Machine learning classification of Conus (Gastropoda: Conidae) using geometric morphometrics-based features

Abstract

Machine learning (ML) classifiers for mollusks often require large datasets and are sensitive to degradation since they rely on color, patterns, and texture. This study evaluates whether geometric morphometric (GM) data can serve as effective features for mollusk classification. Five Conus species were analyzed using 16 shell landmarks per specimen, followed by Generalized Procrustes Analysis to remove non-shape variation. Three feature sets (raw coordinates, engineered features, and principal components) were used to train five classifiers: perceptron, k-nearest neighbor (KNN), and support vector machines (SVM) with linear, polynomial, and RBF kernels. Performance was evaluated using 10-fold cross-validation and compared with a baseline convolutional neural network (CNN) trained on shell images. Models trained on raw landmark coordinates and principal components achieved accuracies up to 100%, while engineered features performed substantially worse (62.6–75.2%). The CNN achieved a mean accuracy of 93.84%. The results show that GM-based features provide highly effective representations of shell shape and perform better than a simple CNN trained from scratch on raw images.