, and steep hotspot boundaries. It enhances the identification capability of high-intensity fishing areas, reduces false positive responses in background areas, and strengthens the spatial pattern stability in rolling multi-month predictions, thus providing a more reliable spatial prediction method for dynamic fisheries management, fishing pressure assessment, and maritime surveillance deployment. This invention can be widely applied to practical scenarios such as dynamic marine spatial planning, fisheries management decision support, fishing pressure assessment, and patrol monitoring deployment. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the steps of the present invention.

[0045] Figure 2This is a schematic diagram of the network model architecture of the present invention.

[0046] Figure 3 This is a schematic diagram of the attention-guided residual predictor framework of the present invention.

[0047] Figure 4 This is a schematic diagram of the multi-scale spatial continuity module framework of the present invention.

[0048] Figure 5 This is a schematic diagram of the context-aware router framework of the present invention.

[0049] Figure 6 This is a schematic diagram showing the spatial prediction and residual results of fishing effort for different models in representative months according to the present invention.

[0050] Figure 7 This is a schematic diagram illustrating the prior art defects of the present invention. Detailed Implementation

[0051] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0052] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0053] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0054] like Figures 1 to 7 As shown, this invention proposes a method and system for spatial prediction of monthly AIS fishing effort based on decoupled spatiotemporal deep learning, including the following:

[0055] Currently, spatial prediction of monthly fishing effort based on AIS is constrained by factors such as marine environment, fleet operation inertia, strong clustering of zero-expansion data, and accumulation of rolling errors. This easily leads to problems such as blurred hotspot boundaries, false positives in the background, and long-term spatial drift, severely affecting the accuracy of operational area identification. Figure 7 As shown. The main technical problems currently faced are as follows:

[0056] (1) False positives in background areas. The AIS fishing effort has a significant zero expansion characteristic, with fishing activity being weak or close to zero in most sea areas. In the end-to-end prediction process, existing models are prone to generating unreasonable diffusion responses in low-activity background sea areas, causing areas that are actually close to zero to be predicted as weak hotspots or false hotspots, affecting the reliability of fishing pressure hotspot identification and management decisions.

[0057] (2) Maintaining hotspot boundaries is difficult. AIS fishing efforts typically exhibit strong spatial clustering and steep hotspot boundaries, meaning that a few high-intensity fishing areas are concentrated, while the intensity in the surrounding areas decreases rapidly. When processing this type of data, ordinary spatiotemporal prediction models tend to treat it as a general continuous environmental field and perform smooth extrapolation, resulting in the underestimation of high-intensity hotspots and the blurring of hotspot boundaries, making it difficult to accurately maintain the morphology and intensity details of the actual fishing hotspots.

[0058] (3) Accumulation of rolling prediction errors. In the scenario of continuous multi-month rolling prediction, the prediction results of the previous month will be used as part of the input of the next month, and the prediction error will be gradually propagated over time. Due to the lack of an explicit anchoring mechanism for the inertia of historical operations, existing models are prone to problems such as hot spot position shift, morphological diffusion and background error amplification, which reduce the spatial stability of long-term prediction.

[0059] (4) Coupling of environmental drivers and historical inertia. The spatial distribution of fishing fleets on a monthly scale is not random, but rather the result of the combined effects of marine environmental conditions and the fleet's historical operational inertia. On the one hand, environmental variables such as sea surface temperature, chlorophyll concentration, ocean currents, salinity, sea surface height, and dissolved oxygen can influence potential fishing ground migration and operational choices; on the other hand, historical fishing experience can lead fleets to continue utilizing known high-yield fishing grounds and establish sustained presence or localized expansion in adjacent waters. Existing spatiotemporal prediction models typically directly concatenate environmental variables with historical fishing activities and input them into a unified backbone network, lacking explicit distinction between the two types of processes, making it difficult for the models to separately constrain the generation of new hotspots and the continuation of historical hotspots.

