A sample model for applying feature selection and machine learning techniques to estimate and manage crayfish populations

Abstract

Sustainable fishing and aquaculture production require not only data collection but also its proper application. Another factor influencing the study’s direction is dataset size. Regardless of dataset size, the most critical aspect of data use is feature selection. This study focuses on estimating sustainable fishing and aquaculture production by selecting relevant features and appropriate algorithms for both sexes and all individuals in the crayfish population living in the Atikisar Reservoir. For this purpose, the Mean method was used for data preprocessing, Pearson correlation analysis for feature selection, and Multiple Linear Regression (MLR), Support Vector Machine Regression (SVMR), Gradient Enhancement Regression (GBR), Sampled Sum (Bootstrap Combination, Bagging) Regression (BAR), Random Forest Regression (RFR), and k-Nearest Neighbors Regression (k-NNR) algorithms were used for prediction and stock management. Additionally, the data obtained from the population were evaluated within the framework of the classical length-weight relationship. The MLR, GBR, and RFR algorithms showed the highest coefficients of determination (R² = 0.98) for female individuals; the MLR, GBR, BAR, RFR, and k-NNR algorithms had the highest R² values (0.96) for male individuals; and the MLR, GBR, RFR, and k-NNR algorithms achieved the best coefficients of determination (R² = 0.98) for the entire population. Analyzing crayfish populations using the GBR algorithm will enable us to make predictions with the highest accuracy level and to manage populations, after processing the data with the Mean method from data preprocessing methods, and performing feature selection with Pearson correlation analysis. Regression analyses and machine learning algorithms, applied to overall and sexual length and weight of crayfish caught in the Atikhisar Reservoir, revealed positive allometric growth in the TL-TW and meat yield relationship for females, males, and the entire population.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-45535-9.

Introduction

With the rapid increase in the world population, there are various difficulties in meeting the needs of qualified animal proteins. When the nutritional regimes in many countries are examined, and the economic conditions are considered, crayfish are an important food source in eliminating these difficulties. Furthermore, considering global climate change, projections for 2050 predict that total global food demand will increase by + 30% to + 62%, and the population at risk of hunger will increase by −91% to + 30%¹. Crayfish has a high nutritional value due to its high moisture and carbohydrate content, especially protein, low fat content, maximum digestibility, low lipid value, and balanced and sufficient amounts of essential amino acids. Crayfish meat is a good source of low-calorie protein, as well as a rich source of minerals such as sodium, potassium, calcium, and magnesium. In addition, crayfish are more advantageous than other cultivated species due to their water and environment requirements, their tolerances, and the fact that they can be cultivated extensively/semi-intensively. Although crayfish epidemics, overfishing, and increased water pollution have caused fluctuations in crayfish production in recent years, the crayfish has an important place among the species produced in aquaculture². Due to the projected increase in the world population, global warming, and fluctuations in crayfish production volumes, both natural stocks and cultured production plans should be made more carefully, improved, and developed, and the obtained data should be used more successfully. The current status of a population allows us to assess its past and future². Population or stock indices³ such as mortality rates, age/size distribution, number of age groups in the population or stock, reproductive success, growth rate, and population or stock size not only help us identify factors such as disease and/or problems in the population’s past and present state, nutrient status, predator presence, and cannibalism in carnivorous populations, but also help us predict the future of the population, implement appropriate stock management, and ensure sustainability. Monitoring population indices is a continuous process necessary for the sustainability and assessment of populations and species, and is both time-consuming and costly. Because a precise determination of a population size is only possible by capturing all individuals within that population, the meat yield of individuals, particularly under extensive conditions, is an important parameter used in population estimation. When estimating crayfish populations, metric and meristic characteristics such as length-weight are determined, and calculations are made using various formulas^4,5.

