Enhancing crayfish sex identification with Kolmogorov-Arnold networks and stacked autoencoders

Abstract

Crayfish play an important role in freshwater ecosystems, and sex classification is crucial for analyzing their demographic structures. This study performed binary classification using traditional machine learning and deep learning models on tabular and image datasets with an imbalanced class distribution. For tabular classification, features related to crayfish weight and size were used. Missing values were handled using different methods to create various datasets. Kolmogorov-Arnold networks demonstrated the best performance across all metrics, achieving accuracy rates between 95 and 100%. Image data were generated by combining at least five images of each crayfish. Autoencoders were employed to extract meaningful features. In experiments conducted on these extracted features, support vector machines achieved 84% accuracy, and multilayer perceptrons achieved 82% accuracy, outperforming other models. To enhance performance, a novel architecture based on stacked autoencoders was proposed. While some models experienced performance declines, Kolmogorov-Arnold networks showed an average improvement of 3.5% across all metrics, maintaining the highest accuracy. To statistically evaluate performance differences, McNemar's and Wilcoxon tests were applied. The results confirmed significant differences between Kolmogorov-Arnold networks, support vector machines, multilayer perceptrons, and naive Bayes. In conclusion, this study highlights the effectiveness of deep learning and machine learning models in crayfish sex classification and provides a significant example of hybrid artificial intelligence models incorporating autoencoders.