[0060] To address these issues, a spatiotemporal deep learning framework tailored to the characteristics of AIS fishing effort data needs to be constructed. This framework should explicitly decouple environment-driven hotspot generation from the historical inertia-driven dwell-diffusion process, and integrate the two types of predictions through an adaptive fusion mechanism to improve hotspot identification, background false detection suppression, and rolling prediction stability in monthly-scale predictions. Existing research has begun to apply fisheries big data and deep learning methods to the spatiotemporal prediction of fishing activities.

[0061] The AIS fishing effort spatiotemporal prediction network of this invention adopts a low incremental complexity design. To ensure that the model improves the prediction ability of fishing hotspots while avoiding a significant increase in the number of parameters and computational cost, the network structure reuses shared spatiotemporal coding features as much as possible. Lightweight convolutional operations, such as 1×1 convolution, 3×3 convolution, and dilated convolution, are introduced into each prediction module. At the same time, the number of feature channels is strictly controlled in the attention-guided residual predictor, spatial persistence module, and context-aware router to effectively control the model complexity. Compared with PredRNN as the backbone network, this invention only increases the number of parameters and multiply-accumulate operations by a limited amount, and can improve the prediction performance of monthly AIS fishing effort and hotspot areas with lower additional computational overhead.

[0062] In relevant embodiments of the present invention, the relevant statements are as follows:

[0063] The Spatiotemporal Long Short-Term Memory network is represented as STLSTM.

[0064] Attention-Guided Residual Predictor (ARP).

[0065] Multi-scale Spatial Persistence Module (MSSP).

[0066] Context-Aware Router (CAR).

[0067] Spatiotemporal sequence prediction model: PredRNN (Predictive Recurrent Neural Networks)

[0068] like Figure 2 As shown, the overall design of this invention is divided into two parts: a shared spatiotemporal coding stage and a generation-retention-route prediction stage. The shared spatiotemporal coding stage is used to extract the temporal variation characteristics and spatial distribution characteristics of historical marine environmental variables and AIS fishing efforts, providing a unified spatiotemporal feature representation for subsequent predictions. The generation-retention-route prediction stage focuses on modeling the generation of potential fishing hotspots and the persistence and diffusion of historical fishing areas, and performs pixel-level adaptive fusion of the two types of prediction results based on the contextual information of different spatial locations. This further enhances the network's ability to represent the structure and spatial change process of fishing hotspots, effectively addressing the zero-expansion fishing effort spatial field and steep hotspot boundaries. The following describes the content of each module of the network.

[0069] In one embodiment of the present invention, the Spatiotemporal Long Short-Term Memory (STLSTM) network module includes the following:

[0070] The Spatiotemporal Long Short-Term Memory (STLSTM) network is used to encode multivariate spatiotemporal sequences over a continuous 12-month period. This module transmits information between time steps and network layers through hidden states and shared spatiotemporal memories, extracting marine environmental conditions, historical fishing activities, and their cross-monthly variation patterns, providing basic spatiotemporal features for subsequent fishing effort prediction.

[0071] The hidden state fusion module integrates the features of the hidden states output from different STLSTM layers to fuse shallow spatial information and deep temporal information. This module can simultaneously preserve the local boundaries, spatial clustering characteristics, and long-term temporal trends of fishing hotspots, resulting in shared fused spatiotemporal features.

[0072] In one embodiment of the present invention, the attention-guided residual predictor (ARP) module includes the following:

[0073] The Attention-Guided Residual Predictor (ARP) enhances feature responses relevant to the generation of potential fishing hotspots by applying channel and spatial attention processing to fused spatiotemporal features. This module can identify areas where new hotspots, hotspot migrations, or hotspot expansions may occur based on the joint characteristics of marine environmental conditions and historical fishing activities, and generate potential hotspot generation responses. Simultaneously, residual connections are used to preserve spatial details in the original features, reducing the loss of hotspot boundary information during feature transformation.

[0074] In one embodiment of the present invention, the multi-scale spatial persistence module (MSSP) includes the following:

[0075] The Multi-Scale Spatial Persistence Module (MSSP) extracts the persistence of historical fishing hotspots and their neighborhood diffusion trends by fusing attention-enhanced features with the spatial field of AIS fishing effort from the last historical month. This module utilizes local convolution and dilated convolution to acquire information at different spatial scales, and then superimposes the predicted increments onto the final observation field using a residual update method, thereby preserving the location of the main historical fishing zones and their local variation characteristics.

