ADR002 - ModelTrainingAndThirdPartyAnalysis.md

Title: Model Training And Third Party Analysis

Status:

Proposed

Context:

Challenge: Conservation efforts often require efficient and accurate species identification, which can be challenging and time-consuming.

Opportunity: Leveraging AI for species identification through camera traps can significantly enhance the speed and accuracy of data analysis.

Rationale:

1. AI for Conservation: Integrating AI models enables automated species identification from camera trap footage, reducing manual effort and providing real-time insights.
2. Flexible Deployment: Offering both Raspberry Pi and AWS Snowcone as hardware options allows users to choose based on their specific needs, considering factors like budget and computational requirements.
3. Collaboration and Data Exchange: Integration with labeling platforms, citizen science initiatives, and global biodiversity data exchange platforms enhances the collaborative nature of conservation efforts.

Decision:

1. Third-Party Services: Selection of Roboflow, Edge Impulse, and TensorFlow Lite for data preprocessing and model training to leverage proven, efficient solutions.
2. Local Execution: Allowing users to trigger model execution on local CPUs provides flexibility and autonomy in deploying the system.
3. Dual Hardware Variants: Choosing Raspberry Pi and AWS Snowcone as hardware variants provides a spectrum of deployment options to cater to diverse user needs.
4. Integration with Platforms: Collaborating with Wildlife Insights, TrapTagger, Trapper, iNaturalist, Cantrap DP, and GBIF ensures a comprehensive and interconnected ecosystem.

Implementation:

1. Data Preprocessing and Training: Utilization of third-party services involves integrating APIs and SDKs for seamless data processing and model training.
2. Prediction Engine: Deployment of trained models on selected hardware variants (Raspberry Pi or AWS Snowcone) for efficient and accurate species predictions.
3. Model Repository: Implementation of a secure central repository for storing and managing model checkpoint files, ensuring accessibility and security.
4. User Interaction: Designing a user-friendly interface for triggering model execution and facilitating interactions with camera trap labeling platforms.
5. Security Measures: Implementation of encryption and access controls for safeguarding model checkpoint files in the repository.

Consequences:

1. Increased Efficiency: The AI-based system significantly reduces the time and effort required for species identification, enabling more efficient conservation monitoring.
2. User Empowerment: Offering a choice between Raspberry Pi and AWS Snowcone empowers users to make decisions based on their specific requirements and constraints.
3. Collaborative Impact: Integration with various platforms fosters collaboration with experts, citizen scientists, and global biodiversity initiatives, amplifying the impact of conservation efforts.
4. Data Exchange Standardization: Implementing data exchange standards like Cantrap DP enhances interoperability and ensures compatibility with broader conservation data ecosystems.