Explainable Image Classification of Indian Butterflies Through Post-hoc Methods

Deep learning offers powerful image classification capabilities, but its application in Indian butterfly identification is less. This paper focuses on explainable butterfly species classification (BSC) and butterfly family identification (BFI) of Indian butterflies through deep convolutional neural networks (CNNs) combined with post-hoc explanation methods. To the best of our knowledge, this is the first work toward explainable BSC and BFI over Indian butterflies. A dataset of Indian butterfly images is prepared for this research work. This dataset is dedicated for academic purposes and it is termed as KYBDD-5k (Know Your Butterfly Diversity Dataset). For the Indian BSC task, a total of 3730 training images, 800 validation images, and 799 test images are considered from KYBDD-5k. On the other hand, Indian BFI, a total of 3570 training images, 765 validation images, and 766 test images have been considered from KYBDD-5k. To explore the explainable BSC and BFI of Indian butterflies, we consider four CNN architectures and three post-hoc explanation methods. It has been observed that the reliability of such CNN-based classifiers in automatic Indian BSC and BFI has not been extensively studied in the ecological informatics literature. Our experimental evaluation suggests one to consider ResNet-34 and Smooth Grad-CAM++ for explainable BSC. Similarly, EfficientNet-B0 & Grad-CAM is best in explainable BFI as this setup not only maintains a good accuracy but also offers concise, useful insights in the form of heatmaps.