Applications

synapse-sr is built for the uses named in the SIH problem statement: crop monitoring, urban analysis, water mapping and disaster assessment. Each needs detail finer than 10 m and an honest statement of how far that detail can be trusted. Every example below works on any Result.

Crop monitoring

import synapse_sr

r = synapse_sr.super_resolve("fields.tif")
idx = r.indices()
ndvi, savi, evi, ndre = idx["ndvi"], idx["savi"], idx["evi"], idx["ndre"]   # NDRE uses the 20 m red edge

fields = synapse_sr.boundaries(r, "field")        # 0..1 parcel-boundary strength at 2 m
trusted = r.support == 2                          # observation-determined pixels
print("mean NDVI on trusted pixels:", float(ndvi[trusted].mean()))

On the development split (128 GeoSR pairs with a 2 m NAIP-derived reference), the NDVI edge F1 (field boundaries) is 0.599 for SYNAPSE Pro v2 (Pro v1: 0.597). The comparison figures are SEN2SR 0.548, LDSR-S2 0.570 and bicubic 0.477. An independent held-out evaluation is pending.

Urban analysis

import synapse_sr

r = synapse_sr.super_resolve("city.tif")
edges = synapse_sr.boundaries(r, "urban")         # buildings, roads
built = r.indices()["ndbi"]                       # built-up index (20 m SWIR context)

The urban benchmark covers 12 Indian cities (Google Open Buildings v3), with the same classifier for every product and leave-one-city-out evaluation. The table gives the fraction of individual buildings detected (measured with Pro v1):

Building size	Sentinel-2 10 m	Bicubic	SEN2SR	LDSR-S2	Satlas	SYNAPSE Pro v1
< 40 m²	0.556	0.581	0.599	0.620	0.636	0.635
40–100 m²	0.677	0.725	0.709	0.740	0.715	0.766
100–400 m²	0.812	0.855	0.831	0.867	0.783	0.894
> 400 m²	0.932	0.961	0.952	0.960	0.841	0.970

Exact 2 m footprint outlines remain beyond every product. Pixel IoU is 0.35–0.37 for all of them, including native 10 m (Satlas 0.28).

Water and flood mapping

import synapse_sr
r = synapse_sr.super_resolve("scene.tif")

water = r.indices()["ndwi"] > 0                   # 2 m water mask
shore = synapse_sr.boundaries(r, "water")         # shoreline / flood-front strength

Disaster and change assessment

import synapse_sr

before = synapse_sr.super_resolve("before.tif")
after = synapse_sr.super_resolve("after.tif")

flood = synapse_sr.change(before, after, "ndwi")          # water advance
burn = synapse_sr.change(before, after, "nbr")            # burn scar (20 m SWIR context)
damage = synapse_sr.change(before, after, "brightness")   # debris, collapse, bare soil
print(damage.area_km2, "km2 changed;", damage.unreliable_fraction, "of pixels excluded as unreliable")

change only flags pixels that are valid and observation-supported on both dates, so detected change rests on measured evidence.

In a known-truth benchmark (collapsed structures, debris strips, flood advance), every product was held at the same 0.5 % false-alarm rate on unchanged ground. SYNAPSE had the best F1: 0.331, against 0.319 for SEN2SR and 0.303 for native 10 m. That lead comes from precision, not stability: SYNAPSE varies slightly more than SEN2SR between two looks at unchanged ground. LDSR-S2 and Satlas ESRGAN vary much more, so they need a higher threshold and miss events. LDSR-S2 missed 25 % of debris events. Satlas ESRGAN detected no collapses and a quarter of the floods.

Available indices

Index	Formula	Bands	Use
NDVI	(N − R) / (N + R)	2 m	vegetation, crop vigour
SAVI	1.5 (N − R) / (N + R + 0.5)	2 m	sparse vegetation
EVI	2.5 (N − R) / (N + 6R − 7.5B + 1)	2 m	dense canopy
GNDVI	(N − G) / (N + G)	2 m	chlorophyll
NDWI	(G − N) / (G + N)	2 m	open water
NDRE	(N − B05) / (N + B05)	2 m × 20 m	crop stress, nitrogen
NDBI	(B11 − N) / (B11 + N)	20 m context	built-up
NBR	(N − B12) / (N + B12)	20 m context	burn severity
MNDWI	(G − B11) / (G + B11)	20 m context	water, flood

Indices that use a 20 m context band inherit that band's 20 m spatial detail.