A closer look at changing terrain
Presidential Slide Explorer brings the lab’s work into geospatial analysis. The project focuses on comparing imagery of the Presidential Range and investigating possible rock and landslide changes over time.
Designed around investigation
- Compare imagery from different dates within a consistent geographic area.
- Review potential changes alongside terrain and image-quality context.
- Distinguish candidate surface changes from clouds, shadows, snow, and seasonal vegetation.
- Keep a human analyst in the loop before interpreting a change as a landslide.
The investigation workspace
The original project brief calls for a map bounded to the Presidential Range, a change heat map, and historical playback using two or three cropped observations.
| View | Research question |
|---|---|
| Bounded range map | Where is the candidate within the study area? |
| Change overlay | Which areas changed between usable observations? |
| Dated crops | Is the apparent change visible and persistent? |
| Quality overlay | Which pixels were excluded, and why? |
| Terrain context | Does the landform support the interpretation? |
| Candidate evidence | What remains unexplained or unverified? |
A change overlay should represent the underlying measurement. It must not imply landslide probability unless a calibrated model and validation support that interpretation.
Rule out look-alikes
The next development pass calls for auditing the actual imagery processor before claiming these checks are implemented.
Clouds, snow, and shadow
Inspect pixel-level quality masks, questionable edges, seasonal differences, and poorly illuminated slopes. A mostly clear scene can still hide the candidate. Seek additional usable observations before interpreting change.
Roads, buildings, and construction
Include the Auto Road, Cog corridor, summit buildings, and other developed areas as reference layers. Proximity should prompt review; excluding every area near infrastructure could hide a real slide.
LiDAR and terrain context
A bare-earth elevation model can add slope, terrain shape, and hillshade context. One survey describes terrain; it does not establish when a change happened. Comparing elevation across dates requires compatible references, alignment, and uncertainty checks.
Keep the evidence inspectable
For each candidate, the planned record should preserve observation dates, imagery provenance, usable coverage, quality exclusions, change measurements, terrain context, and alternative explanations. Two or three crops support playback; additional observations may be needed to assess baseline conditions and persistence.
Separate processing from the website
The intended deployment prepares imagery and candidate outputs on a local processing machine, then uploads finished crops, overlays, and records to the web host. The website should remain available when that machine is offline and display the last successful processing date.
Local GPU analysis was deferred until imagery preparation and a trustworthy baseline are established. A future model should be evaluated alongside that baseline using reviewed examples and look-alikes.
Current stage
The Slide Explorer web app now searches real Sentinel-2 observations, displays aligned dated crops and playback, applies scene-classification masks, and screens for vegetation loss through a configurable NDVI-decrease overlay. Connected changed regions can be exported as bounding-box GeoJSON and saved with review notes.
The comparison uses a fixed 512 × 512 grid over the study area, not native-resolution detection. It excludes invalid, cloud, shadow, snow, and water classes using Sentinel-2 SCL. Seasonal changes and classification errors can still produce false positives. There is no trained landslide classifier, LiDAR layer, or field validation in this release.
Interpretation matters
A visible change is not a confirmed landslide. The explorer is a research tool, not an operational hazard-warning service or a substitute for field verification.