Collaborative Forecasting and Analysis of Fish Catch in Hokkaido From Multiple Scales by Using Neural Network and ARIMA Model

Fishery catch forecasting is a crucial aspect of aquatic research because of its relevance to establishing effective fishery management and resource allocation systems. In this study, we aim to forecast and analyze fish catch by collaboratively processing data using methods at multiple scales. To this end, we propose two computational fishery catch forecasting models. A neural network model based on the multi-timescale features of a convolutional neural network and a long short-term memory neural network is proposed and implemented to forecast short-term measures for the daily catch in the eastern ports of Hokkaido, Japan. Similarly, we propose a long-term catch forecasting and analysis model combining the autoregressive integrated moving average (ARIMA) method and a neural network to explore short-term water temperature and long-term catch dependence in the case of sparse data; we implement this method to investigate the total monthly catch in Hokkaido. The experimental results demonstrate that the proposed methods were able to effectively forecast and analyze fishery catch based on different data scales, volumes, and other complex situations. This is also the first work in the field that considers multiple perspectives.


I. INTRODUCTION
Coastal fishing is an important primary industry in Japan. Hokkaido, an island prefecture in northern Japan, is the largest fishing region in the country, accounting for approximately one-quarter of the coastal fishery products of Japan [1]. However, after peaking in 1984 at 1.282 billion tons, Japanese fishery and aquaculture production declined rapidly and, although this downward trend leveled off after 1995, it has continued to the present day [2]. In this context, accurate catch forecasts can help fishery operators make decisions and perform efficiently. In particular, fishermen and ports can effectively arrange fishing work, and relevant industrial chains such as logistics can be supported to improve logistical efficiency and ensure the freshness of aquatic products [3]. This can greatly relieve pressure on fishery workers. Therefore, catch forecasting is among the most important tasks in the fishery industry [4].
The associate editor coordinating the review of this manuscript and approving it for publication was Eunil Park .
Research on fishery catch forecasting has considered both long-term catch forecasting on a scale of months or years and short-term forecasting on a scale of days. For long-term catch projections, Komiya studied annual catch changes in spiny lobsters and forecasting issues [5]. Leathwick et al. used data on water temperature and salt concentrations with a decision tree to forecast snapper catches and fishing probabilities off the coast over short periods of time [6]. Moreover, Kokaki et al. developed a fish catch forecasting method using a state-space model that described the probabilistic behavior of fish inside nets [7]. Long-term catch forecasting can reflect macrolevel trend changes in the catch, whereas short-term catch forecasting can reflect specific changes in the catch in detail. Therefore, a combination of long-term and shortterm forecasts can comprehensively reflect changes in fish catches, and the two approaches can confirm and complement each other to simulate and summarize the patterns to the greatest extent. Particularly, if changes over different time scales are combined with changes over different geographical scales, macro-and microlevel changes in fishery catches can be demonstrated more effectively. However, in existing works, long-term and short-term catch forecasting studies have often been conducted independently. Hence, analyzing and studying catch data from multiple perspectives remains challenging.
Moreover, the fishery industry is subject to problems involving limitations of available data in practical applications. In fact, in Japan, particularly in Hokkaido, numerous challenges arise in catch data collection and collation. Due to work cycles or actual work conditions, fishery practitioners responsible for catch data often do not have complete and detailed records, leading to a lack of real data and inaccuracies. This situation is particularly severe in the case of short-term catch data, for which it is not easy to achieve the desired forecasting by relying only on the analysis a single catch data type.
To solve these problems, it is necessary to integrate long-term and short-term catch data to develop a more comprehensive analytical approach. To this end, we performed data selection and considered the data management approach of the fishery industry in the Hokkaido region of Japan (Fig. 2). From the figure, it may be observed that catch data are managed in the Hokkaido region via a cascading accounting process, in which data are first collected by individual fishing vessels or ports, then aggregated to the regional fishing associations, and finally pooled from each region to construct the total catch for the Hokkaido region. This method of data management creates a difference between data recorded at high and low levels. Because the high-level data representing a large scope is composed of lower-level data representing a small scope, the statistical period for aggregating the data is longer for high-level data than for lower-level data. In other words, high-level data are generally counted only when low-level data are accumulated to a certain extent. Consequently, long-term data tend to correspond to a wide range of geographical data records. By contrast, shortterm data tend to have a relatively small geographical scope.
To the best of our knowledge, the present work is the first to integrate long-and short-term data processing for fishery data. The proposed approach is implemented in the aforementioned forecasting study on the catch in Hokkaido. Monthly catch data for the entire Hokkaido province and daily catch data for the eastern ports of Hokkaido were used. The proposed approach is illustrated in Fig. 1. We also consider the combination of long-and short-term features according to the actual situation of the data; a model based on neural networks and autoregressive integrated moving average (ARIMA) [8] is proposed to handle short-term characteristics and longterm catch data, and was implemented to forecast long-term catch data. Similarly, a neural network model with a data splitter able to extract long-and short-term features from the data was used to forecast short-term catch data.

