YOLO Goes Green: Unveiling Sundarbans Biodiversity using Leaf Image Classification

Abstract

In the Sundarbans mangrove forest, the identification of plant species is crucial for ecological research, tourism, and conservation efforts. This study presents an innovative image-based plant recognition model for the Sundarbans, leveraging deep convolutional neural networks (DCNN). A comprehensive dataset of 28,081 high-resolution images, representing 36 different plant species native to the Sundarbans, was compiled to facilitate the model's development. Moreover, transfer learning was applied to a pre-trained YOLOv8 model, enabling the system to use previously learned features while adapting to the unique characteristics of Sundarbans flora. The results demonstrate that transfer learning is highly effective in identifying Sundarbans plant species. In addition, the high performance of the model has been further explained through the incorporation of the Grad-CAM technique. This approach allows researchers and conservationists to visually understand which parts of an image contribute the most to the prediction of the model, thus fostering greater trust and insight into the decision-making process of the neural network. Furthermore, six different YOLO pre-trained models were compared, with the YOLOv8x-cls model outperforming the other variants. In particular, the YOLOv8x-cls model achieved a top-1 accuracy of 98.38% and a top-5 accuracy of 99.76%.