Stages 04 – 05

Processing & AI

A delivered scene is raw material, not an answer. SatView runs it through a real processing chain, then reads it too — with models you control, on hardware you own.

04 · Process

Build the chain step by step, run it asynchronously

A pipeline is a chain of steps assembled in the browser — each step names the one whose output it takes, so a chain can branch. Jobs run on the Python backend, are tracked in the interface, and their outputs land back on the map as layers.

30+ native operations

Grouped by intent in the Processing Center: analysis (spectral indices, image statistics, change detection), geometry (reproject, extract ROI, orthorectification), filtering, pre-processing (band extraction, contrast enhancement, cloud masking, atmospheric correction DOS1) and fusion (image arithmetic, pansharpening, mosaic).

SAR & interferometry

SAR calibration, speckle filtering (Lee, Frost, Refined Lee) and InSAR coherence as pipeline steps; InSAR displacement on Sentinel-1 SLC pairs through ESA SNAP. Range-Doppler terrain correction is on the roadmap. Validation →

Classification & segmentation

K-means and threshold-index classification, morphological operations, watershed, graph-based and SLIC superpixel segmentation, connected components — each producing a labelled raster you can carry downstream.

Orfeo Toolbox & ESA SNAP

Industry-standard remote sensing toolchains driven through their CLI and GPT interfaces — not reimplemented, actually invoked.

Cloud-native

Process straight from a COG URL without downloading the full scene first, and push results back as cloud-optimized GeoTIFF.

Super-resolution

Upscaling for detail recovery on lower-GSD sources, available as a pipeline step like any other operation.

The Processing Center: an image library on the left, the full pipeline operation catalogue grouped into analysis, geometry, filtering, pre-processing, fusion, SAR, segmentation, classification and export, and a spectral indices parameter form set to NDVI with band assignments and a Run Processing button.

Processing Center · operation catalogue and job parameters

Pipeline Builder with a five-step chain named Port watch: atmospheric correction (DOS1), cloud masking, change detection with CVA and auto-threshold, object detection with an ONNX model, and conversion to a cloud-optimised GeoTIFF, each step taking the previous one as input.

Pipeline Builder · a port-watch chain — atmospheric correction, cloud mask, change since the last pass, object detection, cloud-optimised export

K-means land-cover classification of a Sentinel-2 scene of the Port of Rotterdam drawn over the satellite basemap: water, basins, container terminals, industrial ground and vegetation in six colours.

Process from URL · a public Sentinel-2 scene of Rotterdam read straight from its COG and classified (K-means, 6 classes)

Enhance before you analyse

HyperResolution sharpens the detail already in the image with a wavelet transform — the Mexican-hat method of Coupinot et al. (1992) — at one or several scales, with presets from ultra-light to vehicle-scale. It adds no invented pixels.

The same view of the Maasvlakte container terminal side by side, before and after HyperResolution with the Light preset: container rows, trucks and cars come out sharper on the right.

HyperResolution · before and after, Light preset, on the Maasvlakte container terminal

04 · Process — the science

Image science, checked against references

Every operation is a published method, and every one has been run on real imagery and compared with an independent implementation or a case whose answer is known. The table below gives the gap, not a promise.

Radar that reads like a map

Speckle is the grain of every radar image. SatView filters it with the methods SAR analysts use — Lee, Frost and Refined Lee (Lee 1981), the last one choosing, around each pixel, the half-window on the same side of an edge, so façades and arcades stay sharp while the noise goes.

Piazza San Marco in Venice seen by the Capella X-band radar at 0.2 m, on the SatView map: on the left the calibrated backscatter with its grainy speckle, on the right the same scene after the Refined Lee filter, smoother water and squares while the arcades of the Procuratie and the campanile keep sharp edges.

Venice, Piazza San Marco · Capella σ⁰ at 0.2 m, calibrated in SatView — raw (left) and Refined Lee 7×7 (right). Public Capella Open Data scene, 16 Aug 2024

Image science in 34 seconds · a real recording of SatView, every frame drawn by the application from the processed scenes: Venice radar de-speckled, Sentinel-1 calibrated and geocoded, NDVI, and the measured gaps

A full Sentinel-1 IW GRD scene over the Loire valley and the Orléans forest drawn in greyscale over the satellite basemap: radar backscatter in decibels, rivers dark, towns bright, field patterns visible across about 250 km.