II. FISH CATCH FORECASTING A. PROBLEM FORMULATION
Given time-series data D = y 1 , y 2 , y 3 , . . . , each y s is the n-dimensional real vector. Our objective is to establish the estimation of y s+T +1 using T -length sequence data y s+1 , y s+2 , . . . , y s+T , where s = 0, 1, 2, . . .. That is, our goal is to generate a rolling forecast of a series of future signals. Concerning the dimension of each datum, we use two cases for n = 2 and n = 1. In the case of n = 2, each datum is presented by y s = (y temp s , y catch s ) for daily catch 7824 VOLUME 10, 2022 forecasting, In the case of n = 1, each datum is presented by y s = y catch s in monthly catch forecasting. Where y temp s and y catch s ) present the sea temperature data and salmon catch volume data, respectively.

B. FORECASTING OF DAILY CATCH OF PORT USING NEURAL NETWORK
In the process of applying fish catch forecasting, real time-series data often include a mixture of short-and longterm patterns. This is particularly true for port-catch data, which are highly cyclical. Long-term characteristics reflect the labor cycle, seasonal climate, and other changes, whereas short-term features reflect the effects of weather and human activities, among other influences. Accurate time-series forecasting is not possible without considering both types of cyclical patterns. In this study, we aim to solve this problem using a machine learning approach.
Hochreiter et al. proposed the long short-term memory (LSTM) architecture as a machine learning method based on recurrent neural networks (RNNs) [9], with certain information mining capability for long-range temporal data [10]. LSTM models are widely used in the recognition of speech, emotion, and human activity, as well as in load forecasting [11]- [14]. Behavior forecasting and the recognition of activity, speech, and emotion have also been performed by fusing convolutional neural network (CNN) [15] and LSTM [16]- [19] models. The common bases of these methods are the extraction of high-dimensional features with the short sequence feature abstraction ability of CNN models and the subsequent synthesis of highdimensional short-sequence features for temporal forecasting, which is suitable for processing temporal data with local relevance.
On the basis of these methods, we use a specialized neural network model with a structure consisting of a convolutional layer, a data splitter, two recurrent LSTM layers, and a fully connected layer. Input layer is consisted with n × T nodes. We use T length sequence data X s = {x s+1 , x s+2 , . . . , x s+T } (s = 0, 1, 2, . . .) for the casen = 2. The whole data set generated by the given time-series data is presented by {X s } s>0 . Concretely, x 1 s and x 2 s are the variables of the input nodes for the temperature and catch volume data, respectively.
The model uses a convolutional layer to extract highdimensional features, a data splitter to divide the data into different time scales, and a recurrent layer to capture the complex long-term dependency patterns. The splitter divides the data into two time scales; one is input to the LSTM layers and the other is processed by skip-LSTM layers. Finally, a fully connected layer is used to integrate the outputs of the two LSTM layers and generate the final prediction results. This network architecture enables better learning of the periodicity of the input time-series data (Fig. 3). The technical details of each section are described in detail later (Section II B 1-4).

1) CONVOLUTIONAL LAYER
The CNN model performs a convolution operation by generating convolutional check information in a receptive field of an appropriate size, which can express the original data at a higher, more abstract level. Therefore, we use the 1D CNN, which is widely used to process sequence data [20], as the convolutional layer (without a pooling layer) to extract the short-term features of the time series, where the input matrix is X ∈ R n×T , and the k-th filter sweeps through the input and produces C k = c k s+1 , c k s+2 , . . . , c k s+T (k = 1, 2, 3, . . . , K ), where K is the number of filters, as shown in Eq. (1).
where * represents the convolution operation, W k ∈ R n is the weight vector of the convolution kernel, and b k represents the offset. ReLU is used as the activation function, and ReLU (x) = Max (0, x). The structure of a 1D CNN is shown in the convolutional layer section of Fig. 3, with the input time-series data from the input layer, which is a multidimensional matrix X ∈ R n×T . It is convolved from top to bottom as shown by the arrows in the figure, with red representing one filter and green representing another, which can be followed by other filters. We make each vector C k of length T by zero-padding for the input matrix X and the output matrix C of the convolutional layer of size K × T .