Artificial intelligence (AI) is computer software that simulates how humans think to perform complex tasks such as analyzing, reasoning, and learning. Machine learning (ML) is a subset of AI that uses algorithms trained on data to produce models that can perform such complex tasks. ML has led to its applications in different academic and industrial fields, including nanotechnology⁶, healthcare⁷, energy efficiency⁸, biomass⁹, and environmental protection¹⁰, due to its capacity to process complex data to perform various tasks, including unsupervised learning, supervised learning, and reinforcement learning. Some researchers have also proposed a combination of evolutionary techniques and machine learning tools to estimate a target parameter¹¹.

The learning mechanisms of predictive models can be divided into two categories: traditional predictive models and machine learning predictive models. When traditional forecasting models use simple rules or methods to predict future values, such as the naive method (NM), moving average (MA), autoregressive (AR), and autoregressive integrated moving average (ARIMA), machine learning forecasting models use more complex computational methods and model structures to extract underlying patterns from data, such as multilayer perceptrons (MLP), recurrent neural networks (RNN), convolutional neural networks (CNN), and transformers¹². MLP is a type of neural network machine learning model that has attracted considerable attention¹³. It is a fully connected feedforward artificial neural network and has been used as a benchmark to test the predictive performance of other predictive models^14,15.

Data preprocessing is the process of evaluating, filtering, organizing, and coding data so that a machine learning algorithm can understand and use the resulting output. This can significantly reduce the processing power and time required to train or infer a machine learning or AI system. Therefore, this study aims to identify the most accurate machine learning technique for predicting crayfish growth performance by selecting the most suitable quantitative features using data preprocessing operations, for instance, missing value and feature selection. This will make management of the crayfish stocks more efficient, effective, and sustainable.

Materials and methods

Study area and data collection

The dataset used in this study was created by using narrow-clawed crayfish data obtained from Atikhisar Reservoir, which was built on Sarıçay Creek between 1971 and 1975 for irrigation and flood control purposes in Çanakkale-Türkiye. Map of the study area generated using ArcGIS Desktop version 10.8 (https://www.esri.com/en-us/arcgis/products/arcgis-desktop/overview) (Fig. 1). A total of 6470 narrow-clawed crayfish were captured, of which 2606 (40.28%) were female and 3864 (59.72%) were male. According to Atay¹⁶, gender determination was performed macroscopically, and the gender ratio was determined from the frequency distribution of length from months. Each sample has 20 attributes, which are presented in Table 1. The lengths of body segments extending from the rostrum to the telson, which are used to determine comparative population growth, marketable crayfish size, meat yield, systematic differentiation, and differences between male and female crayfish, were measured using a digital caliper with a precision of 0.01 mm. Weights were weighed using a scale with a precision of 0.01 g for broodstock and a scale with an accuracy of 0.0001 g for juveniles^2,17–21.

Fig. 1.

Atikhisar reservoir.

Table 1.

Attribute information.

Number	Attribute name	Type	Description
1	SN	numeric	Sample number
2	CD	date	Catch date
3	G	string {m, f}	Gender
4	TL	numeric	Total length
5	CL	numeric	Cephalothorax length
6	CW	numeric	Cephalothorax width
7	AL	numeric	Abdominal length
8	AW	numeric	Abdominal width
9	SL-R	numeric	Scissor length-right
10	SL-L	numeric	Scissor length-left
11	SW-R	numeric	Scissor width-right
12	SW-L	numeric	Scissor width-left
13	CH	numeric	Cephalothorax weight
14	AH	numeric	Abdominal weight
15	AMH	numeric	Abdominal meat weight
16	SH-R	numeric	Scissor weight-right
17	SH-L	numeric	Scissor weight-left
18	SMH-R	numeric	Scissor meat weight-right
19	SMH-L	numeric	Scissor meat weight-left
20	TW	numeric	Total weight

Length-weight relationship and meat yield

In crayfish, the relationship between weight and length is calculated as in Eq. 1. By taking the logarithm of both sides of Eq. 1, the relationship between weight and length is made linear (Eq. 2). When checking the significance of the calculated b value, the test statistic is calculated using Eq. 3⁴. To determine meat yield, the abdomen, claws, and scissors were cut with the help of a scalpel, and the meat inside was weighed directly to determine its weight²¹.

