Advanced Convolutional Neural Network for Reptile and Amphibian Classification: A Step Towards Biodiversity Monitoring and Conservation

Classification of reptiles and amphibians holds a very crucial place in the respective biodiversity conservation, ecological related studies, and herpetology. In this work, a custom CNN is presented that classifies reptiles and amphibians into 10 species by using 6,045 images taken from the Kaggle dataset. It ranges from having data in different classes-Turtle/Tortoise class with 1,862 images, to 210 images categorized under the Chameleon class. This reflects certain challenges faced by this dataset for class imbalance and underrepresentation. Adaptation of preprocessing methodologies has been done by resizing all images to 224x224 pixels, normalizing images, treating class imbalance, and further dividing into three subsets, with 64% of the images in the training subset, 16% in the validation subset, and 20% in the test subset. The CNN architecture proposed consists of four convolutional layers and two dense layers, optimized with Adam optimizer, having a learning rate of 0.001, training for 50 epochs with augmentation techniques. It reached a general accuracy of 86%, while achieving the topmost F1-scores within well-represented classes like Turtle/Tortoise at 0.98 and Crocodile/Alligator at 0.96. At the same time, some defects in minority classes are represented, namely Gecko at 0.64 and Chameleon at 0.76. This work underlines the generalization capability of the model across diverse species and raises the issue of objectively better representation of the minority classes. The results will be useful in the development of AI applications for biodiversity monitoring and the conservation of species. Improvements in feature extraction, more advanced architectures, and integration of multimodal data are certain aspects left for future work to extend the real-world applications for some limitations.