2) DATA SPLITTER
LSTM models can remember historical information and understand long-term dependencies. However, longer input results in more information being contained within the system, and when the sequence is excessively long, instability and gradient disappearance occur during the training of a single LSTM. Thus, long-term interdependence cannot be captured. The proposed model mitigates this problem by using a data splitter and bypassing the loop layer, which leverages the real-world cyclical pattern. The data splitter intervals (p ≥ 1) longer data from the input data. In And the original data are used as the . D T is input into the LSTM layer, and the split data D skip T are then input into a new LSTM layer, S-LSTM.

3) RECURRENT LAYER
In the convolutional layer, the output enters the loop layer at the same time as the jump loop layer. The loop layer is an LSTM network that refers to the gate function, which is used to mine the time-series change rules of longer intervals in the time series. The structure is shown in Fig. 4.
D T refers to the data value of the time t input sequence. c t refers to a memory cell or cell state, which is the core of the network and controls the transmission of information. The model structure of a neural network for forecasting fish catch volume is presented. The convolutional layer, data splitter, and recurrent layer are introduced as the hidden structure between the input and output layers. The variables and parameters of the network are presented in the below. The settings of the parameters were appropriately determined for the purpose of fish catch forecasting. The convolutional layer, data splitter, recurrent layer, and fully connected layer are referred to in Section II-B-1, Section II-B-2, Section II-B-3 and Section II-B-3), respectively. i t refers to an input gate, which determines the amount of information that the current D T retains in c t . f t refers to a forget gate, which determines the amount of c t−1 of the cell state from the previous moment is saved to the current c t . o refers to the output gate, which determines the amount of output h t transmitted by c t to the current state. h t−1 refers to the state of the hidden layer at time t − 1. The corresponding formulae for the aforementioned process are provided in Eqs.
where W Di , W Df , W Do , and W Dc refer to the weight matrix of the input gate, forget gate, output gate, and cell state, The computational process of the S-LSTM layer is represented by the LSTM layer; simply replace the input data with the matrix D skip T and its output is denoted as h skip t . We set the output dimensionality of the S-LSTM layer toL, so that the range of all weight matrices W is Finally, a fully connected layer is used to combine the output of the LSTM and S-LSTM layers, as shown in Eq. (8).
W LSTM and W S−LSTM refer to the weight matrix of the LSTM and S-LSTM, respectively. b refers to the offset. The output y catch s+T +1 of the fully connected layer is the forecasting result, which represents the port catch volume data.

4) LOSS FUNCTION AND OPTIMIZATION
The mean square error (MSE) [21] is the default loss function for many forecasting tasks and is calculated as shown in Eq. (9).
whereŷ catch s+T +1 is an estimation of y catch s+T +1 . MSE denotes the MSE of the forecasting result for m days in a row. The optimization strategy adopted in this work is the same as that in the general time series forecasting model; we choose Adam [22] as the optimizer.