[0076] In one embodiment of the present invention, the context-aware router (CAR) module includes the following:

[0077] The Context-Aware Router (CAR) adaptively fuses hotspot generation responses and historical persistent responses by calculating pixel-level routing weights. This module generates generation branch weights and persistent branch weights based on attention enhancement features and the spatial field of the last observed fishing effort, enabling different ocean grids to select appropriate predicted responses according to their environmental conditions and historical operational status.

[0078] In one embodiment of the present invention, the following is included corresponding to step S1:

[0079] The predictive network of this invention first takes 12 consecutive months of multivariate spatiotemporal data as input. The data for each month includes potential temperature, chlorophyll concentration, eastward current velocity, northward current velocity, salinity, sea surface height, dissolved oxygen, and AIS-estimated fishing hours, for a total of 8 variable channels. The processed input data size is 12×8×64×96.

[0080] Simultaneously, the spatial field of AIS fishing effort is extracted from the last time step of the input sequence, denoted as... , serving as a spatial anchor point for historical fishing activities.

[0081] The input sequences are encoded in a Spatiotemporal Long Short-Term Memory (STLSTM) network in chronological order. This scheme employs a multi-layer STLSTM to process the input data monthly, transferring information between different time steps and network layers through hidden states and shared spatiotemporal memory, thereby extracting features of marine environmental changes, historical fishing activities, and their cross-monthly evolution. After encoding all historical sequences, the hidden states output by different STLSTM layers at the final time step are fused to obtain a shared spatiotemporal feature fusion. This feature contains both shallow spatial distribution information and deep temporal dependency information, providing the foundational features for the subsequent two prediction branches.

[0082] In one embodiment of the present invention, the content corresponding to step S2 includes the following:

[0083] The fused spatiotemporal features are first input into the Attention-Guided Residual Predictor (ARP). The ARP sequentially calculates channel attention and spatial attention through the Convolutional Block Attention Module (CBAM) to enhance the variable channels and ocean grid positions related to the generation of fishing hotspots, resulting in attention-enhanced features H.

[0084] Subsequently, the attention-enhancing features are processed through convolutional mapping to generate potential hotspot responses. Meanwhile, ARP sets up residual connections, which fuse the original fused features with attention features after adjusting them through 1×1 convolution, thereby preserving hotspot boundaries and local spatial details.

[0085] In one embodiment of the present invention, the content corresponding to step S3 includes the following:

[0086] At the same time, the network will enhance the attention of the feature H and the last observed fishing effort field. The input is a multi-scale spatial persistence module (MSSP). The two inputs are first concatenated along the channel dimension, and then processed separately through a local persistence branch and a neighborhood diffusion branch. The local persistence branch uses a 3×3 convolution to extract spatial features of recent fishing hotspots persisting within a local area; the neighborhood diffusion branch uses a dilated convolution to expand the receptive field while extracting information about the movement or diffusion of fishing activities to surrounding areas. The outputs of the two branches are fused to generate the spatial increment of fishing effort. and with By summing the residuals, we obtain the historical persistence / diffusion response. .

[0087] In one embodiment of the present invention, the following is included corresponding to step S4:

[0088] The context-aware router (CAR) combines the attention-enhanced feature H with the last observed field. Further concatenation is performed, and continuous branch weights are generated through convolution operations and the Sigmoid activation function, respectively. and generating branch weights Then, the two weights are respectively compared with... and Element-wise multiplication and element-wise summation of the weighted results enable pixel-level adaptive fusion of the two prediction branches. Independent Sigmoid weights allow hotspot-generated responses and historical persistent responses to function simultaneously at the same spatial location.

[0089] In one embodiment of the present invention, step S6 includes the following:

[0090] Finally, the fusion results are subjected to non-negativity constraints using the ReLU activation function to obtain the spatial field of the AIS-estimated fishing effort for the following month. The output results maintain the same spatial grid size as the input data and can be further denormalized to fishing hours. Based on the set hotspot threshold, a distribution map of high fishing effort areas is generated.

