Technical Summary: AwakeForest

Problem Statement

Forest imagery analysis is a complex domain requiring tightly coupled vision tasks, including object detection, spatial delineation, and quantitative analysis. These tasks must operate under significant variation in geographic regions, sensor types, and acquisition conditions. A critical gap exists in the current tooling landscape: practitioners lack a unified platform that is simultaneously geospatial-native, cloud-optimized, and ML-integrated for end-to-end workflows.

Existing solutions fall into two categories, neither of which fully addresses the needs of large-scale forest management:

1. Generic Annotation Tools (e.g., CVAT, Label Studio, AnyLabeling): These support AI-assisted labeling but treat geospatial coordinates and massive orthomosaics as secondary. They are not designed for the high-resolution, spatially extensive imagery (ranging from gigabytes to hundreds of gigabytes) typical of ecological monitoring.
2. Geographic Information Systems (GIS) (e.g., QGIS): These are geospatial-native and excellent for spatial indexing and management but lack integrated, interactive machine learning annotation workflows with plug-and-play model inference capabilities.

Consequently, practitioners often face high costs in creating labeled datasets and struggle with domain shifts that limit the reuse of existing models and annotations, forcing repetitive adaptation and dataset rebuilding with fragmented tooling.

Methodology and System Architecture

AwakeForest is presented as an interactive, end-to-end platform designed to bridge this gap. It integrates model-assisted inference, automatic annotation, and human-in-the-loop refinement within a single, consistent geospatial context.

High-Level Architecture

The system employs a modular, service-oriented architecture (Figure 1) comprising four distinct tiers:

• Client Tier (Next.js): Provides the user interface for project management, dataset registration, and model selection. It utilizes Leaflet for interactive geospatial map visualization.
• Service Tier (FastAPI): Acts as the central coordinator for user authentication, project organization, and the orchestration of inference requests. It manages the logic and tiling services.
• Model Inference Tier: Consists of independent services registered by users via API endpoints. These services are task-specific (e.g., detection, segmentation) and can include models like YOLO (for detection) and SAM/SAM2 (for segmentation and prompt-based labeling).
• Data Tier (MinIO/Supabase): Stores large datasets using Cloud Optimized GeoTIFFs (COGs) in MinIO (S3-compatible). Data persistence and spatial indexing are handled via Supabase (PostgreSQL with PostGIS extension).

Core Workflow and Features

The platform operates through a project-based structure linking datasets with registered models (Figure 2):

1. Dual-View Interface:
   • Global View: Displays the entire orthomosaic, allowing users to pan, zoom, and view annotation overlays for real-time progress tracking across large forest scenes.
   • Patch-Level View: Divides the dataset into fixed-size image patches. Users can select models to generate predictions on the current patch, manually annotate missed objects, assign labels, and verify quality.
2. Unified Geospatial Representation: To ensure consistency, AwakeForest decouples annotations from specific image tiles. Intermediate results are stored with absolute spatial attributes and standardized geometry formats based on geospatial coordinates. This allows for stable indexing, cross-model comparison, and incremental updates regardless of resolution or imagery source.
3. Plug-and-Play Model Integration: Users can register external FastAPI endpoints for inference. The system supports prompt-based segmentation (using bounding boxes or text prompts like "palm tree") to rapidly generate instance masks, leveraging the few-shot capabilities of models like SAM3.
4. Analysis-Ready Outputs: The system produces outputs that can be directly used for downstream analysis. Verified results are promoted to final annotations, while intermediate predictions remain traceable. Outputs include confidence-thresholded prediction densities and spatial summaries (e.g., kernel density estimation of instance locations).

Key Contributions

• Unified Workflow: AwakeForest provides the first integrated system that combines geospatial-first data management, cloud-optimized orthomosaic access, and interactive ML annotation.
• Scalability: The platform is engineered to handle large orthomosaics (several GB to hundreds of GB) through on-demand tile streaming via TiTiler, eliminating the need for preprocessing large files for interactive exploration.
• Human-in-the-Loop Refinement: It facilitates a seamless transition from automated inference to human verification, allowing for iterative model updates and dataset refinement within a cohesive spatial context.
• Comparative Advantage: As shown in Table 1, AwakeForest is the only tool among major competitors that simultaneously supports geospatial-native operations, spatial indexing, cloud optimization, and AI-assisted labeling.

Results and Demonstration

The system was demonstrated on the PALMS dataset, which consists of high-resolution UAV-derived orthomosaics covering large, spatially continuous forest regions with dense, heterogeneous canopy structures.

• Workflow Execution: Users loaded PALMS orthomosaics and interactively explored them at multiple scales. Pretrained detection models generated initial proposals for palm instances (bounding boxes and crown centers).
• Refinement: Users invoked SAM-based segmentation using these boxes as prompts to obtain high-resolution instance masks without manual drawing. Text prompts (e.g., "palm") were also used to retrieve candidate instances and generate masks.
• Output Generation: The demonstration showed the ability to filter predictions by confidence thresholds and visualize prediction densities via kernel density estimation. Finalized outputs included spatial summaries such as palm center distributions and instance counts over selected regions.

Significance and Claims

The paper positions AwakeForest as a practical solution for forest management and ecological monitoring, specifically addressing the need for end-to-end workflows that handle the scale and complexity of modern forest imagery.

• Practical Utility: The platform enables users to move seamlessly from large imagery exploration to verified, analysis-ready outputs, reducing the manual burden of labeling and the friction of switching between disparate tools.
• Modesty in Claims: The authors explicitly state that the work focuses on demonstrating the platform's capabilities in a representative scenario (RGB forest imagery) rather than providing quantitative benchmarks on predictive accuracy or annotation efficiency.
• Future Directions: The paper acknowledges that the quality of results depends on the suitability of the employed pretrained models, particularly under domain shifts. Future work is identified as extending support to additional sensing modalities (multispectral, LiDAR, temporal data) and enabling collaborative, multi-user workflows.

In summary, AwakeForest illustrates how an integrated, end-to-end workflow can support detection, segmentation, spatial summarization, and iterative refinement in a unified environment, filling a critical gap in the tooling available for large-scale forest analysis.