PadhaikoAll subjects

Passive Remote Sensing

Remote Sensing and Geospatial Data Processing (ENCT 437)

Open in the Padhaiko app

3.1 Radiometry, Spectrometry and Hyperspectral Sensing

Passive remote sensing measures naturally available solar radiation reflected from Earth's surface or thermal radiation emitted by the surface and atmosphere. Quantitative interpretation of passive imagery begins with radiometry—the exact physical measurement of electromagnetic energy flow across space, area, solid angle, and wavelength.

Fundamentals of Radiometry

Core Radiometric Quantities and SI Units
QuantitySymbolDefining RelationSI UnitPhysical Meaning
Radiant Energy (joule)Total electromagnetic energy carried by photons
Radiant Flux (Power) (watt)Rate of energy transfer per unit time
Radiant IntensityFlux emitted by a point source per unit solid angle
IrradianceFlux incident onto a unit surface area from all directions
Radiant ExitanceFlux leaving a unit surface area into the full hemisphere
RadianceFlux per unit projected area () and per unit solid angle (); invariant along a lossless ray

Adding the subscript yields spectral quantities per unit wavelength interval—for example, spectral radiance () and spectral irradiance (). Because radiance is conserved along an unattenuated optical path (independent of distance to an extended target), satellite sensors directly measure at-sensor spectral radiance.

Spectral Radiance and Reflectance Measurements

In field or laboratory spectroscopy, spectral reflectance factor (or ) is the dimensionless ratio of radiance reflected by a target to that reflected by a calibrated, nearly ideal Lambertian white reference panel (e.g., Spectralon) under identical illumination and viewing geometry:

For hemispheric illumination , an ideal Lambertian surface satisfies .

Field Spectroradiometers and Spectral Reflectance Library Integration

enabling empirical validation, spectral unmixing endmember selection, and mineral/species identification.

Principles of Optical Dispersion and Sensor Evolution

To measure spectra across multiple wavelengths, imaging spectrometers separate incoming polychromatic light using optical dispersion:

Evolution from Panchromatic to Contiguous Hyperspectral Sensing
Sensor ClassBand CountTypical Bandwidth ()Spectral CoverageKey Trade-off & Application
Panchromatic (PAN)Very wide ()Integrated visible–NIR brightnessHighest spatial resolution (more photons/pixel); pan-sharpening
Multispectral (MS)Broad ()Discrete atmospheric windowsBroad land-cover mapping, vegetation indices (Landsat OLI)
SuperspectralModerate ()Includes narrow Red-Edge, cirrus & water-vapor bandsPrecision agriculture, coastal water quality (Sentinel-2, MODIS)
Hyperspectral (Imaging Spectroscopy)Narrow, contiguous ()Continuous spectrum ()Fine mineral absorption, invasive species, gas plumes (EnMAP, PRISMA, EMIT)

3.2 Radiometric Calibration

Raw optical satellite images store quantized Digital Numbers (DN)—unsigned integers (e.g., for -bit or for -bit products). Because DN values depend on detector gain, solar elevation, Earth–Sun distance, and sensor degradation, raw DNs cannot be compared across dates or sensors without radiometric calibration.

Conversion of DN to Top-of-Atmosphere (TOA) Radiance and Reflectance

Definition

DN to TOA Radiance and TOA Reflectance Equations

  1. DN to TOA Spectral Radiance (): Using sensor-specific multiplicative gain () and additive offset () coefficients supplied in the product metadata (e.g., Landsat MTL file):

(Equivalently in min–max calibration form: .)

  1. TOA Radiance to TOA Planetary Reflectance (): Normalizing for solar illumination geometry and Earth–Sun distance (in Astronomical Units, ):

where is the mean exo-atmospheric solar irradiance in that band and is the solar zenith angle ( is solar elevation). (In modern Landsat 8/9 OLI Level-1 products, is also given directly via reflectance rescaling coefficients: .)

