A simple guide to the experiment
A better order for checking landslides.
We added a model to sort the places flagged by Tanvi's rules, so a person can inspect likely landslides first.
See what changedWider ground context helps separate hillside scars from riverbeds. The rings show nearby ground at several scales. Illustration only.
Three steps. A person makes the final check.
-
Rules find suspicious places
Tanvi's method looks for before/after satellite vegetation change and applies slope rules.
-
Our model sorts the list
We added a ranking that puts the candidates most likely to be landslides near the top.
-
A human verifies
A person inspects the candidates to decide which are landslides, starting with the higher-ranked places.
Tanvi's original workflow used rules followed by manual inspection. It did not use a trained machine-learning model.
More slides in the first checks.
The cleaned list contains 2,120 locations, including 1,181 labelled landslides.
In the latest recorded run, we look at the first 424 candidates: 20% of that list.
Both bars use the same 424-check scale and the same candidate list. Random ordering is a calculated expectation, not a measured result from Tanvi's own inspection order.
About 84 real slides per 100 early checks. The model's first 424 candidates contain 354 labelled landslides and 70 labelled non-slides.
That retrieves about 30% of the 1,181 known slides in this list. The 84% figure describes these early checks; it is not overall accuracy or district-wide recall.
Four models,
one ordering.
The headline model is an ensemble: it combines four models' scores to rank each candidate.
- Logistic Regression
- Random Forest
- XGBoost
- LightGBM
Inputs include terrain, before/after satellite information, changes, and summaries of nearby ground at several scales. That wider context helps tell hillside scars apart from riverbeds.
What the models look at, and why.
Height above drainage stood out most in the recorded LightGBM diagnostic. Adding the surrounding ground gave the biggest overall improvement. Here are the main inputs and our interpretation of why they help.
- Height above drainage (HAND)
-
How high the patch sits above its drainage channel.
Helps separate riverbed changes from changes higher on a hillside.
- Vegetation change (NDVI)
-
How much greenness changed between images.
Vegetation loss can reveal a fresh scar, though construction and erosion can look similar.
- Slope
-
How steep the ground is.
Adds a clue about whether the location resembles landslide terrain. Steepness alone is insufficient.
- Terrain shape and roughness
-
Whether ground is smooth, uneven, curved or channelled.
Gives more context than the colour of one patch.
- Before/after satellite bands
-
Changes in colour and infrared reflectance.
Adds information about the changed surface beyond vegetation alone.
- Neighbourhood averages and variation
-
Measurements across 30 m, 70 m and 150 m windows.
Reduces reliance on one possibly unrepresentative pixel by describing the surrounding ground.
Two patches can both lose vegetation. One sits beside a stream on relatively flat ground; another sits on a steep hillside. Combining features helps distinguish them.
What we actually checked. The importance calculation used one LightGBM model evaluated on its training data. We have not established separate importance rankings for all four models. Treat the ranking as exploratory, and these explanations as plausible interpretations rather than proof of physical causes.
Background: NASA on vegetation indices and USGS on landslide terrain.
What this can
and cannot tell us.
The practical claim is a better inspection order. This has not been established as an end-to-end replacement detector or an early-warning system. The latest model uses after-event information, so it cannot establish warning before an event.
It can only sort the places already flagged. The original rules may miss landslides outside this list. The ranking cannot recover them, and district-wide recall is unknown.
Fresh independent evaluation is still needed. Tests withheld whole geographic areas and left a buffer from training points. But this remained repeated exploratory use of one historical inventory.