1

2

L, total length (TL); W, total weight (TW); a and b, constant parameters of the Eq. 

3

Sx, standard deviation of log (L) values; Sy, standard deviation of log (W) values); n, the number of individuals used in the calculation; r², coefficient of determination of the log (L) and log (W) values.

Data pre-processing

Data preprocessing is the first step in machine learning; here, data is transformed and encoded so that the machine can quickly navigate or parse through it. In other words, it can be interpreted as the model algorithm immediately analyzing the data’s characteristics²². In the case of missing data in a data set, one solution is to simply ignore the record with missing data. However, in a set with a large amount of missing data, this method significantly increases the error rate. Instead, methods such as Hot/Cold Deck, Expectation Maximization, Carry Forward Last Observation, Multiple Imputation, Mean, Median, Mode, and the last observation can be used to impute missing data²³. If a variable follows a normal distribution, the Mean method is the appropriate choice for filling in missing values ​​because they are equivalent. Therefore, the Mean method was used to fill in missing values within the dataset using values from the same class for each class²⁴.

Feature selection

The features in a dataset are one of the most important factors affecting prediction performance. Therefore, it is necessary to accurately identify a sufficient number of features that represent the original dataset, which will reduce training time, improve data quality, and increase model performance. Thus, eliminating unnecessary/irrelevant features from the dataset will both reduce the learning algorithm’s runtime and create a better classification model^25,26. Correlation analysis is a statistical method that reveals the direction, degree, and significance of the relationship between variables. The coefficient indicating the direction and degree of the relationship is called the correlation coefficient and is denoted by “r.” The correlation coefficient takes values ​​between − 1 and + 1. An r value close to −1 indicates a negative relationship between variables, while values ​​close to + 1 indicate a positive relationship. As the correlation coefficient value approaches “0,” the relationship between two variables weakens. A value close to ± 1 indicates a perfect correlation: As one variable increases, the other tends to increase (if positive) or decrease (if negative). A coefficient value between ± 0.50 and ± 1 indicates a strong correlation. A value between ± 0.30 and ± 0.49 indicates a moderate correlation. A value below 0.29 indicates a small correlation. A value of zero indicates no correlation. Pearson’s correlation is the most commonly used correlation metric to measure the degree of relationship between linearly related variables. In this study, well-known feature selection method, such as the Pearson correlation coefficient, was applied to rank the contents of total length (TL), cephalothorax length (CL), cephalothorax width (CW), abdominal length (AL), abdominal width (AW), scissor length-right (SL-R), scissor length-left (SL-L), scissor width-right (SW-R), scissor width-left (SW-L), cephalothorax weight (CH), abdominal weight (AH), abdominal meat weight (AMH), scissor weight-right (SH-R), scissor weight-left (SH-L), scissor meat weight-right (SMH-R), and scissor meat weight-left (SMH-L), and total weight (TW) of crayfish samples according to their effects on the growth of crayfish. In this study, a threshold of |r| ≥ 0.5 was chosen for the Pearson correlation coefficient to ensure that the relationships between variables are considered moderate to high and practically significant. This threshold value is consistent with classifications commonly used in the literature and was determined to prevent weak relationships from affecting the analysis results²⁷.

