# Phase 2: Image Dataset Acquisition - Implementation Plan

## Overview

**Goal:** Gather labeled plant images matching our 2,064-plant knowledge base from Phase 1.

**Target Deliverable:** Labeled image dataset with 50,000-200,000 images across target plant classes, split into training (70%), validation (15%), and test (15%) sets.

---

## Prerequisites

- [x] Phase 1 complete: `data/final_knowledge_base.json` (2,064 plants)
- [x] SQLite database: `knowledge_base/plants.db`
- [ ] Python environment with required packages
- [ ] API keys for image sources (iNaturalist, Flickr, etc.)
- [ ] Storage space: ~50-100GB for raw images

---

## Task Breakdown

### Task 2.1: Research Public Plant Image Datasets

**Objective:** Evaluate available datasets for compatibility with our plant list.

**Actions:**
1. Research and document each dataset:
   - **PlantCLEF** - Download links, species coverage, image format, license
   - **iNaturalist** - API access, species coverage, observation quality filters
   - **PlantNet (Pl@ntNet)** - API documentation, rate limits, attribution requirements
   - **Oxford Flowers 102** - Direct download, category mapping
   - **Wikimedia Commons** - API access for botanical images

2. Create `scripts/phase2/research_datasets.py` to:
   - Query each API for available species counts
   - Document download procedures and authentication
   - Estimate total available images per source

**Output:** `output/dataset_research_report.json`

**Validation:**
- [ ] Report contains at least 4 dataset sources
- [ ] Each source has documented: URL, license, estimated image count, access method

---

### Task 2.2: Cross-Reference Datasets with Plant List

**Objective:** Identify which plants from our knowledge base have images in public datasets.

**Actions:**
1. Create `scripts/phase2/cross_reference_plants.py` to:
   - Load plant list from `data/final_knowledge_base.json`
   - Query each dataset API for matching scientific names
   - Handle synonyms using `data/synonyms.json`
   - Track exact matches, synonym matches, and genus-level matches

2. Generate coverage matrix: plants × datasets

**Output:**
- `output/dataset_coverage_matrix.json` - Per-plant availability
- `output/cross_reference_report.json` - Summary statistics

**Validation:**
- [ ] Coverage matrix includes all 2,064 plants
- [ ] Report shows percentage coverage per dataset
- [ ] Identified total unique plants with at least one dataset match

---

### Task 2.3: Download and Organize Images

**Objective:** Download images from selected sources and organize by species.

**Actions:**
1. Create directory structure:
   ```
   datasets/
   ├── raw/
   │   ├── inaturalist/
   │   ├── plantclef/
   │   ├── wikimedia/
   │   └── flickr/
   └── organized/
       └── {scientific_name}/
           ├── img_001.jpg
           └── metadata.json
   ```

2. Create `scripts/phase2/download_inaturalist.py`:
   - Use iNaturalist API with research-grade filter
   - Download max 500 images per species
   - Include metadata (observer, date, location, license)
   - Handle rate limiting with exponential backoff

3. Create `scripts/phase2/download_plantclef.py`:
   - Download from PlantCLEF challenge archives
   - Extract and organize by species

4. Create `scripts/phase2/download_wikimedia.py`:
   - Query Wikimedia Commons API for botanical images
   - Filter by license (CC-BY, CC-BY-SA, public domain)

5. Create `scripts/phase2/organize_images.py`:
   - Consolidate images from all sources
   - Rename with consistent naming: `{plant_id}_{source}_{index}.jpg`
   - Generate per-species `metadata.json`

**Output:**
- `datasets/organized/` - Organized image directory
- `output/download_progress.json` - Download status per species

**Validation:**
- [ ] Images organized in consistent directory structure
- [ ] Each image has source attribution in metadata
- [ ] Progress tracking shows download status for all plants

---

### Task 2.4: Establish Minimum Image Count per Class

**Objective:** Define and track image count thresholds.

**Actions:**
1. Create `scripts/phase2/count_images.py` to:
   - Count images per species in `datasets/organized/`
   - Classify plants into coverage tiers:
     - **Excellent:** 200+ images
     - **Good:** 100-199 images (target minimum)
     - **Marginal:** 50-99 images
     - **Insufficient:** 10-49 images
     - **Critical:** <10 images