[0091] Introduction to attention-guided residual predictors, multi-scale spatial persistence modules, and context-aware routers:

[0092] This invention proposes a collaborative design of an Attention-Guided Residual Predictor (ARP), a Multi-Scale Spatial Persistence Module (MSSP), and a Context-Aware Router (CAR). By constructing a generation-retention-routing decoupled prediction mechanism, it works in conjunction with the existing STLSTM spatiotemporal encoder and hidden state fusion process to achieve separate extraction and pixel-level adaptive fusion of potential hotspot generation information, historical hotspot persistence, and neighborhood diffusion information. The Attention-Guided Residual Predictor, Multi-Scale Spatial Persistence Module, and Context-Aware Router are described below:

[0093] In one embodiment of the invention, the attention-guided residual predictor is as follows: Figure 3 As shown, this module mainly consists of a Convolutional Block Attention (CBAM) module, a convolutional prediction branch, and a 1×1 convolutional residual branch. First, shared spatiotemporal features are input into the CBAM. The CBAM uses global average pooling and global max pooling to calculate channel attention, enhancing important feature channels. Then, the channel-weighted features are used to calculate the channel average and maximum values, and spatial attention is generated through 7×7 convolution to enhance key ocean grid locations. The spatial attention weighting result is added to the original input features as residuals to obtain the attention-enhanced feature H.

[0094] The attention-enhancing feature H is fed into the convolutional prediction branch and the residual branch, respectively. The convolutional prediction branch sequentially uses 3×3 convolution, group normalization, LeakyReLU, 3×3 convolution, LeakyReLU, and 1×1 convolution to map the multi-channel features into a single-channel hotspot generation result; the residual branch directly generates a single-channel response through a 1×1 convolution. The hotspot generation response is obtained by adding the two branches together.

[0095] ;

[0096] This design enhances features related to potential new hotspots, hotspot migration, and hotspot expansion, while preserving the underlying spatial information in the attention-enhanced features. ARP ultimately outputs the attention-enhanced feature H and the hotspot generation response. .

[0097] In one embodiment of the present invention, the multi-scale spatial persistence module is as follows: Figure 4 As shown, this module uses the spatial field of AIS fishing effort for the last historical month. As spatial anchors, H is used to extract local persistence information and neighborhood diffusion information of historical hotspots. First, H is connected to... The data is concatenated along the channel dimension to form a continuous prediction context. .

[0098] The concatenated features are fed into two parallel branches. The local persistence branch uses ordinary 3×3 convolution, group normalization, and LeakyReLU to extract the local structure within the last observed hotspot and its immediate neighborhood. The neighborhood diffusion branch uses 3×3 dilated convolution with a dilation rate of 2, group normalization, and LeakyReLU to expand the receptive field without increasing the kernel size, and is used to extract information about the movement or diffusion of fishing activities into surrounding waters. The outputs of the two branches are concatenated along the channel dimension and then used to generate a single-channel spatial variation through a 1×1 convolution. Subsequently, this change is added to the residual of the last observed fishing effort field. This design allows the model to learn the necessary intensity variations and spatial diffusion based on the most recent real fishing activity distribution, thereby preserving the main historical fishing zones and reducing large-scale unconstrained shifts in hotspot locations.

[0099] In one embodiment of the present invention, a context-aware router such as Figure 5 As shown, this module is used to calculate pixel-level fusion weights for hotspot generation responses and historical persistent responses based on attention enhancement features and the last observed fishing effort field. First, H is compared with... The routing context is formed by concatenating the data at the channel level.

[0100] The routing context is sequentially processed by a 3×3 convolution, LeakyReLU, and a 1×1 convolution, outputting two channels. Then, a sigmoid function is used to generate pixel-level weights between 0 and 1. The first channel is used to generate branch weights. The second channel is the continuous branch weight. The two weights are multiplied element-wise by the prediction results of their respective branches and then summed:

[0101] ;

[0102] Here, ⊙ denotes element-wise multiplication. The two weights are generated by independent Sigmoid functions, and their sum is not required to be 1. Therefore, hotspot generation responses and historical persistent responses can play a role simultaneously in the same spatial location, or they can be enhanced or suppressed separately according to the local spatiotemporal background. After CAR completes the weighted fusion, the result is returned to the main network; the main network then applies non-negativity constraints through the ReLU function to obtain the spatial field of AIS fishing effort for the next month.