C. FORECASTING OF MONTHLY CATCH IN HOKKAIDO USING ARIMA
In the wide range of catch forecasting, the data volume limitation owing to the large time unit does not satisfy the most favorable conditions for neural network models, which characteristically require large amounts of data to perform training. However, these wide-range, long-time unit data, which can represent the trend changes in the catch on a macroscopic scale, are more linear and thus more suitable for processing by traditional statistical methods [23]. Therefore, we use the ARIMA model to forecast the monthly catch for the entire Hokkaido area.
Box and Jenkins introduced the ARIMA model in 1970. Also referred to as the Box-Jenkins methodology, ARIMA is composed of a set of activities for identifying, estimating, and diagnosing ARIMA models with time-series data.
The ARIMA model has three parameters, as indicated in Eqs. (10): p, d, and q. p represents the number of lags used in the forecasting model itself and is also known as the autoregressive (AR) term [24]. d refers to the time-series data that require several orders of differential differentiation to be stable and is also known as the integrated term. q represents the number of lags of the forecasting errors used in the forecasting model and is also known as the moving average (MA) term [25]. It is expressed as follows.
y t = µ + ϕ 1 y t−1 + · · · + ϕ p y t−p + θ 1 e t−1 + · · · + θ p e t−p , where ϕ • denotes the coefficient of AR and θ • denotes the coefficient of MA. In contrast to the daily catch forecast, we use the daily sea temperature over the month to forecast the monthly catch in the wide-area catch forecasting.
Seawater temperature has an unavoidable influence on the capture forecasting problem, but the nature of the ARIMA-based autoregression model leads to an inability to resolve other characteristic variables. Therefore, to enable the model to learn the characteristic information resulting from the water temperature data, we include an LSTM layer in ARIMA to resolve the monthly water temperature data and to enhance the forecasting accuracy, with the structure depicted in Fig. 5. First, we extract the water temperature and fish catch data as two inputs. For example, to forecast the catch data of month t + 1, we use the catch data of the previous month 1 to month t and the daily seawater temperature data of month t + 1.
The data normalization process has often been required for the use of the ARIMA model. Then, we substitute (y catch t − y catch min )/(y catch max − y catch min ) with y catch t as the data normalization. Also, we substitute (y Subsequently, we input the water temperature and fish catch data into the LSTM and ARIMA, respectively, and use two algorithms to process the two data separately. Finally, we use a fully connected layer to integrate the outputs of the two networks as the final outputŷ catch t+1 .

III. EXPERIMENTS
A. DATA

1) DATA SOURCES
We use data from JIJI Fishery News, Fisheries Research Institute, and NEAR-GOOS. Specifically, we use the daily port catch data from JIJI Fishery News with the surface seawater temperature data from NEAR-GOOS as local-area data for the forecasting studies, whereas the monthly Hokkaido catch data from the Fisheries Research Institute and the corresponding surface water temperature data (also from NEAR-GOOS) are used as wide-area data. The details of the data are listed in Table 3.

2) TIME SPACE DISTRIBUTION OF DATA
• Short-term catch volume forecasting We use catch data from June to November of each year from the set-net fishing periods from 2005 to 2015. The location we selected is the eastern part of Hokkaido, where the catch volume is relatively high, and selected four representative ports (as indicated in Fig. 6): Nemuro, Habomai,  Ochiishi, and Rausu, as well as 50 m surface water temperature data at the same locations and at the same times.
• Long-term catch volume forecasting We select the monthly total catch data for Hokkaido from September to February of each year for the period 2000 to 2016, as well as data on the surface water temperature at 50 m for each area and time.

B. EXPERIMENTAL DETAILS 1) SHORT-TERM CATCH VOLUME FORECASTING
For this part of the experiment, as the total amount of data spanned the period 2005 to 2015, we use the port catch and water temperature data for the 2005 to 2014 period from each of the four ports as the training set for the neural network input. Moreover, we use the data for the remaining year, 2015, as the testing set to verify the experimental results. The inputs to the network are the catch and water temperature data from day 1 to day n, whereas the output datum is the catch on day n + 1.

2) LONG-TERM CATCH VOLUME FORECASTING
Data covering the period from September 2000 to February 2015 are used as the training set, and data from September 2015 to February 2016 are used as the testing set. The catch data from month 1 to month n are fed into the ARIMA network as inputs; the daily surface water temperature data from month n + 1 are fed into the neural network, whereas the catch data from month n + 1 are used as the output.
Notably, for the month n + 1 water temperature data, we divide the water temperature data into those of the Sea of Japan and the Pacific Ocean according to the actual sea area of the Hokkaido area (Fig. 7) and input these data into the neural network. Because the distribution of currents in the Sea of Japan differs from that in the Pacific Ocean, and the water temperature is affected by these currents [26], different states occur, which have different effects on the catch. Therefore,  separating the water temperature data of these two bodies of water is more favorable for catch forecasting.
The water temperature data of the Sea of Japan were calculated by averaging the water temperature data from the sea area of Japan and those of Okhostk, whereas the water temperature data of the Pacific Ocean were calculated by averaging the water temperature data from the sea areas of Nemuro, Eastern Erimo, and Western Erimo.

3) COMPREHENSIVE ANALYSIS OF LONG-TERM AND SHORT-TERM CATCH VOLUMES
In this section, we present a comprehensive comparison of long-term and short-term data. Using techniques such as correlation analysis, we analyze the direct links between ports and between ports and catch volumes across Hokkaido. We also explore ways to link long-term data with short-term data.