Sentinel-1 σ⁰ in dB · a full IW scene (26 442 × 16 669 px) calibrated with ESA's own tables, then placed on the map from its 210 ground control points

NDVI of the Port of Rotterdam from Sentinel-2 drawn over the satellite basemap: water and quays in red, parks, dunes and green belts in green.

NDVI · Sentinel-2 L2A over the Port of Rotterdam — water and concrete in red, vegetation in green

Validation on real scenes

Operation	Reference	Test	Gap
Sentinel-1 radiometric calibration (σ⁰)	ESA calibration tables, bilinear as specified by ESA	Full IW GRD scene, Sentinel-1A, 29 Jun 2025 — compared with an independent implementation on 20 480 pixels	≤ 0.000003 dB
Geocoding of a GRD from its ground control points	Pixel predicted from the neighbouring control points	Around 152 control points of the same scene	94.8 % identical pixels (0.9 % if shifted by one)
Refined Lee speckle filter	Lee (1981), independent implementation	Capella X-band scene, 1 028 196 pixels compared	≤ 0.00001 % (rounding)
Lee speckle filter	Lee (1981); Lopes, Touzi & Nezry (1990)	Equivalent number of looks on open sea	0.78 → 3.9, mean kept within 1.1 %
InSAR coherence	Touzi et al. (1999) estimator	Radar pairs of known coherence 1 / 0.6 / 0	1.000 / 0.604 / 0.190 (expected 1.000 / 0.604 / 0.189)
RPC orthorectification	Copernicus DEM + EGM2008 geoid	Image with a known sensor model; height used recovered from the result	DEM + 45.4 m geoid, within 0.5 m
Spectral indices (NDVI, EVI)	Same formula computed outside SatView	Sentinel-2 L2A, 21 Jun 2025	NDVI 0.1176 vs 0.118; EVI 0.1086 vs 0.1088

What a method does not do is said as plainly: Refined Lee lowers the mean by about 9 % on single-look data (Lee is preferred there); geocoding from control points is not a full terrain correction; a coherence computed from detected (GRD) images is labelled as a backscatter correlation, not interferometric coherence.

05 · Analyze

Inference that produces geometry, not just pictures

Every AI job returns a raster and a companion GeoJSON reprojected to WGS84 — so results are map layers, queryable objects and exportable evidence, not screenshots.

Object detection — ONNX models run with a sliding window, or classic blob detection. 15 classes out of the box (plane, ship, storage tank, harbor, helicopter, large/small vehicle, roundabout, bridge, sports fields and more) — bring your own model for anything else. Output: heatmap raster plus GeoJSON, with a confidence-threshold slider and per-class counts.
Change detection — Change Vector Analysis with Otsu thresholding, polygonized into vector change masks.
Semantic segmentation — ONNX models or spectral K-means, producing a labelled raster with its legend.
Detection layer — results rendered on the map as a GeoJSON layer, alongside the imagery they came from.
Auto-indexing — every detection feeds a searchable index, so “show me every ship detected in this area last month” becomes a query rather than a manual review.

Bring your own models — SatView ships with no detection engine baked in. Your administrator plugs in an open-source model, or one under its own specific license, through the Model Manager — upload, version and swap ONNX weights whenever you want. Nothing is locked to a vendor's model zoo, and SatView never bundles, sells or redistributes third-party AI models itself.
AI Vision — a multimodal model describes the scene and answers questions about it.
Natural-language query — plain English translated into catalog search parameters.

Where the models run matters. Inference happens on your Python backend, on your hardware. Imagery is not uploaded to a third-party inference API unless you explicitly configure one for AI Vision or natural-language query.

Object detection results over an airport: 99 objects detected and outlined on the map — 85 planes, 10 roundabouts, 2 bridges, a ground-track field and a soccer-ball field — with a confidence-threshold slider and a per-class checklist covering all 15 supported classes.