Machine learning methods used for numerical prediction

Machine learning is the process of using mathematical algorithms to derive insights and make predictions from a given dataset. This process involves training models on existing data to make predictions and decisions, rather than relying solely on pre-programmed instructions^28,29. Machine learning is broadly divided into three types: supervised learning, unsupervised learning, and reinforcement learning. Supervised learning, which involves learning a function that maps an input to an output based on sample input-output pairs, is a machine learning model. Supervised learning problems are categorized as regression and classification problems. Regression analysis includes various machine learning methods that allow predicting a continuous outcome variable (y) based on the value of one or more predictor variables (x). The most important distinction between classification and regression is that classification predicts discrete class labels, while regression facilitates the prediction of a continuous quantity³⁰. In the present study, six different machine learning algorithms were used to create regression models. Multiple linear regression (MLR), support vector machine regression (SVMR), Gradient Boosting Regression (GBR), Sampled Sum (Bootstrap Aggregation, Bagging) Regression (BAR), Random Forest Regression (RFR), and K-Nearest Neighbors Regression (k-NNR) are popular machine learning algorithms. Some of these algorithms, such as Support Vector Machines and Random Forest, can also be used for classification; however, in the scope of this study, MLR, SVMR, GBR, BAR, RFR, and k-NNR are discussed from a regression perspective. The study’s dataset includes a total of 6470 crayfish specimens. The dataset is divided into two subsets: training and testing. The training set is the subset used to train the machine learning models. Conversely, the test set is the subset used to test the model and evaluate it against an unknown test set. The data in this study were randomly split into the training set, comprising 70% of the data, and the test set, comprising 30% of the data. In the present study, while other hyperparameters were left at their default values, only the number of trees and the randomness parameter were adjusted to evaluate the baseline performance of the Random Forest model. The number of trees (n_estimators) in the RandomForestRegressor model was set to 100. This value is a commonly used parameter in the literature that balances the generalization performance and computational cost of the Random Forest model. To prevent the model from producing randomly dependent results and ensure the reproducibility of the experiments, the random_state parameter was fixed at 42³¹.

The models were made using an Apple MacBook Air M1 Chip, 8 GB, 256 GB MacOS Ventura 13.3.1 on the operating system Windows 10 Pro operating system with implemented using Python language with Scikit-Learn, Keras, and Tensorflow libraries (Fig. 2).

Fig. 2.

TL-TW relationships of the crayfish population in the Atikhisar Reservoir.

Performance metrics

In this study, R-squared (coefficient of determination) (R²), root mean square error (RMSE), mean absolute error (MAE), relative absolute error (RAE), and root relative square error (RSE) values ​​were used to measure the predictive performance of the models. Using these metrics, it is possible to determine which technique is most suitable for this dataset. In linear regression, R² (Eq. 5) is a measure of how closely the data points align with the fitted line. It is also known as the coefficient of determination. R² is a metric that represents the predictive performance of regression models. It is a positive value between 0 and 1. The ideal value for R² is 1. The closer the R² value is to 1, the better the model. RMSE (Eq. 6) measures how far the regression line of the forecast error is from the data points. MAE (Eq. 7) is the error rate of the growth forecast model. RAE (Eq. 8) takes the total absolute error and normalizes it by dividing it by the total absolute error of the simple estimator. RSE (Eq. 9) calculates the square root of the sum of the squared differences between the predicted value and the true value divided by the sum of the squared differences between the true values and the mean value³². The lower the calculated values ​​for the four-error metrics and the closer the coefficient of determination to 1, the more accurate the results. Although these four different metrics and coefficients of determination are not sufficient in terms of calculated values alone, they are significant when evaluated together².

4

5

6

7

8

Results

In the data set, the TL value of the population varied between 22.77 mm and 164.97 mm. The average TL in female individuals was 106.90 ± 17.08 mm, the average TL in male individuals was 108.61 ± 13.65 mm, and the overall average TL value of the population was calculated as 107.92 ± 15.79 mm. When the TW distributions of the population were examined, it was found that they varied between 0.17 g and 130.35 g, and the female, male, and population mean TW values ​​were 29.40 ± 13.59 g, 38.55 ± 14.56 g, and 34.86 ± 16.35 g, respectively. Length-weight relationships of the crayfish population in the Atikhisar Reservoir were examined logarithmically using TL-TW linear regression analysis for males, females, and mixed females and males (Table 2). TL-TW relationship graphs were given in Fig. 3 for the entire population, including females, males, and females/males.

Table 2.