[0103] In one embodiment of the present invention, the proposed generation-residence-routing decoupled spatiotemporal deep learning network is an end-to-end prediction network that can be optimized by minimizing the error between the predicted fishing effort and the actual observation. Considering that the study area includes land and filled grids, the present invention adopts a land-sea mask constraint loss, calculating the prediction error only within the effective ocean grid to avoid invalid regions affecting network parameter updates. The current training configuration defaults to using the mask mean square error loss. The formula for calculating the mask mean square error loss is as follows:

[0104] ;

[0105] in, Let be the predicted fishing effort of the model at the (i,j)th grid position, and be the land-sea shield. When the grid belongs to the effective ocean area... When the grid belongs to a land area or a filled area, This scheme also retains the mask mean absolute error as an optional training loss, and its calculation formula is as follows:

[0106] ;

[0107] Specifically, the mask mean square error imposes a stronger penalty on larger prediction biases, thus prompting the model to focus on optimizing significant errors in areas of high fishing effort; in contrast, the mask mean absolute error is more robust to extreme outliers.

[0108] The present invention also includes related embodiments, including the following:

[0109] The parameter selection in this invention includes the following:

[0110] During training, the input sequence length was set to 12 months, and the prediction length was set to 1 month. Each month included 7 marine environmental variables and 1 AIS-estimated fishing hours channel, with a spatial grid size of 64×96. The network used a two-layer STLSTM with a hidden feature dimension of 64. The maximum number of training epochs was set to 100, and the batch size for both training and validation was 2, employing a 4-step gradient accumulation method, resulting in an equivalent batch size of 8.

[0111] The optimizer uses Adam, and the initial learning rate is set to ,

[0112] The weight decay coefficient is set to During training, the ReduceLROnPlateau learning rate scheduler is used. If the validation set loss does not improve for eight consecutive rounds, the learning rate is reduced to 0.5 of its original value. The minimum learning rate is set to... Meanwhile, the gradient norm is limited to within 5, and training is stopped early if the validation set loss fails to improve for 20 consecutive rounds.

[0113] Table 1 Ablation experiments under single-step prediction

[0114]

[0115] Table 2 Ablation experiments under rolling 12-month predictions

[0116]

[0117] To gain a deeper understanding of the contributions and synergistic effects of each module, we designed structural ablation experiments, as shown in Tables 1 and 2, to evaluate the impact of the Attention-Guided Residual Predictor (ARP), the Multi-Scale Spatial Persistence Module (MSSP), and the Context-Aware Router (CAR) on model performance. In both single-step prediction and the 12-month rolling prediction for 2024, the complete model achieved the best results, demonstrating that the three modules can synergistically improve prediction accuracy and spatial structure preservation capabilities from the perspectives of hotspot generation, historical persistence, and adaptive fusion, respectively.

[0118] After adding MSSP separately, the error index of single-step prediction decreased, R... 2 The scores for SSIM, CSI, and F1 all improved, with SSIM in rolling prediction increasing from 0.7842 to 0.8021, indicating that this module can enhance the modeling of historical hotspot persistence and neighborhood diffusion characteristics. Adding ARP alone resulted in a decrease in the index and increased volatility, suggesting that it is difficult to function stably independently. Further adding CAR resulted in the complete model achieving optimal results in both prediction scenarios, demonstrating that ARP, MSSP, and CAR, through the synergistic effects of generation, residence, and routing, jointly improved the model's prediction accuracy and spatial pattern stability.

