Caspian Eye Prototype

Date A

Date B

Zone

Pilot zones

10 km

Move the pointer over the map to read coordinates and classes.

Field statistics

What the radar record shows

Open water mapped from Sentinel-1 for the selected zone. The October series compares the same season each year, so seasonal flooding does not drive the trend.

Open water each October, 2017–2026

km², Sentinel-1 orbit 108, VH threshold per scene. For the whole AOI, sea state and farmland add year-to-year noise; the pilot zones are steadier.

Show table

Former sea floor that stayed dry, by first dry year

km² of pixels that were water in Oct 2017 and land in every later October from that year

Show table

2026 season: open water from two radars

km², counted inside the area NISAR imaged on every date.

Show table

Phase 1 deliverable · Coverage and data audit

NISAR is already observing the Kura Delta

Query of NASA's Common Metadata Repository for NISAR Level-2 GCOV granules intersecting the pilot area.

Acquisition timeline by track and frame

Each mark is one GCOV product; color shows its release level. Rows are sorted by how much of the pilot area the frame covers.

NISAR quicklooks against the Sentinel-1 water map

NISAR browse images are 8-bit visualizations (~183 m), not calibrated γ⁰. Dark pixels were split with Otsu's method and compared with the nearest Sentinel-1 water map.

Full GCOV processing: track 165, frame 69

Calibrated L-band γ⁰ (HH, HV) read directly from each HDF5 product (about 1.4 GB) by byte range, reprojected from UTM 38N to the 20 m analysis grid. Water: a per-scene two-class model on smoothed HH and HV (5×5), trained on reference areas that Sentinel-1 mapped as water, or as land, on all six 2026 dates. Flooded-vegetation candidate: HH > −3 dB and HH−HV > 10 dB. In a test on 25 June, a single HV threshold reached only 78–80% agreement with Sentinel-1, because at 24 cm the smooth, dry semi-desert looks almost as dark as water. NISAR and Sentinel-1 maps line up to within one 20 m pixel.

Quality control

Validation against optical imagery

Sentinel-1 water maps compared pixel by pixel with Sentinel-2 water (NDWI > 0) on near-simultaneous cloud-free dates. Clouds, shadows and no-data pixels are excluded using the Sentinel-2 scene classification.

Where the two sensors disagree

Almost all disagreement is radar water that Sentinel-2 does not flag; optical-only water is under 1% of compared pixels.
On the open sea the two maps agree completely (checked on two offshore blocks, 16 Sep 2026).
Inland farmland is the largest source: about 165 km² of radar-only water on smooth bare fields in the north-west of the AOI on 11–16 Sep 2026. The radar mistakes these for water.
In Qızılağac Bay the shallow water is covered by algae and appears bright green in the optical layer. NDWI falls below zero there, while the radar still sees a water surface. Here the radar is the more reliable sensor.

Scientific method, as implemented

How each change signal is produced

STEP 1

Data selection

One relative orbit (108, descending) so every date has the same viewing geometry. October scenes for the 2017–2026 trend; monthly scenes for 2026.

STEP 2

Pre-processing

Sentinel-1 RTC γ⁰ (VV, VH) and NISAR L2 GCOV γ⁰ (HH, HV) on one 20 m UTM 39N grid. 3×3 boxcar on linear power to reduce speckle.

STEP 3

Classification

C-band: open water where smoothed VH is below a per-scene Otsu threshold. L-band: per-scene two-class model on HH and HV, trained on stable reference areas. Both then get a 3×3 majority filter. Flooded-vegetation candidate from double bounce: VV > −3 dB and VV−VH > 9 dB (C-band), HH > −3 dB and HH−HV > 10 dB (L-band).

STEP 4

Comparison

Pixel-wise transitions between any two dates, only where both dates are valid. Areas summed at 0.0004 km² per pixel inside each pilot polygon.

STEP 5

Confidence and validation

Per-pixel confidence from distance to the threshold (≥2 dB = high). Agreement with Sentinel-2 NDWI, NISAR GCOV and NISAR quicklooks reported above.

Limitations of this prototype

NISAR results come from one track (165, descending) and 10 dates since June 2026; most products are Provisional and may change after recalibration. The 2017–2026 trend relies on Sentinel-1, because NISAR data start in late 2025.
Over the sea, L-band HV sits at the instrument noise floor (about −22 dB), and the mean HH of water and of land differ by only 4–7 dB. NISAR therefore needs the two-channel model and reference areas; it also shows a noise-floor gradient across the swath.
Smooth bare or freshly worked fields can fall below the VH water threshold. This produces scattered inland changes on farmland that are not flooding. A land-cover or elevation mask is the planned fix.
Wind-roughened water can look like land in VV; the classifier uses VH for this reason, but very rough sea states can still cause misses. Smooth wet sand and mudflats can look like water.
The flooded-vegetation class is a radar signature candidate. Settlements and some infrastructure produce the same double-bounce signal and are not masked yet.
The October trend uses one scene per year, so a single unusual day (storm surge, strong wind) can shift that year's value. The production design uses several scenes per season.
Pilot zone boundaries are working polygons drawn for this demo. They need to be agreed with local institutions.

Delivery plan status

From this prototype to a monitoring product

Weeks 1–2

Coverage and data audit

Done. 4 NISAR tracks, 10 frame stacks and their dates catalogued for the AOI.

Weeks 3–4

Initial SAR analysis

Done: NISAR GCOV for 10 dates and Sentinel-1 for 15 dates, with water and change maps and area statistics.

Weeks 5–6

Web prototype

This page: date comparison, zone statistics, explanation panel. Next: a hosted map service.

Weeks 7–8

Validation and demo

Optical check done. Needs field data and review by SAR hydrology experts.

What we are asking for

Technical guidance: the recommended NISAR products and processing chain for wetland inundation (GCOV vs. GSLC coherence) and the expected cadence over the Caspian coast.
Scientific review: feedback on the threshold and double-bounce approach for L-band hydrology in reed-dominated deltas.
Pilot validation: comparison of Kura Delta results with independent geospatial and ecological data, with local environmental institutions.