| --- |
| license: gemma |
| library_name: big_vision |
| base_model: google/paligemma-3b-pt-224 |
| tags: |
| - agricultural-ai |
| - vlm |
| - paligemma |
| - crop-disease-detection |
| - rwanda |
| - d2p-agri-platform |
| metrics: |
| - accuracy |
| pipeline_tag: image-to-text |
| --- |
| |
| # IsokoLink: VLM Crop Disease Diagnostic Module |
|
|
| *IsokoLink* is a fine-tuned vision-language model (VLM) specialized for identifying and localizing diseases in staple crops (*Maize and Beans) within smallholder farming systems in **Rwanda*. |
| |
| Developed as the core AI engine for the *Digital-to-Physical (D2P) Agri-Platform*, this model transforms visual field data into verifiable quality scores that anchor financial transactions and fair trade for farmers. |
| |
| ## Model Details |
| |
| - *Model Type:* Vision-Language Model (VLM) |
| - *Base Model:* [Google PaliGemma 3B (224px)](https://huggingface.co/google/paligemma-3b-pt-224) |
| - *Language(s):* English |
| - *Fine-tuning Method:* Parameter-Efficient Fine-Tuning (PEFT) targeting attention layers only. |
| |
| ## Intended Use |
| |
| ### Direct Use |
| IsokoLink is designed to be used by farmers,hub operators and extension officers at *aggregation hubs* (e.g., Nyabugogo) to: |
| 1. Identify specific crop diseases from mobile imagery. |
| 2. Localize disease clusters using bounding boxes. |
| 3. Generate an objective "Quality Certificate" for crop batches. |
| |
| ## Training Data & Methodology |
| |
| ### Dataset |
| The model was trained on a hybrid dataset: |
| • *PlantVillage:* Baseline labels for common agricultural diseases. |
| • *Musanze Field Dataset:* Locally collected images from smallholder farms in Rwanda’s Northern Province to account for regional environmental variations and lighting. |
| |
| ### Target Classes |
| • maize streak virus |
| • maize stalk borer |
| • bean leaf disease |
| • healthy |
| |
| ### Training Hyperparameters |
| • *Batch Size:* 8 |
| • *Learning Rate:* 0.005 (Cosine schedule with 10% warmup) |
| • *Training Steps:* 64 |
| • *Sequence Length:* 128 tokens |
| • *Hardware:* NVIDIA T4 GPU |
| |
| ## Performance & Limitations |
| |
| ### Accuracy Target |
| The model is optimized to achieve a classification accuracy of *≥ 85%* for the target crops. |
| |
| ### Limitations |
| • *Resolution:* As a 224px model, it may struggle with very small, microscopic symptoms (e.g., early-stage fungal spores) without close-up photography. |
| • *Scope:* Currently limited to Maize and Beans; performance on other Rwandan staples like Irish potatoes or Cassava is not guaranteed. |
| • *Hallucination:* Like all VLMs, the model may occasionally generate bounding boxes for non-existent symptoms in extremely noisy backgrounds. |
| |
| ## Ethical Considerations |
| IsokoLink was built to address the *structural trust deficit* between farmers and middlemen. By providing objective evidence of crop health, the model aims to prevent under-pricing of healthy yields. Users are advised that AI diagnostics should supplement, not entirely replace, the expertise of agricultural agronomists in critical food security decisions. |
| |
| ## How to Get Started |
| To run inference in JAX/big_vision, load the paligemma-3b-pt-224.f16.npz weights and use the prompt: |
| detect maize streak virus ; maize stalk borer ; bean leaf disease ; healthy |
| |
| --- |
| *Project Context:* This module is part of the D2P Agri-Platform, aiming to modernize Rwandan agriculture through transparent, AI-backed market linkages. |