C. EVALUATION CRITERIA 1) METHODS FOR COMPARISON
The methods used in our comparative evaluation are as follows.
• XGboost [27] is a boosting algorithm. • AR refers to the autoregressive mode. • ARIMA indicates the autoregressive integrated moving average model.
• TCN [28] is the temporal convolutional network, which combines dilated convolution and residual block.
• CNN-LSTM is the network model produced by combining the cell of CNN and that of LSTM.
• LightGBM [29] refers to a light gradient boosting machine, a distributed gradient boosting framework based on a decision tree algorithm.
• LSTM refers to a long short-term memory network model.
• S-LSTM is the proposed neural network model with a data splitter.
• L-ARIMA is the proposed time-series data forecasting model based on LSTM and ARIMA. All these methods are widely used for forecasting timeseries data, covering the scope of statistics and machine learning. Among them, AR and ARIMA belong to traditional statistical methods, XGBoost and LightGBM are both gradient boosting decision trees (GBDTs) in traditional machine learning methods, and TCN belongs to deep learning. Specifically, AR and ARIMA models, as traditional statistical models, are used even more widely in forecasting efforts such as price forecasting [30], wind speed forecasting [31], and even COVID-19 situations [32]. On the other hand, XGBoost applies to both classification and regression and is used in prediction work in industries such as electricity [33] and health care [34], as well as in web text classification, malware classification [27], etc. Similarly, LightGBM has performed well in financial forecasting and cancer patient classification [35] [36]. Although TCN is a recently proposed model, it is also widely used in the fields of the weather forecasting [37], runoff forecasting [38], etc.
In addition, to validate the model's performance combining ARIMA and LSTM (L-ARIMA) is proposed in this paper. Two additive hybrid methods(Additive-ARIMA-LSTM, Additive-ETS-LSTM) and two multiplicative hybrid methods (Multiplicative-ARIMA-LSTM, Multiplicative-ETS-LSTM), which directly combine linear and non-linear models, are used for comparison.
The experiments were conducted using the Tensor-Flow [39] machine learning framework for the Python programming language.

2) METRICS
In this study, we used one metric, namely the root mean squared error (RMSE), to measure the magnitude of the error in the forecasting results as a quantitative criterion. The metrics are defined as follows in Eqs. (11).
whereŷ catch t represents the forecast value, y catch t represents the true value, and m represents the total amount of test data. The RMSE is the square root of the ratio of the sum of the squares of the deviations of the observations from the true value to the number of observations m. It is generally used to measure the deviation of the observations from the true values. Table 4 presents the forecasting errors for the catch at each of the four ports in eastern Hokkaido as predicted by the neural network. We also calculated the error percentage for comparison purposes because of the different scales of the catch volume data from port to port. This corresponds to the error after normalization of the data from the four ports and is more indicative of the model's predictive capability at each port. From the table, the S-LSTM model reduced forecasting error compared to other models commonly used for time-series data forecasting. This model greatly alleviates the unavoidable gradient disappearance and explosion problems in recurrent neural network optimization and improves forecasting accuracy by simultaneously increasing the perceptual range of the model to the data. It can be observed that the VOLUME 10, 2022 forecasting of the network model for the catch at each of the four ports fluctuated to an extent but generally remained within a specific range. The model exhibited slightly better forecasting for Nemuro compared to the other three ports, which was also reflected in the visualization results. A visualization of the forecasting is shown in Fig. 8, with a comparison of the forecast and true values in 2015 for the four ports of Nemuro, Habomai, Ochiishi, and Rausu. The orange curve represents the forecast results and the blue curve represents the true changes in catch. The x-axis represents the time and the y-axis represents the catch (in tons). It can be seen that the network fit the real data well in the part where the capture was zero. In the period from October to November, when the catch volume changed dramatically, the forecasting of the capture volume data from Nemuro exhibited the best fit.

A. RESULTS OF NEURAL NETWORK
Overall, although the network showed good forecasting of the capture volume, it did not exhibit a good forecasting ability for the peak portion of the catch volume, particularly for Habomai and Rausu. Owing to the high sensitivity of the RMSE error to outliers, the RMSE [%] values for Habomai and Rausu in Table 4 were also higher than those for the other two ports. This may be attributed to insufficient feature volume data. Although the water temperature data were also included as a feature quantity in this experiment, it may be possible to capture the variation patterns of the catch quantity more effectively by incorporating richer feature quantities, such as wind speed and weather.
We also performed statistical tests of the forecast results. Concretely, t-test and likelihood-ratio test were applied for the forecast results by S-LSTM and the other methods. The p-values of the t-test results are presented in Table 5, while the p-values of the likelihood-ratio test are shown in Table 6. As a result, the p-values were greater than 0.05. Therefore, the forecast results by S-LSTM were not significantly different from the target data. Similarly, the p-values of the likelihood-ratio test are less than 0.05, and the goodness of fit to the target data of the proposed method S-LSTM is better than the other methods.