[0119] Qualitative comparisons include the following:

[0120] This invention provides an intuitive visual comparison with a representative baseline model. For example... Figure 6 As shown, PredRNN recovers hotspot boundaries and local intensity more smoothly in some months, leading to some diffusion of the spatial range of the main fishing zone and generating more diffuse responses in low-activity background areas. On the other hand, PredRNN-GPR can better maintain the position and aggregation structure of the main nearshore fishing zone, and the recovery of hotspot boundaries and local intensity changes is clearer. Especially in months of high activity and rapid increase, the method of this invention can accurately capture the main hotspot areas; in months of low activity and rapid decrease, it can also effectively reduce false alarm responses in the background area. Regions R1 and R2 in the figure further demonstrate that the method of this invention exhibits better visual consistency in the recovery of hotspot location, boundary, and intensity.

[0121] This invention compares the proposed PredRNN-GPR with seven representative spatiotemporal prediction models, including PredRNN, PFGNet, PredRNN-V2, exPreCast, SeaCast, SwinLSTM, and PredFormer. Tables 3 and 4 provide a quantitative comparison of each model under single-step prediction and a rolling 12-month prediction scenario for a given year, and qualitative analysis of the models is conducted using spatial prediction results for representative months. This invention uses mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), coefficient of determination (R²), structural similarity index (SSIM), critical success index (CSI), and F1 score as performance indicators. Lower values ​​for MAE, MSE, and RMSE indicate smaller prediction errors, and lower values ​​for R² indicate smaller prediction errors. 2 Higher values ​​for SSIM, CSI, and F1 indicate better prediction performance. CSI and F1 are calculated using a fixed threshold of 0.3175 determined based on the validation set. The results in the table represent the mean ± standard deviation of 5 independent training iterations.

[0122] Table 3. Quantitative Comparison of Single-Step Forecast and 2024 Rolling 12-Month Forecast

[0123]

[0124] Table 4. Quantitative Comparison of Rolling Forecast and 2024 Rolling 12-Month Forecast

[0125]

[0126] In the single-step prediction scenario, PredRNN-GPR achieved the best results across all evaluation metrics, with MAE, MSE, and RMSE of 0.0884, 0.0255, and 0.1596, respectively. 2 The RMSE and SSIM scores reached 0.7138 and 0.8440, respectively, while the CSI and F1 scores reached 0.6717 and 0.8036, respectively. Among them, PFGNet performed closest to the method of this invention in terms of continuous spatial field error index, and PredRNN performed closest in terms of hotspot identification index. Compared with PredRNN, the method of this invention reduced RMSE by 0.0046, increased R² by 0.0165, and increased CSI and F1 by 0.0237 and 0.0172, respectively, indicating that the generation-residence-routing structure can simultaneously improve continuous spatial field reconstruction and high-fishing-effort hotspot identification.

[0127] Under the 12-month rolling forecast scenario in 2024, PredRNN-GPR still achieved the best average results across all metrics, with MAE, MSE, and RMSE of 0.0971, 0.0291, and 0.1703, respectively; R² and SSIM of 0.6737 and 0.8224, respectively; and CSI and F1 of 0.6407 and 0.7808, respectively. Compared to PredRNN, the method of this invention reduced RMSE by 0.0035, increased R² by 0.0131, and improved CSI and F1 by 0.0111 and 0.0082, respectively. Although rolling forecasts are affected by the accumulation of autoregressive errors, resulting in overall performance lower than single-step forecasts, the method of this invention still maintains the highest average continuous spatial field index and hotspot identification index, indicating its good stability in multi-month rolling forecasts.

[0128] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for spatial prediction of monthly AIS fishing effort based on decoupled spatiotemporal deep learning, characterized in that, A prediction network was established to predict monthly AIS fishing efforts. The prediction network uses a spatiotemporal long short-term memory network to extract the temporal variation and spatial distribution characteristics of historical marine environmental variables and AIS fishing efforts. Then, the generation of potential fishing hotspots and the persistence and diffusion of historical fishing areas were modeled and predicted based on the established models. The two prediction results were then fused at the pixel level according to the context information of different spatial locations. Finally, a non-negative constraint was applied using the ReLU function to obtain the prediction results of future monthly AIS fishing efforts.

