# Action Plan: Fix False Imminent Triggers (CI-Only + Confidence Intervals)

**Problem**: Noise/clouds cause false imminent triggers (model learns on noisy data)  
**Solution**: Better smoothing + uncertainty quantification to filter noise  
**Effort**: 4-5 hours implementation + 30 min training  

---

## Root Cause Analysis

Your graph shows: Smooth blue LOESS curve (real field state) vs. Jagged red line (noisy measurements)

**Current model problem:**
- Feature engineering uses raw noisy data
- Model learns "this noise pattern = harvest signal"
- When clouds/sensor errors create similar noise → False trigger

**Fix:**
1. Derive features from SMOOTHED curve only (remove noise at source)
2. Add "stability" feature (harvest = smooth decline, noise = jagged)
3. Add "decline rate" feature (harvest = consistent slope)
4. Add confidence intervals to identify uncertain predictions (= noise)

---

## Step-by-Step Implementation

### STEP 1: Update Feature Engineering (Section 5)
**What**: Replace 7 features with new CI-only features  
**How**: Use 21-day median + 7-day mean smoothing as foundation  
**Features**:
- Smoothed CI (from smooth curve, not raw)
- 7d velocity (from smooth curve)
- 7d acceleration (from smooth curve)
- 21d MA (very long-term trend)
- 21d velocity (slow changes only)
- **Decline rate** (NEW - slope of smooth curve, harvest = negative slope)
- **Stability** (NEW - smoothness metric, harvest = high stability)

**Code**: See `CI_ONLY_IMPROVEMENTS.md` → "Solution 1: Aggressive Smoothing"

**Expected result**: Model learns real patterns, not noise

### STEP 2: Add Monte Carlo Dropout (Confidence Intervals)
**What**: Run prediction 30 times with dropout ON, get uncertainty  
**Why**: High uncertainty = model unsure = probably noise  
**How**: Keep dropout active during inference, ensemble predictions

**Code**: See `CI_ONLY_IMPROVEMENTS.md` → "Solution 2: Add Confidence Intervals"

**Expected result**: Each prediction has mean + 95% CI

### STEP 3: Filter by Uncertainty
**What**: Only alert on HIGH probability + LOW uncertainty  
**Why**: Filters out noise-driven false positives  
**How**: Use threshold like `prob > 0.5 AND std < 0.10`

**Code**: See `CI_ONLY_IMPROVEMENTS.md` → "Solution 3: Use Uncertainty to Filter"

**Expected result**: False positive rate drops 30-50% without losing real harvests

### STEP 4: Retrain & Evaluate
**Runtime**: ~30 minutes on GPU (standard)

---

## What NOT to Do (Yet)

❌ **Don't add temperature data yet**  
❌ **Don't add rainfall data yet**  
❌ **Don't add soil moisture yet**  

Reason: Fix CI-only first. Once this works perfectly, external data will add value. Adding too many features now would confuse the problem.

---

## Expected Performance

| Metric | Before | After | Change |
|--------|--------|-------|--------|
| Imminent AUC | 0.8793 | 0.90-0.92 | +1-3% |
| False positive rate | ~15% | ~3-5% | -70% |
| **Recall** (catches real harvests) | 100% | 85-90% | -10-15% |

**Trade-off**: You lose 10-15% of early warnings to filter 70% of false positives. Acceptable trade.

---

## Testing Strategy

After implementation, test on same 6 sequences you've been using:

```
For each sequence:
  1. Plot imminent probability + confidence bands
  2. Plot uncertainty over time
  3. Verify:
     - Cloud dips show HIGH uncertainty
     - Real harvest shows LOW uncertainty
     - False triggers disappeared
```

---

## File Location

All documentation is now in:  
`python_app/harvest_detection_experiments/`

Main files:
- `CI_ONLY_IMPROVEMENTS.md` ← Implementation details + code
- `README_EVALUATION.md` ← Navigation guide
- Other `.md` files for reference

---

## Timeline

- **Day 1**: Read CI_ONLY_IMPROVEMENTS.md, plan implementation
- **Day 2-3**: Implement Step 1 (new features)
- **Day 4**: Implement Steps 2-3 (Monte Carlo + filtering)
- **Day 5**: Retrain + test
- **Day 5+**: Evaluate results, iterate

Total: **3-4 focused days** of work

---

## Success Criteria

✅ Model trained without errors  
✅ Uncertainty bands visible in plots  
✅ Cloud dips show high uncertainty  
✅ Real harvest shows low uncertainty  
✅ False positive rate < 5%  
✅ Recall > 85% (still catches most real harvests)