B. RESULTS OF ARIMA
The errors of the experiments are shown in Table 7. The upper two tables are comparisons of the RMSE from 2015/09 to  2016/02. The optimal estimation was not obtained by using LSTM. A plausible result was obtained by L-ARIMA. The averaged error for 10 estimations by using L-ARIMA is presented for each month in the lower table. The experimental results indicate that the forecasting error is reduced after the LSTM layer, which is capable of handling the water temperature data, is added to the ARIMA model. The variation characteristics of water temperature help the network model to learn the variation patterns of the catch data. Although certain individual data are far from the real data, the forecast data still fit the real data effectively in general and could fit the trend of the real data appropriately. Moreover, compared with pure ARIMA and other forecasting methods, the L-ARIMA method achieve higher forecasting accuracy despite data volume limitations. Notably, in the case of the LSTM model used only for monthly catch forecasting, the neural network model is not able to perform learning and forecasting effectively, owing to the small amount of data.

C. ANALYSIS OF LONG-TERM AND SHORT-TERM FISH CATCH DATA
We use monthly catch data from September to early December 2015 for Hokkaido and the port catch data to analyze the correlation between the long-term and short-term catches, as shown in Fig. 10. First, although the catches of the four ports are not quantitatively the same, the trends in the timeline  were very similar, and the increasing and decreasing changes in the catches of each port were consistent. This is also reflected in the correlation distribution of catches between ports, as shown in Fig. 9. We analyze and compare the correlation of catch data for the four ports from September 1, 2015, to November 30, 2015. Fig. 9 (a) is the correlation plot generated from the real values of port captures, and Fig. 9 (b) is the correlation plot generated from the results obtained by our proposed network structure with two kinds of LSTM. In Fig. 9 (b), the correlations are reduced due to the errors in forecasting (especially between Habomai and Ochiishi). However, a correlation among the catches of the four ports is evident nonetheless. The variations in the catches among the four ports in the same region are influenced by geographical characteristics and are consistent to a certain extent. For this reason, the correlations can contribute to the prediction of catches in different ports in the same region. And as additional information, the reference data between a related pair of two ports in Fig. 9 might be available for the total forecasting of port catch in eastern Hokkaido.
Second, the monthly catch in Hokkaido is a cumulative value, and the magnitude of its change reflected the amount of real-time catch by combining the monthly catch of the entire island of Hokkaido and the changes in the catch data of its eastern ports. It can be observed that at the beginning of October, when the rapid increase in the monthly catch in Hokkaido approached its peak, the catch of each port also increased. Moreover, when the catch of each port decreased after the second half of October, the change in the monthly catch in Hokkaido also tended to level off. Although the port catch data are slightly inconsistent between the forecast results and the real values, it is nonetheless evident that the real-time changes in port catches are in line with the long-term trend shown by the monthly catches.
According to the aforementioned analysis, there are correlations in both the spatial and time scales between the monthly Hokkaido catch as long-term data and the port catch as short-term data. The analysis and forecasting of the two data types also play a positive role in corroborating one another and help in the analysis of the changes in the catches from different perspectives.

V. CONCLUSION
Two models were used to study the catch of Hokkaido. A neural network model was used to forecast the catch from the local area, short-term data, represented by the daily catch of the port, and the ARIMA model was used to study the catch forecasting from the larger regional area, long-term data, represented by the monthly catch of the entire island of Hokkaido. The results demonstrated the feasibility of the proposed approach adopting two models to forecast the data on different scales; the models were able to perform forecasting of catches with a certain accuracy. Furthermore, the forecasting results can reflect the changing patterns of the catch on different scales by forecasting and analyzing the catch from different angles, which can aid practitioners in understanding the overall data trends and reflect detailed short-term changes to guide daily fishing work.
In future work, we plan to start from the correlation of longand short-term data and explore the possibility of using longand short-term data to complement each other in a capture forecasting problem to achieve scaling of the data over time scales and thus improve forecasting accuracy.