2. The method for spatial prediction of monthly AIS fishing effort based on decoupled spatiotemporal deep learning according to claim 1, characterized in that, The method also includes the following: Step S1: Encode the multivariate spatiotemporal sequence of several consecutive months using a spatiotemporal long short-term memory network; provide basic spatiotemporal features for subsequent fishing effort prediction; then integrate the hidden states output by different spatiotemporal long short-term memory network layers to obtain shared fused spatiotemporal features; Step S2: An attention-guided residual predictor is used to process the fused spatiotemporal features with channel attention and spatial attention to generate potential hotspot generation responses. At the same time, residual connections are used to preserve the spatial details in the original features and reduce the loss of hotspot boundary information during feature transformation. Step S3: Use a multi-scale spatial persistence module to fuse the attention enhancement features with the AIS fishing effort spatial field of the last historical month to extract the persistence of historical fishing hotspots and the neighborhood diffusion trend. Step S4: Calculate pixel-level routing weights using a context-aware router and adaptively fuse hotspot generation responses and historical persistent responses; including generating generation branch weights and persistent branch weights based on attention enhancement features and the last observed fishing effort space field. Step S5: Apply the ReLU activation function to the non-negative constraint of the adaptive fusion result to obtain the monthly AIS fishing effort prediction result, including the spatial field of the AIS estimated fishing effort for the next month, and calculate the error.

3. The method for spatial prediction of monthly AIS fishing effort based on decoupled spatiotemporal deep learning according to claim 2, characterized in that, Step S1 includes the following: In step S1, multivariate spatiotemporal data is used as an input sequence and encoded in the spatiotemporal long short-term memory network in chronological order. The spatiotemporal long short-term memory network takes multivariate spatiotemporal data from several consecutive months as input. The multivariate spatiotemporal data for each month includes potential temperature, chlorophyll concentration, eastward current velocity, northward current velocity, salinity, sea surface height, dissolved oxygen, and AIS-estimated fishing hours. At the same time, the AIS fishing effort spatial field is extracted from the last time step of the input sequence as the final observed fishing effort spatial field, which is used in subsequent multi-scale spatial persistence modules and context-aware routers. In step S1, the spatiotemporal long short-term memory network transmits information between different time steps and network layers through hidden states and shared spatiotemporal memory, thereby extracting marine environmental changes, historical fishing activities and their cross-month evolution characteristics. After encoding all historical sequences, the hidden states output by different layers of the spatiotemporal long short-term memory network at the last time step are fused to obtain shared spatiotemporal features. These shared spatiotemporal features contain both shallow spatial distribution information and deep temporal dependency information, providing basic features for the subsequent two prediction branches.

4. The method for spatial prediction of monthly AIS fishing effort based on decoupled spatiotemporal deep learning according to claim 3, characterized in that, Step S2 includes the following: In step S2, the attention-guided residual predictor includes a convolutional block attention module, a convolutional prediction branch, and a convolutional residual branch; In step S2, the shared spatiotemporal features are input into the convolutional block attention module; the convolutional block attention module uses global average pooling and global max pooling to calculate channel attention and enhance important feature channels; Subsequently, the channel average and maximum values ​​are calculated for the channel-weighted features, and spatial attention is generated to enhance key ocean grid locations; Then, the spatial attention-weighted result is added to the original input features by residual addition to obtain the attention-enhanced features; Furthermore, the attention-enhanced features are fed into the convolutional prediction branch and the convolutional residual branch, respectively; the convolutional prediction branch maps the multi-channel features into single-channel hotspot generation results; the residual branch directly generates single-channel responses; the hotspot generation response is obtained by adding the two branches together. Finally, the attention-guided residual predictor outputs attention-enhanced features and hotspot generation responses.

