# RESAMPLING.md
*Methodological scaffold for epistemically grounded resampling and resolution changes*
## 1. Purpose
Resampling is not a formatting change — it is a **statement about what information is preserved when the spatial lattice is reconstructed.**
If a DEM encodes real-world surface energy, slope discontinuity, microtopography, or drainage structure, resampling directly affects *physical truth*, not just visual appearance.
Most silent GIS errors occur here because interpolation is assumed to be neutral.
It is not.
Each resampling method encodes a **theory of what is “important” in the data.**
This methodology ensures that resampling reflects the **intended scientific meaning** of the raster.
---
## 2. When Resampling Methodology is Required
Epistemic escalation is required when:
* the raster encodes *physical signal* (terrain/hydrology, hazards, geophysics),
* slope/flow/gradient will be derived downstream,
* values represent measurement rather than classification,
* spatial frequency structure matters (fine-scale terrain / headwaters / breaks),
* up/down-sampling alters interpretive resolution.
This corresponds to the **Resolution / Resampling epistemic risk class** in `EPISTEMIC_RISK_CLASSES.md` and follows the governance flow in `IMPLEMENTATION_PLAN.md`.
---
## 3. Epistemic Prompting Questions
Before selecting any resampling method, the agent should surface:
1. **What property of the raster must be preserved?**
(classification fidelity, terrain realism, gradient continuity, statistical representativeness, etc.)
2. **Does the raster encode *physical measurement* or *categorical abstraction*?**
3. **Is high-frequency terrain signal important?**
(headwaters, gullies, local maxima/minima → hydrology sensitive)
4. **Are we upsampling or downsampling (and is this reversible)?**
5. **Will smoothing alter scientific interpretation downstream?**
6. **Under what conditions should a future revision occur?**
(e.g., realigned DEM, corrected LiDAR base, multi-scale analysis)
---
## 4. Mapping Schema → Resampling Context
| Schema Field | Resampling Interpretation |
|---------------------|-------------------------------------------------------------|
| `domain` | always `resampling_justification` |
| `intent` | what information must survive the spatial transform |
| `assumptions` | physical vs categorical meaning of pixels |
| `candidate_methods` | NOT “nearest/bilinear/cubic” → but the *property preserved* |
| `selected_method` | which preservation type aligns with intent |
| `epistemic_status` | residual uncertainty + when to revisit choice |
---
## 5. Structured Reasoning Workflow
1. Identify whether raster encodes *signal* or *category*
2. Determine which *aspect of truth* must survive (value, gradient, topology, shape)
3. Consider candidate preservation strategies
4. Reject inappropriate resampling types explicitly
5. Select the method aligned to epistemic intent
6. Document uncertainty + revision triggers
---
## 6. Worked Example (DEM → slope → hydrology)
**Scenario:**
A DEM being resampled before slope and flow routing are computed.
```
epistemic_justification:
domain: resampling_justification
intent:
description: "Preserve microtopography and slope fidelity required for accurate drainage direction."
context: "Upsampling DEM before hydrology workflow; terrain signal drives flow routing."
assumptions:
known:
- "DEM contains high-frequency terrain detail relevant to drainage structure."
- "Hydrologic derivatives depend on local slope accuracy."
uncertain:
- "Input DEM may include sensor noise at sub-pixel scale."
dependencies:
- "Smoothing may erase headwater initiation points."
candidate_methods:
considered:
- "classification-preserving (nearest-value semantics)"
- "gradient-preserving (bilinear value continuity)"
- "smoothness-prior (cubic / spline)"
rejected:
- method: "smoothness-prior"
reason: "Introduces artificial terrain continuity → erases micro-basins."
- method: "classification-preserving"
reason: "Appropriate for categorical rasters, not continuous terrain."
selected_method:
name: "gradient-preserving interpolation (bilinear)"
rationale: "Minimizes smoothing while preserving local slope continuity; hydrologically stable."
tradeoffs: "Very small blurring of extremely small discontinuities; acceptable within DEM resolution limits."
epistemic_status:
confidence_level: "medium"
residual_uncertainty_sources:
- "Edge-case microtopography may still be lost if resolution gap is large."
conditions_for_revisit:
- "If later LiDAR integration provides finer source DEM."
- "If drainage enforcement requires explicit sink/breach review after resampling."
```
---
## 7. When Resampling Escalation Is *Not* Required
No methodology justification is needed when:
* data is categorical and transformation is classification-preserving,
* resampling is for map display only,
* resolution is unchanged (copy),
* values are symbolic (e.g., mask/grids, boolean layers),
* output is not used for measurement-derived inference.
---
## 8. Forward Compatibility
This scaffold is intentionally **property-first**, not algorithm-first.
A future model may:
* discover new interpolation families,
* derive adaptive kernels from local geomorphology,
* learn hybrid frequency-aware methods,
…and the schema will still apply, because **the epistemic commitment is to what is preserved, not how.**
This is how the agent remains capable of *scientific evolution.*
---
## Related References
- **Risk Registry**: `EPISTEMIC_RISK_CLASSES.md`