TL-TW relationship equations and correlation values in the crayfish population of Atikhisar Reservoir (N♀♀= 2606; N♂♂= 3864; N♀♀♂♂= 6470) ^a(p < 0,05), ^b(p < 0,01).

Gender	Log y = Log a + b x Log X	a ± Sh	b + Sh	r + Sh
♀♀	LogTW=−3.75739 + 3.03366 LogTL	0.00018 ± 0.0236	3.03366 ± 0.014b	0.943 ± 0.079a
♂♂	LogTW=−3.80196 + 3.088597 LogTL	0.00016 ± 0.0345	3.088597 ± 0.034b	0.928 ± 0.085a
♀♀♂♂	LogTW=−3.87024 + 3.116796 LogTL	0.00013 ± 0.024	3.116796 ± 0.014b	0.943 ± 0.079a

Fig. 3.

Pearson correlation matrix.

According to the regression analysis, positive allometric growth was detected in terms of TL-TW relationships for females, males, and the entire population.

To determine whether there is a relationship between total weight (TW) and other attributes in the study, the Pearson Correlation method was used, and is shown in Fig. 3. This way, by excluding irrelevant data from the calculations, a potential bias in the calculation can be avoided. The values ​​in the matrix in Fig. 3 range from − 1 to 1. A value of 1 indicates a very strong linear relationship, while a value of −1 indicates a very strong inverse relationship. 0 indicates no relationship. For example, based on the matrix, it can be concluded that TW has a strong linear relationship (R = 0.98) with cephalothorax weight (CH), while there is a very weak linear relationship between abdominal width (AW; R = 0.61) and scissor weight-left (SMH-L; R = 0.60). Similar inferences can be made for the other attributes.

The features with high correlation values (correlation_target[abs(correlation_target) > 0.5]) with TW are listed as follows: (1) TL, (2) CL, (3) CW, (4) AL, (5) AW, (6) SL-R, (7) SW-R, (8) SL-L, (9) SW-L, (10) CH, 11) AH, 12) AMH, 13) SH-R, 14) SMH-R, 15) SH-L, 16) SMH-L. Machine learning methods were first applied to female individuals in the population, then to male individuals, and then to the entire population using features obtained from the Pearson correlation matrix.

MLR, GBR, and RFR stand out as the algorithms with the best coefficient of determination (R² = 0.98), while k-NNR and BAR stand out as the algorithms with the best error values for the female crayfish (♀♀). The SVMR model has the lowest coefficient of determination. After finding the MSE and MAE values using the metrics method in the SciKit-Learn library, the sqrt function in the Math library was used to find the RMSE value. R² and 4 different error metric results for MLP, SVMR, GBM, BAR, RFR, and k-NNR models are given in Table 3.

Table 3.

Performance metric results for TW based on measurement values ​​in the female crayfish (♀♀) (R², coefficient of determination; MAE, mean absolute error; RMSE, root mean square error; RAE, relative absolute error; RSE, root relative square error; MLR, Multiple linear regression; SVMR, support vector machine regression; GBR, Gradient Boosting Regression; BAR, Sampled Sum (Bootstrap Aggregation, Bagging) Regression; RFR, Random Forest Regression; k-NNR, k-Nearest Neighbors Regression).

Regression models	Performance metrics
	R 2	MAE	RMSE	RAE	RSE
MLR	0.98	0.85	1.94	0.08	0.0002
SVMR	0.96	0.94	2.91	0.09	0.0002
GBR	0.98	0.78	1.75	0.71	0.0004
BAR	0.97	0.89	2.25	0.08	0.0003
RFR	0.98	0.79	2.10	0.07	0.0002
k-NNR	0.97	1.32	2.25	0.12	0.0002

MLR, GBR, BAR, RFR, and k-NNR stood out as the algorithms with the best coefficient of determination (R²=0.96), while RFR stood out as the algorithm with the best error value for the male crayfish (♂♂). The SVMR model had the lowest coefficient of determination. R² and four different error metrics for the MLR, SVMR, GBMR, BAR, RFR, and k-NNR models are presented in Table 4.

