Towards transparency in AI: Explainable bird species image classification for ecological research

Birds are indicators of biodiversity and ecosystem health and play an essential role in maintaining the balance of natural ecosystems. However, urbanization, deforestation, and technological advancements have severely affected bird habitats, leading to a significant decline in species diversity. Manual detection and recognition of bird species based on morphological and behavioral characteristics during birdwatching and surveys are challenging and require expertise, making deep learning a promising alternative. Deep learning techniques offer the advantage of automatic identification, which can significantly enhance the efficiency and accuracy of species recognition tasks. However, the black-box nature of these models presents a significant issue, as it is difficult to understand their internal decision-making processes, leading to concerns about their reliability and trustworthiness. This study addresses these issues by employing Explainable Artificial Intelligence (XAI) to enhance the transparency of deep learning models for bird species image classification. In this paper, a three-stage XAI-based approach is proposed, involving transfer learning, Local Interpretable Model-Agnostic Explanations (LIME), and Intersection over Union (IoU) scores to assess model performance. Six pretrained models are evaluated on the CUB 200-2011 dataset, with EfficientNetB0 achieving the highest accuracy (99.51%) and IoU score (0.43). Despite high accuracy, models such as InceptionResNetV2 and DenseNet201 showed lower IoU scores, raising trustworthiness concerns. This study underscores the importance of XAI in ensuring the transparency and reliability of Artificial Intelligence (AI) models in ecological applications. • Conducted the first known study on bird species image classification using XAI, offering unique insights into the domain of ecology. • Proposed a new approach to evaluate the reliability and trustworthiness of deep learning models specifically for bird species image classification. • Enhanced the transparency and interpretability of model predictions, enabling users to understand the reasons behind classification results. • Utilized LIME to generate feature heatmap images to identify key features influencing model decisions. • Considered six pre-trained models to leverage their learned features for accurate bird image classification.