5. The method for spatial prediction of monthly AIS fishing effort based on decoupled spatiotemporal deep learning according to claim 4, characterized in that, Step S3 includes the following: In step S3, the multi-scale spatial persistence module uses the AIS fishing effort spatial field of the last historical month as a spatial anchor point to extract local persistence information and neighborhood diffusion information of historical hotspots, including the following: Attention-enhancing features and spatial anchors are concatenated along the channel dimension to form a continuous predictive context; The spliced ​​features are fed into two parallel branches. The local persistence branch uses convolution, group normalization, and LeakyReLU to extract the local structure in the last observed hotspot and its immediate neighborhood. The neighborhood diffusion branch uses dilated convolution, group normalization, and LeakyReLU to expand the receptive field without increasing the convolution kernel size, and is used to extract information about the movement or diffusion of fishing activities to the surrounding sea areas. Then, the outputs of the two branches are concatenated along the channel dimension and convolution is used to generate a single-channel spatial variation. The change was then added to the residual of the last observed fishing effort field.

6. The method for spatial prediction of monthly AIS fishing effort based on decoupled spatiotemporal deep learning according to claim 5, characterized in that, Step S4 includes the following: In step S4, the context-aware router calculates pixel-level fusion weights for the hotspot generation response and historical persistent response based on the attention enhancement features and the last observed fishing effort field; including the following: First, attention-enhancing features and spatial anchors are concatenated along the channel dimension to form a routing context; Then the context-aware router outputs two channels, and then generates pixel-level weights between 0 and 1 using the Sigmoid function; the first channel is the weight for generating branches, and the second channel is the weight for continuing branches. Then, the two weights are multiplied element-wise with the prediction results of the corresponding branches and added together to complete the weighted fusion.

7. The method for spatial prediction of monthly AIS fishing effort based on decoupled spatiotemporal deep learning according to claim 6, characterized in that, Step S5 includes the following: In step S5, after the context-aware router completes the weighted fusion, it returns the result to the main network. The main network then applies non-negative constraints through the ReLU function to obtain the monthly-scale AIS fishing effort prediction results, which include the spatial field of the AIS fishing effort for the next month. Finally, based on the land and filled grids included in the spatial distribution characteristics, the land-sea mask constraint loss, which only calculates the prediction error within the effective ocean grid, is used to optimize the error between the prediction results and the actual observations.

8. A monthly-scale AIS fishing effort spatial prediction system based on decoupled spatiotemporal deep learning, used to execute the monthly-scale AIS fishing effort spatial prediction method based on decoupled spatiotemporal deep learning as described in any one of claims 1 to 7, characterized in that, Includes the following: The spatiotemporal long short-term memory network module is used to encode multivariate spatiotemporal sequences over several consecutive months, providing basic spatiotemporal features for subsequent fishing effort prediction. Then, the hidden states output by different spatiotemporal long short-term memory network layers are integrated to obtain shared fused spatiotemporal features. The attention-guided residual predictor module is used to perform channel attention and spatial attention processing on the fused spatiotemporal features to generate potential hotspot generation responses. At the same time, residual connections are used to preserve spatial details in the original features and reduce the loss of hotspot boundary information during feature transformation. A multi-scale spatial persistence module is used to fuse attention-enhanced features with the spatial field of AIS fishing effort in the last historical month to extract the persistence of historical fishing hotspots and neighborhood diffusion trends. The context-aware router module is used to calculate pixel-level routing weights and adaptively fuse hotspot generation responses and historical continuous responses. This includes generating generation branch weights and persistence branch weights based on attention enhancement features and the spatial field of the last observed fishing effort; The constraint module uses the ReLU activation function to apply non-negative constraints to the fusion results of adaptive fusion, obtaining monthly AIS fishing effort prediction results, including the spatial field of AIS estimated fishing effort for the next month. The error between the prediction results and the actual observations is optimized by using the land-sea mask constraint loss, which only calculates the prediction error within the effective ocean grid.

9. A monthly-scale AIS fishing effort spatial prediction system based on decoupled spatiotemporal deep learning, comprising an electronic device, wherein the electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a method for spatial prediction of monthly AIS fishing effort based on decoupled spatiotemporal deep learning as described in any one of claims 1 to 7.

10. A monthly-scale AIS fishing effort spatial prediction system based on decoupled spatiotemporal deep learning, comprising a computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a method for spatial prediction of monthly AIS fishing effort based on decoupled spatiotemporal deep learning as described in any one of claims 1 to 7.

Images