Object detection · 99 detections, confidence threshold and per-class breakdown

Detection, change and classification on a real port

Detect Object run on the current map view over the Maasvlakte container terminal in Rotterdam: 28 objects outlined with oriented boxes — 19 ships at the quay and 9 large vehicles — with the confidence threshold and the per-class counts.

Detect Object on the current view · 28 objects on the Maasvlakte terminal — 19 ships, 9 large vehicles

Change detection between two Sentinel-2 images of the Port of Rotterdam, September 2020 and June 2025: 153 changed areas outlined in orange over the port, with a legend giving 3.77 % of the area changed.

Change detection · Sentinel-2, September 2020 against June 2025 — 153 changed areas, 3.8 % of the port

AI Vision Analysis panel on a satellite image of the Port of Rotterdam container terminal: scene type industrial at 95 % confidence, a description of the terminal, detected objects — container stacks, 25 gantry cranes, 4 container vessels, 3 tugboats — and an assessment.

AI Vision · a multimodal model reads the scene — industrial, 25 gantry cranes, 4 container ships, 3 tugs — and gives its assessment

ML Models window listing the object-detection model in use: an oriented-box YOLOv8 model trained on the DOTA dataset, 170 MB, 1024-pixel input, with its classes, replaceable by the administrator.

ML Models · the detection model in use and its classes — swap in your own ONNX weights

Exploitation

Reading the results

Time Series Viewer

Swipe comparison between acquisitions across a date range.

InSAR viewer

Reads back the coherence and displacement products as map layers.

Content search

Query the index of detected objects, with results on the map and in a list, or filtered to dark-vessel and recurring-entity candidates.

Intel reports

Generate a report and export it to PDF with cover page, tables and classification banner.

Content Search panel over the Maasvlakte quay: nine ships found in the index for the current map view, listed with their confidence and marked on the map, each flagged with a warning when no AIS position matches it.

Content search · every ship detected on the quay, from the index — flagged when no AIS position matches

How detections are checked against AIS broadcasts, and how dark-vessel and recurring-entity candidates are built, is on the Intelligence page.

Visualisation

My map · your own images, dated, on the map and on the relief

Every image in your library is drawn where it belongs, newest on top. A timeline steps through your acquisition dates and shows the map as it stood on each of them; colours can be harmonised between neighbouring images, and a click on any point lists the images of yours that cover it.

The same map opens in 3D: your images are draped over the Copernicus DEM GLO-30 relief, with the height factor of your choice (×1, ×1.5, ×2). The shapes drawn by the other modules on the 2D map follow, a click gives the altitude and the image history of the point, and the view exports as a PNG carrying its date, its method and the data credits. It is a situational view, not a measurement: a 30 m surface model, loaded around your images only.

The My map bar over the satellite basemap: a Sentinel-2 image of the Mont Blanc massif drawn in place, a timeline with one tick per acquisition date, and the Harmonise colours, 3D, Images and Point history buttons.

My map · a library image in place, with the timeline of acquisition dates

The 3D view of My map at a past date: a Sentinel-2 image of 26 September 2023 draped over the Copernicus GLO-30 relief of the Mont Blanc massif, the Mer de Glace and the Chamonix valley, on a dark neutral ground.

3D view · Sentinel-2 image of 26 September 2023 over the GLO-30 relief — Mont Blanc massif, Mer de Glace, Chamonix valley

A click in the 3D view opens a card with the coordinates of the point, its altitude of 2408 m on the GLO-30 surface above the EGM2008 geoid, and the two dated images of the library that cover it, each selectable to show it alone.

Point history · altitude of the clicked point and the dated images that cover it

The PNG exported from the 3D view: the whole massif seen from the north-west, with a caption band stating the date shown, that it is a situational view, the relief source and height factor, and the Copernicus and MapLibre credits.

PNG export · the view with its date, method and credits

Images: Copernicus Sentinel-2 data (2023, 2025), true-colour product, processed in SatView. Relief: Copernicus DEM GLO-30. The July 2025 scene shown in the first, third and fourth pictures is saturated on snow in the source product.

Now make it run without you

Monitoring rules turn this chain into a standing watch — re-acquiring on cadence, or retrying when cloud or quality fall short.

Autonomous monitoring Full feature catalog