THE STUDY WINDOW
60 m analysis cellsAmaravati, in focus
Preparing the observation map…
VegetationExposedWaterBuilt
16.53° N 80.515° E INDIA
Land, seen a little closer.
A small window into Amaravati’s changing landscape.
Real NASA HLS inputs. Model trained against ESA WorldCover 2021 reference labels. Loading evidence…
AREA COMPARISON
km² · valid classificationsWhat the observations show
FIELD REVIEW
16Sectors to investigate
COMPARE OBSERVATIONS
Change is an observation, before it is an explanation.
Compare similar-season snapshots. Crop calendars, construction and weather can all change land cover.
Land-cover transitions
Only cells valid on both dates and above the model-score threshold.
Vegetated-reference anomalies
An Isolation Forest compares current optical indices with the 2021 vegetated reference. This is spectral novelty, not a diagnosis of crop stress.
Planting dates and verified farm records are required before interpreting this as a crop-season alert.
Sector-level comparison
From satellite light to a land-cover label.
10
Measurements per cell
Six reflectance bands and NDVI, NDWI, NDMI, NDBI describe the optical signal.
96
Decision trees
The fitted forest combines tree votes to assign one of four cover classes.
4
Cover classes
Vegetation, exposed ground, water and built surfaces. Crop species are not identified.
The model classifies the supplied observation dates. Future land-use ideas in area reports are conditional planning scenarios, separate from this trained classifier.
Held-out confusion matrix
Rows: WorldCover reference. Columns: ReGrow prediction. Each square contains a number of test cells.
Why are there 16 squares?
4 reference classes × 4 predicted classes = 16 possible outcomes. The diagonal counts matches; other squares count disagreements. These squares are not geographic areas.
The map separately has 16 clickable inspection sectors (a 4 × 4 geographic grid). Each sector groups many 60 m cells. Those sectors are not the ~960 m spatial blocks used to separate training and testing.
Which inputs matter?
Impurity-based feature importance. Correlated features share influence.
Overall agreement is only part of the story.
Precision asks how often a predicted class matches the reference. Recall asks how much of a reference class was found. F1 balances both. Rare classes need more verified examples.
Exposed-ground recall is 26.5% on just 34 test cells. Built-surface F1 is 0.481. Vegetation and water perform better against this reference map. These are study-wide 2021 results; later dates and individual sectors have no independent accuracy assessment.
A traceable training pipeline
↓ Training report01Observe
NASA HLS S30 surface reflectance, with Fmask screening.
02Reference
ESA WorldCover 2021 modal labels on the same 60 m grid.
03Separate
Spatial blocks for fitting, validation and held-out testing.
04Evaluate
Compare bounded-depth models, then report untouched test blocks.
Validation is inside the development set. After selecting settings, the final model refits on all development cells. Held-out test blocks remain separate. Spatial separation reduces nearby-pixel leakage; no buffer gap was applied, so residual spatial dependence is possible.
Validation trials
Model-score threshold
Hide lower-scoring classifications across the dashboard. The score is tree-vote support and is not a calibrated probability.
84,656 observations. Every row traceable.
21,164 fixed 60 m cells across four satellite dates. The CSV retains 82,859 input-quality-valid rows and 1,797 excluded rows, with screening flags, source IDs, ten model inputs and dated reference labels. Quality validity is separate from your model-score threshold. The latest included observation is 19 April 2026; this is not a live feed.
Extract the .gz file before opening the CSV. Filter input_quality_valid = 1 before using optical measurements. Later-year reference labels remain dated 2021, not new ground truth.
Observation manifest
Original catalogue identifiers remain attached to each snapshot.
What this prototype can establish
The water measurements
| Measurement | Available here? | Interpretation |
|---|---|---|
| Surface-water extent | Yes · optical classification | Mapped water area at an observation date |
| Water depth / reservoir level | No measurements supplied | Needs a gauge or a suitable SWOT observation |
| Local soil moisture | No sensor record supplied | NDMI is a vegetation moisture index, not measured soil moisture |
| Groundwater / soil chemistry | No field records supplied | Requires local measurements and expert assessment |