Acknowledgment

I would like to express my sincere gratitude to all those who supported and guided me throughout the development of the FeatherFind: AI-Powered Bird Finder Chatbot.

First, I am deeply thankful to our academic mentors and faculty members at the Informatics Institute of Technology, whose knowledge and encouragement provided the foundation for this project. A special thanks goes to our project advisor. I also wish to extend my appreciation to my dedicated team members, whose collaboration, technical skills, and commitment played a vital role in the success of this team project. Your contributions made the integration of image recognition, NLP-based query processing, and dynamic bird description generation both efficient and innovative.

Last but not least, I’m grateful for the opportunity to work on a solution that not only applies cutting-edge technologies such as RASA, BERT, and GPT-2, but also contributes to a real-world use case in biodiversity and education.

Introduction

This project on "FeatherFind" defines the purpose, scope, and feature set of a system designed for bird identification and access to information. It follows with functional requirements on features such as image-based bird identification, Range prediction, Bird finder using user prompt and Bird information generation with true chatbots

The proposed Bird Identification and Exploration platform targets both amateur and professional ornithologists by addressing key issues in studies about birds in Sri Lanka. Inculcating higher capabilities of machine learning models would make the application innovate bird identification and exploration for users, making the overall experience accessible and educational, and deepening the appreciation for birdlife.

The chatbot leverages state-of-the-art technologies such as RASA for conversational AI, BERT for intent classification, and GPT-2 for dynamically generating descriptive bird information. It also incorporates an image recognition module to identify bird species from user-uploaded images and a range prediction system to provide location-based bird sightings.

FeatherFind aims to enhance birdwatching experiences, support biodiversity awareness, and demonstrate the practical application of AI in wildlife education and conservation. This project reflects not only technical capability but also a commitment to building meaningful AI solutions that interact seamlessly with users in real-time.

methodology

Research methodology - The research design is based on positivism; it dwells upon the study of things that can be measured and observed. That works in your project, where you will use AI to identify species of birds and track their migration. The data so collected will give clear factual results, which is in keeping with the concept of positivism itself-scientific methods of comprehending the world.

Development methodology - Flexible approach is required due to the dynamic nature of bird migration patterns, species identification, and the integration of weather and historical data. Considering the above requirements, the Agile development strategy will be the most optimal. Iterative development cycles of the Agile model allow the development team to adapt to new data sources and user feedback. This methodology ensures rapid adaptations to enhance the provided features and migration prediction capabilities. Adopting this method allows each development sprint to focus on a specific feature.

Experiments

Intent Classification: Fine-tuned BERT to classify user queries accurately, achieving over 90% accuracy in detecting intents like bird info, image uploads, and location queries.

Dynamic Text Generation: Used a GPT-2 model to generate natural bird descriptions, with fine-tuning improving coherence and relevance significantly.

Image Classification: Trained a ResNet-based model for species recognition, reaching top-1 accuracy on a diverse bird dataset.

Information Retrieval: Implemented FAISS and fuzzy matching to provide precise, fast retrieval of bird data with high relevance scores.

User Testing: Conducted real-user interaction tests to improve flow, usability, and error handling, resulting in positive feedback and iterative improvements.

Results

Bird Range Prediction - The Bird Range Prediction component consists of three predictive models:

1. Bird Presence Prediction Model - Determines whether a specific bird species is likely to be observed at a given location and time. (Training: 92.10% Testing: 88.14% )
2. Bird Location Prediction Model - Predicts the most probable location where a bird species may be found based on historical observations. (Training: 90.36% Testing: 89.09% )
3. Bird Time Prediction Model - Estimates the most likely time of the year and day a bird species may be observed. (Mean Absolute Error (MAE): 0.7381 Root Mean Squared Error (RMSE): 1.0768 R² Score: 0.9375 )

Image based Recognition - In comparison to the performance of different models in bird classification and detection, MobileNetV2 is always the best. For classifying birds, MobileNetV2 achieved a mean training
accuracy of 94.14% and test accuracy of 95.25%. In the detection of birds, it reached even higher with a mean training accuracy of 97.45% and test accuracy of 95.75%. Comparatively, ResNet50 performed worse with an average training accuracy of 46.89% and test accuracy of 53.17%. Although AlexNet was superior to ResNet50 in average training accuracy of 79.31% and test accuracy of 91.67%, MobileNetV2 performed better both in training and test accuracy and thus was the best for this task.

Bird Info Generator - GPT-2 was easier to fine-tune while still delivering high-quality, human-like responses. With a perplexity of 1.61, a train loss of 1.6029, and a validation loss of 0.4759 after 7 epochs, GPT-2
effectively captures context, coherence, and factual accuracy, making it the best choice for dynamic and informative bird descriptions.

Conclusion

The FeatherFind project successfully demonstrates the potential of integrating AI technologies to create an interactive, intelligent system for bird identification and information retrieval. By combining natural language processing, image classification, and dynamic text generation, the chatbot offers a user-friendly platform that bridges the gap between technology and nature education.

Through continuous experimentation and fine-tuning, FeatherFind achieved high performance in intent recognition, descriptive generation, and visual classification. The project not only deepened technical expertise in tools like RASA, BERT, GPT-2, and CNNs, but also showcased the value of collaboration, problem-solving, and real-world AI application.

Overall, FeatherFind reflects a practical and innovative use of AI to support biodiversity awareness, making it a meaningful step in the journey toward becoming a skilled AI and Data Science professional.