Table 4.

Performance metric results for TW based on measurement values ​​in the male crayfish (♂♂) (R², coefficient of determination; MAE, mean absolute error; RMSE, root mean square error; RAE, relative absolute error; RSE, root relative square error; MLR, Multiple linear regression; SVMR, support vector machine regression; GBR, Gradient Boosting Regression; BAR, Sampled Sum (Bootstrap Aggregation, Bagging) Regression; RFR, Random Forest Regression; k-NNR, k-Nearest Neighbors Regression).

Regression models	Performance metrics
	R 2	MAE	RMSE	RAE	RSE
MLR	0.96	1.29	3.42	0.10	0.0006
SVMR	0.95	1.38	3.76	0.10	0.0001
GBR	0.96	1.29	3.38	0.09	0.0009
BAR	0.96	1.35	3.56	0.10	0.0004
RFR	0.96	1.27	3.57	0.09	0.0005
k-NNR	0.96	1.71	3.50	0.13	0.0007

MLR, GBR, RFR, and k-NNR algorithms have the best coefficient of determination (R²) of 0.98, while SVMR and BAR stood out as the algorithms with the best error values for the entire population (♀♀♂♂). Table 5 shows the R² and four different error metric results obtained with the MLP, SVMR, GBR, BAR, RFR, and k-NNR models.

Table 5.

Performance metric results for TW based on measurement values ​​in the crayfish population (♀♀♂♂) (R², coefficient of determination; MAE, mean absolute error; RMSE, root mean square error; RAE, relative absolute error; RSE, root relative square error; MLR, Multiple linear regression; SVMR, support vector machine regression; GBR, Gradient Boosting Regression; BAR, Sampled Sum (Bootstrap Aggregation, Bagging) Regression; RFR, Random Forest Regression; k-NNR, k-Nearest Neighbors Regression).

Regression models	Performance metrics
	R 2	MAE	RMSE	RAE	RSE
MLR	0.98	1.11	2.27	0.09	0.0002
SVMR	0.97	1.13	2.62	0.09	0.0002
GBR	0.98	1.10	2.45	0.09	0.0004
BAR	0.97	1.13	2.68	0.09	0.0003
RFR	0.98	1.05	2.50	0.08	0.0003
k-NNR	0.98	1.46	2.60	0.11	0.0006

Using six different machine learning methods to predict the results for 2606 female individuals (♀♀) in the population, which are presented in Fig. 4.

Fig. 4.

TW estimation results for the female crayfish (♀♀) (Abbreviations: MLR, Multiple Linear Regression; SVMR, Support Vector Machine Regression; GBR, Gradient Boosting Regression; BAR, Sampled Sum (Bootstrap Aggregation, Bagging) Regression; RFR, Random Forest Regression; k-NNR, k-Nearest Neighbors Regression).

The prediction results obtained by applying machine learning methods to 3864 male individuals in the population are given in Fig. 5.

Fig. 5.

TW estimation results for the male crayfish (♂♂) (MLR, Multiple Linear Regression; SVMR, Support Vector Machine Regression; GBR, Gradient Boosting Regression; BAR, Sampled Sum (Bootstrap Aggregation, Bagging) Regression; RFR, Random Forest Regression; k-NNR, k-Nearest Neighbors Regression).

The prediction results obtained from machine learning methods applied to the entire population, comprising 6470 narrow-clawed crayfish, are shown in Fig. 6.

Fig. 6.

TW estimation results for the entire population (♀♀♂♂) (Abbreviations: MLR, Multiple Linear Regression; SVMR, Support Vector Machine Regression; GBR, Gradient Boosting Regression; BAR, Sampled Sum (Bootstrap Aggregation, Bagging) Regression; RFR, Random Forest Regression; k-NNR, k-Nearest Neighbors Regression).