2. Generate coverage report with distribution histogram

**Output:**
- `output/image_count_report.json`
- `output/coverage_histogram.png`

**Validation:**
- [ ] Target: At least 60% of plants have 100+ images
- [ ] Report identifies all plants below minimum threshold
- [ ] Total image count within target range (50K-200K)

---

### Task 2.5: Identify Gap Plants

**Objective:** Find plants needing supplementary images.

**Actions:**
1. Create `scripts/phase2/identify_gaps.py` to:
   - List plants with <100 images
   - Prioritize gaps by:
     - Plant popularity/commonality
     - Category importance (user-facing plants first)
     - Ease of sourcing (common names available)

2. Generate prioritized gap list with recommended sources

**Output:**
- `output/gap_plants.json` - Prioritized list with current counts
- `output/gap_analysis_report.md` - Human-readable analysis

**Validation:**
- [ ] Gap list includes all plants under 100-image threshold
- [ ] Each gap plant has recommended supplementary sources
- [ ] Priority scores assigned based on criteria

---

### Task 2.6: Source Supplementary Images

**Objective:** Fill gaps using additional image sources.

**Actions:**
1. Create `scripts/phase2/download_flickr.py`:
   - Use Flickr API with botanical/plant tags
   - Filter by license (CC-BY, CC-BY-SA)
   - Search by scientific name AND common names

2. Create `scripts/phase2/download_google_images.py`:
   - Use Google Custom Search API (paid tier)
   - Apply strict botanical filters
   - Download only high-resolution images

3. Create `scripts/phase2/manual_curation_list.py`:
   - Generate list of gap plants requiring manual sourcing
   - Create curation checklist for human review

4. Update `organize_images.py` to incorporate supplementary sources

**Output:**
- Updated `datasets/organized/` with supplementary images
- `output/supplementary_download_report.json`
- `output/manual_curation_checklist.md` (if needed)

**Validation:**
- [ ] Gap plants have improved coverage
- [ ] All supplementary images have proper licensing
- [ ] Re-run Task 2.4 shows improved coverage metrics

---

### Task 2.7: Verify Image Quality and Labels

**Objective:** Remove mislabeled and low-quality images.

**Actions:**
1. Create `scripts/phase2/quality_filter.py` to:
   - Detect corrupt/truncated images
   - Filter by minimum resolution (224x224 minimum)
   - Detect duplicates using perceptual hashing (pHash)
   - Flag images with text overlays/watermarks

2. Create `scripts/phase2/label_verification.py` to:
   - Use pretrained plant classifier for sanity check
   - Flag images where model confidence is very low
   - Generate review queue for human verification

3. Create `scripts/phase2/human_review_tool.py`:
   - Simple CLI tool for reviewing flagged images
   - Accept/reject/relabel options
   - Track reviewer decisions

**Output:**
- `datasets/verified/` - Cleaned image directory
- `output/quality_report.json` - Filtering statistics
- `output/removed_images.json` - Log of removed images with reasons

**Validation:**
- [ ] All images pass minimum resolution check
- [ ] No duplicate images (within 95% perceptual similarity)
- [ ] Flagged images reviewed and resolved
- [ ] Removal rate documented (<20% expected)

---

### Task 2.8: Split Dataset

**Objective:** Create reproducible train/validation/test splits.

**Actions:**
1. Create `scripts/phase2/split_dataset.py` to:
   - Stratified split maintaining class distribution
   - 70% training, 15% validation, 15% test
   - Ensure no data leakage (same plant photo in multiple splits)
   - Handle class imbalance (minimum samples per class in each split)

2. Create manifest files:
   ```
   datasets/
   ├── train/
   │   ├── images/
   │   └── manifest.csv  (path, label, scientific_name, plant_id)
   ├── val/
   │   ├── images/
   │   └── manifest.csv
   └── test/
       ├── images/
       └── manifest.csv
   ```

3. Generate split statistics report

**Output:**
- `datasets/train/`, `datasets/val/`, `datasets/test/` directories
- `output/split_statistics.json`
- `output/class_distribution.png` (per-split histogram)

**Validation:**
- [ ] Split ratios within 1% of target (70/15/15)
- [ ] Each class has minimum 5 samples in val and test sets
- [ ] No image appears in multiple splits
- [ ] Manifest files are complete and valid