Sensor Calibration and Radiometric Normalization

Also includes destriping (equalizing detector-to-detector mean and variance across pushbroom arrays).


3.3 Atmospheric Correction

Top-of-Atmosphere reflectance still contains atmospheric path radiance (scattered light that never touched the ground) and two-way transmittance loss ():

where is Bottom-of-Atmosphere (surface) reflectance, is spherical atmospheric albedo, and are direct and diffuse downward irradiances. Retrieving is essential for quantitative indices and multi-temporal change analysis.

Empirical Atmospheric Correction Methods

  1. Dark Object Subtraction (DOS): Assumes that some pixels in every scene (deep clear water, dense shadow) have near-zero true surface reflectance (), so their non-zero measured radiance (often taken at the histogram 1st percentile minus dark-object reflectance) approximates the additive path radiance . Subtracting from every pixel removes haze (basic DOS1); refined versions (DOS2–DOS4) also model transmittance .
  2. Flat Field (FF) Correction: Divides every pixel spectrum in a scene (especially hyperspectral cubes) by the mean spectrum of a spectrally flat, bright, spatially uniform target (e.g., playa salt pan or concrete tarmac) within the scene to cancel solar irradiance shape and atmospheric absorption bands simultaneously.
  3. Internal Average Relative Reflectance (IARR): Normalizes each pixel spectrum by the mean spectrum of the entire image, requiring no known ground target; effective over arid/geological scenes dominated by bare rock, though sensitive if vegetation covers much of the scene.

Radiative Transfer Models (RTMs): 6S and Sen2Cor

Unlike empirical shortcuts, physically based Radiative Transfer Models solve the electromagnetic radiative transfer equation through layered atmospheric profiles of aerosols, water vapor, ozone, and Rayleigh scattering:


3.4 Geometric Restoration and Resampling

Raw satellite imagery contains spatial distortions caused by Earth rotation, curvature, panoramic/off-nadir viewing, sensor altitude/attitude drift, and terrain relief. Geometric restoration maps image pixel coordinates onto a true map projection .

Systematic vs. Non-Systematic Geometric Errors

Ground Control Points (GCPs), Registration, and Polynomial Transformation

Image registration aligns an image to a map coordinate system (georeferencing) or to another image (image-to-image registration) using well-defined Ground Control Points (GCPs)—identifiable intersections, bridges, or corners whose coordinates are known in both spaces.

A 2D polynomial coordinate transformation relates map coordinates to image coordinates :

Typically is required.

Orthorectification

Polynomial warping alone cannot remove relief displacement in mountainous terrain (like Nepal's Himalayas). Orthorectification uses a rigorous sensor collinearity/RPC camera model together with a DEM to project every pixel onto its true planimetric location at terrain elevation , eliminating topographic tilt and scale distortion.

Resampling (Interpolation) Methods

When mapping an output map grid coordinate back to fractional coordinates in the input image, a resampling kernel assigns the output DN:

Comparison of Spatial Resampling Methods
MethodKernel SupportMathematical PrincipleRadiometric FidelityVisual & Operational Trade-off
Nearest Neighbor (NN)Assigns DN of the closest integer input pixelExact preservation of original DNs (no averaging)Introduces blocky/stair-stepped edges and up to spatial shift; best for classified maps
Bilinear InterpolationWeighted linear average of the surrounding pixels along and Smooths DNs (reduces extreme peaks)Removes jagged steps with slight spatial blurring; fast standard choice for continuous imagery
Cubic ConvolutionFits piecewise cubic splines across surrounding pixelsSlight smoothing plus minor edge overshoot/ringingProduces sharpest, visually pleasing continuous edges; slower; unsuitable before strict spectral unmixing

3.5 Visual Image Interpretation

Even in an era of automated machine learning, human visual image interpretation remains essential for validating classifications, mapping complex geomorphology, and rapid disaster damage assessment.