Discussion

As with all living organisms, knowing the current status of aquaculture and fisheries populations is crucial for population management and sustainability. The sex ratio is a crucial factor in stock assessments. The reproductive balance and self-renewal capacity of species are directly affected by the sex ratio in the ecosystem³³. In the present study, 40.28% of the individuals caught from Atikhisar Reservoir were females and 59.72% of males. The prevalence ratios of female and male freshwater crayfish within the population were found to be significantly different (p < 0.05). The female/male ratio for the entire population was calculated as 0.67/1.00. The excess of males may mainly arise from the differences in the activities of the sexes. Still, it may also arise from the fact that males, which grow faster than females, are more prevalent in the catch composition in areas with intensive fishing³⁴. The sex ratio for different crayfish populations was determined to be 1:1³⁵. Disturbance of the sexual composition of crayfish may indicate increased fishing pressure on the species, and that the reproductive capacity of the species may be destroyed in unsuitable ecological conditions in the future. The differences in sample sizes of male/female individuals depending on the seasons are thought to stem primarily from differences in the activities of the sexes. Groves, explaining the surplus of male individuals in a different way, found that in areas with intensive hunting, male individuals, which grow faster than females, are more prevalent in the prey composition. Furthermore, a numerical difference in the sample population is explained by the greater physical activity of male individuals, the sedentary lifestyle of egg-carrying female individuals, and the difference in molting time between male and female crayfish^36,37.

While correlation explains the degree to which two variables that appear to be independent of each other behave in harmony with each other, Pearson correlation allows us to calculate the effect of a change in one variable on the other variable. To clearly see the impact of feature selection on TW, a Pearson correlation analysis was performed, and it was determined that the effects of CH (R = 0.98), CL (R = 0.90), CW (R = 0.90), TL (R = 0.87), and AL (0.80) values ​​were high, respectively (Fig. 3). These data indicate that using Pearson correlation as the feature selection method is appropriate in this study. Similarly, it has also been noted that, in the analysis of 11 years of data from the Shanghai Stock Exchange, the Pearson correlation coefficient demonstrated better performance because it can quickly identify linear correlation features, such as opening price and trading volume, and efficiently perform feature screening on some data^38,39. Crayfish are known to exhibit agnostic behavior and, in some cases, can be successful against rivals and predators with this behavior⁴⁰. Crayfish use their claws when exhibiting aggressive behaviors. During aggressive fights, crayfish can damage and even sever their extremities, but these extremities can regenerate later. The regenerated extremities may be morphologically smaller than the original. The Pearson correlation matrix showed that the lowest interaction with TW was with SMH-L (R = 0.60). Because agnostic behaviors are commonly exhibited by males in crayfish populations⁴¹, and as the individuals captured in this study were predominantly male (59.72%), this can be considered a natural situation. Considering Leatherland’s³ view on population indices, competition for reproduction and habitat among male crayfish, among other factors, is thought to be the key influence. The female/male ratio in the population and the Pearson correlation matrix provided results that supported each other.