---

## End-Phase Validation Checklist

Run `scripts/phase2/validate_phase2.py` to verify:

| # | Validation Criterion | Target | Pass/Fail |
|---|---------------------|--------|-----------|
| 1 | Total image count | 50,000 - 200,000 | [ ] |
| 2 | Plant coverage | ≥80% of 2,064 plants have images | [ ] |
| 3 | Minimum images per included plant | ≥50 images (relaxed from 100 for rare plants) | [ ] |
| 4 | Image quality | 100% pass resolution check | [ ] |
| 5 | No duplicates | 0 exact duplicates, <1% near-duplicates | [ ] |
| 6 | License compliance | 100% images have documented license | [ ] |
| 7 | Train/val/test split exists | All three directories with manifests | [ ] |
| 8 | Split ratio accuracy | Within 1% of 70/15/15 | [ ] |
| 9 | Stratification verified | Chi-square test p > 0.05 | [ ] |
| 10 | Metadata completeness | 100% images have source + license | [ ] |

**Phase 2 Complete When:** All 10 validation criteria pass.

---

## Scripts Summary

| Script | Task | Input | Output |
|--------|------|-------|--------|
| `research_datasets.py` | 2.1 | None | `dataset_research_report.json` |
| `cross_reference_plants.py` | 2.2 | Knowledge base | `cross_reference_report.json` |
| `download_inaturalist.py` | 2.3 | Plant list | Images + metadata |
| `download_plantclef.py` | 2.3 | Plant list | Images + metadata |
| `download_wikimedia.py` | 2.3 | Plant list | Images + metadata |
| `organize_images.py` | 2.3 | Raw images | `datasets/organized/` |
| `count_images.py` | 2.4 | Organized images | `image_count_report.json` |
| `identify_gaps.py` | 2.5 | Image counts | `gap_plants.json` |
| `download_flickr.py` | 2.6 | Gap plants | Supplementary images |
| `quality_filter.py` | 2.7 | All images | `datasets/verified/` |
| `label_verification.py` | 2.7 | Verified images | Review queue |
| `split_dataset.py` | 2.8 | Verified images | Train/val/test splits |
| `validate_phase2.py` | Final | All outputs | Validation report |

---

## Dependencies

```
# requirements-phase2.txt
requests>=2.28.0
Pillow>=9.0.0
imagehash>=4.3.0
pandas>=1.5.0
tqdm>=4.64.0
python-dotenv>=1.0.0
matplotlib>=3.6.0
scipy>=1.9.0
```

---

## Environment Variables

```
# .env.phase2
INATURALIST_APP_ID=your_app_id
INATURALIST_APP_SECRET=your_secret
FLICKR_API_KEY=your_key
FLICKR_API_SECRET=your_secret
GOOGLE_CSE_API_KEY=your_key
GOOGLE_CSE_CX=your_cx
```

---

## Estimated Timeline

| Task | Effort | Notes |
|------|--------|-------|
| 2.1 Research | 1 day | Documentation and API testing |
| 2.2 Cross-reference | 1 day | API queries, matching logic |
| 2.3 Download | 3-5 days | Rate-limited by APIs |
| 2.4 Count | 0.5 day | Quick analysis |
| 2.5 Gap analysis | 0.5 day | Based on counts |
| 2.6 Supplementary | 2-3 days | Depends on gap size |
| 2.7 Quality verification | 2 days | Includes manual review |
| 2.8 Split | 0.5 day | Automated |
| Validation | 0.5 day | Final checks |

---

## Risk Mitigation

| Risk | Mitigation |
|------|------------|
| API rate limits | Implement backoff, cache responses, spread over time |
| Low coverage for rare plants | Accept lower threshold (50 images) with augmentation in Phase 3 |
| License issues | Track all sources, prefer CC-licensed content |
| Storage limits | Implement progressive download, compress as needed |
| Label noise | Use pretrained model for sanity check, human review queue |

---

## Next Steps After Phase 2

1. Review `output/image_count_report.json` for Phase 3 augmentation priorities
2. Ensure `datasets/train/manifest.csv` format is compatible with training framework
3. Document any plants excluded due to insufficient images