WEEDY RICE DSS
An applied research platform for agricultural segmentation and decision support.
Weedy Rice DSS combines a Python backend, a React/Vite UI, and segmentation-training workflows into one domain system. Its docs emphasize layer separation so model work, data flow, and interface changes can evolve independently while staying aligned. It is a practical research stack rather than a generic app shell.
Overview
The Weedy Rice Decision Support System (DSS) is an applied research platform focused on AI-driven agricultural image segmentation. It bridges the gap between heavy machine learning workflows and an accessible interface.
Architecture
A dual-layer stack combining a Python 3.11 backend for AI processing and model training with a React (Vite) frontend. It utilizes specific Docker bind mounts for accessing massive external imagery datasets locally, while maintaining an AWS Lambda Hybrid deployment pattern for its production infrastructure. It keeps UI state and model logic cleanly separated.
Problem It Solves
Processing, segmenting, and analyzing specialized agricultural imagery requires dedicated, heavy infrastructure that isn't suited for standard web APIs. This system provides the dedicated environment to train models and visualize the results.
Current State
MVP operational. The separation between the backend AI stack, the React UI, and the segmentation-training workflows is fully aligned and documented as a cohesive stack.
MILESTONES
Segmentation Stack Aligned
Backend, React UI, and segmentation-training layers documented as a cohesive research stack.
Thesis Direction v3.5 Locked
Post Phase-1 ablation direction locked: Dataset 2 primary 200-image binary, Mendeley pretrain, U-Net + DeepLabV3+ comparison, multi-tier annotation framework, and 4-week proposal timeline committed to repo.
SAM2 Annotation Infrastructure Complete
Full annotation chain shipped: sam2-backend HTTP API with base64 transport, annotator dockerized as dual Tailscale instances (peem/ra), LRU embedding cache, mask refinement controls (threshold + brightness filter), binary-category frontend remediation.