Any strategy to conserve species and habitats inevitably requires precise knowledge of biodiversity. Indeed, a complex taxonomy makes the conservation of threatened species even more problematic. To address this, precise analytical methods are needed. Currently, these methods utilize morphometric characters combined with multivariate analyses. However, analyses of a small number of selected phenotypic features are ineffective in distinguishing stocks. Crustacean weight can be estimated using multiple linear regression with morphological features. There is a nonlinear relationship between weight and these features in crustaceans. In studies involving artificial intelligence, regression models are known to perform better at capturing nonlinear relationships⁴². For practical production application, choosing the right algorithm is just as important as selecting the appropriate features. This raises the question: which algorithm is more suitable than others? To answer this, the study used multiple linear regression (MLR), support vector machine regression (SVMR), gradient boosting regression (GBR), sampled sum (bootstrap aggregation, bagging) regression (BAR), random forest regression (RFR), and k-nearest neighbor regression (k-NNR) algorithms. Similarly, different algorithms were used to accurately estimate the fishing grounds of Antarctic krill, also known as stock regions⁴³. In this study, six different regression algorithms were used to evaluate female (♀♀) and male (♂♂) individuals, as well as the entire population (♀♀♂♂). The MLR, GBR, and RFR algorithms showed the highest coefficients of determination (R² = 0.98) for female individuals; the MLR, GBR, BAR, RFR, and k-NNR algorithms had the highest R² values (0.96) for male individuals; and the MLR, GBR, RFR, and k-NNR algorithms achieved the best coefficients of determination (R² = 0.98) for the entire population. Although fewer samples and algorithms were used, the study on the meat yield of red swamp crayfish in Wuhu city, Anhui province, China, employed the XGBoost (XGB), RF, and GBR algorithms on a sample of 1,392 individuals, with GBR achieving the best estimation results⁴⁴. In this study, the highest coefficients of determination were found for females, males, and the entire population using the MLR, RFR, and GBR algorithms. Additionally, the BAR and k-NNR algorithms produced the highest coefficients of determination in male individuals, while the k-NNR algorithm performed best in the entire population. The SVMR algorithm, however, had the lowest coefficient of determination across all individuals and models. Among the decision tree regression (DCR), GBR, linear regression (LR), and RFR algorithms used to estimate hematological blood parameters in carp fed probiotic feed, GBR and RFR provided the highest prediction accuracy⁴⁵. The results of recent studies have followed similar patterns to the results of this study. Additionally, k-NNR and BAR had the lowest error values for female crayfish, RFR for male crayfish, and SVMR and BAR across the entire crayfish population, based on performance metrics in this study. In Tables 3 and 4, and 5, an RSE of 0.0002 indicates that R² is mathematically almost equal to 1, that the model error is negligible compared to the natural variance of the data⁴⁶. All machine learning models used in this study were assessed using the same dataset and experimental conditions. The results indicate that the Support Vector Regression (SVR) model performed lower than the other models. The primary reason for this is that the use of a linear kernel in the SVR model restricts its ability to adequately capture the nonlinear relationships present in the dataset, thereby limiting the model’s generalization capabilities⁴⁷. Consequently, SVR appears to be less suitable for the current data structure. Regression analyses and machine learning algorithms, applied to overall and sexual length and weight of crayfish caught in the Atikhisar Reservoir, revealed positive allometric growth in the TL-TW and meat yield relationship for females, males, and the entire population.

Conclusion

This study, the first of its kind in terms of scope and size, focused on selecting features and implementing the correct algorithm. Pearson correlation analysis was used for feature selection, along with six machine learning algorithms. The MLR, GBR, and RFR algorithms produced the best results in sex-based and population-based assessments. When comparing this study’s results with those of current studies, it was observed that the GBR algorithm stands out even with small sample sizes. Sustainable fishing and aquaculture production requires not only the need for data but also the correct use of that data. Determining fishing regulations to sustain fisheries is often complicated, especially when limited biological data is available. As a result, according to the present study, analyzing crayfish populations using the GBR algorithm will enable us to make predictions with the highest accuracy level and to manage populations, after processing the data with the Mean method from data preprocessing methods, and performing feature selection with Pearson correlation analysis. For sustainable fisheries, regulations must rely on biological information supported by fishermen’s knowledge, enforced by expert management, updated as new information emerges, and assessed using appropriate data analysis methods.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Y.G. applied machine learning techniques and checked the manuscript.N.G. collected data from the reservoir, analyzed data and wrote the main manuscript.

Funding

Grants promoted the present study from Atatürk University as a scientific research project (Project ID Number: FCD-2023-12154).

Data availability

All data that support the findings of this study are included within this paper and its Supplementary Information files. You can also access the raw data for the article via the following link.https://docs.google.com/spreadsheets/d/12Nza5vIfj0SSL96hf5LJZ4FNz0YRalo7/edit? usp=sharing\&ouid=113247 593921991620057\&rtpof=true\&sd=true.

Declarations

Competing interests

The authors declare no competing interests.