Eight Elements of Visual Image Interpretation

The Eight Fundamental Elements of Image Interpretation
ElementDefinitionDiagnostic Example
Tone / ColorRelative brightness (grayscale DN) or hue/saturation across bandsDeep clear water appears dark in NIR; turbid sediment-laden rivers appear bright cyan
SizePhysical dimensions of an object in context of image scaleDistinguishes residential houses from large industrial warehouses or aircraft types
ShapeExternal geometric outline or formMeandering rivers vs. straight canals; cloverleaf highway interchanges
TextureFrequency of tonal variation ("smooth" vs. "coarse")Calm water and bare concrete are smooth; mature forest canopies appear mottled/coarse
PatternSpatial arrangement and repetition of featuresGrid street networks, orchard tree rows, terraced Himalayan hillslope farming
ShadowDark silhouette cast opposite solar azimuthReveals vertical height of buildings, cliffs, and trees (though obscures ground beneath)
SiteTopographic or geographic settingAlpine tundra above tree line; mangroves along tidal mudflats
AssociationCo-occurrence of related objects that imply functionCooling towers + rail spurs + coal yard identify a thermal power plant

True-Color Composites (TCC) and False-Color Composites (FCC)

Because computer displays drive Red, Green, and Blue (R–G–B) guns simultaneously, any three spectral bands can be mapped to R–G–B:

AspectTrue-Color Composite (TCC)Standard False-Color Composite (FCC)
Band Mapping (R, G, B), , , ,
Healthy VegetationAppears natural green (often low contrast between species)Appears bright red / magenta because high NIR reflectance drives the Red channel
Clear WaterDark blue or greenish-blueVery dark blue to black (water absorbs NIR and Red)
Urban / Bare SoilGray, brown, tanCyan / light blue-gray (moderate Red/Green, low NIR relative to vegetation)
Primary StrengthIntuitive natural appearance for non-specialistsMaximizes vegetation vigor contrast, crop stress detection, and land–water boundary clarity

Other specialized FCC combinations—such as SWIR–NIR–Red (often called agricultural or false-urban composite)—highlight moisture variations, burned scars, and geological differences.


Worked Out Examples

Example 3.1

3.1 (8 marks)

A Landsat-like sensor records a pixel in the Red band with calibrated digital number . Metadata gives radiance scaling factors and . At acquisition time, Earth–Sun distance is , solar zenith angle is , and mean exo-atmospheric solar irradiance is .

  1. Convert to Top-of-Atmosphere (TOA) spectral radiance .
  2. Convert to TOA planetary reflectance .

Converting DN to TOA spectral radiance:

Converting TOA radiance to TOA reflectance:

With and :

Answer:
  1. .
  2. ().
Example 3.2

3.2 (6 marks)

In a Sentinel-2 NIR band, a deep clear-water dark object has measured TOA reflectance , while a vegetated pixel has and a dry asphalt pixel has .

  1. Using simple DOS1 (assuming and ), estimate the surface reflectance of the asphalt pixel.
  2. Using DOS2 where atmospheric two-way transmittance is estimated as (), compute the surface reflectance of the vegetated pixel.

Applying DOS1 path-radiance subtraction: The dark-object TOA reflectance estimates path reflectance . For the asphalt pixel:

Applying DOS2 with atmospheric transmittance ():

Answer:
  1. Asphalt surface reflectance (DOS1): .
  2. Vegetation surface reflectance (DOS2): .
Example 3.3

3.3 (6 marks)

During geometric resampling, a target pixel maps to fractional image coordinates . The four nearest integer pixel Digital Numbers are:

Compute the output DN using (1) Nearest Neighbor and (2) Bilinear Interpolation.

Nearest Neighbor assignment: Rounding to the nearest integer coordinates gives (distances squared: to ; to ). Thus:

Bilinear Interpolation ( neighborhood): Let and .

  • Interpolate along on row :
  • Interpolate along on row :
  • Interpolate along between row and row :
Answer:
  1. Nearest Neighbor DN .
  2. Bilinear Interpolation DN .

Exercises

Exercise — Passive Remote Sensing & Preprocessing

Question 1

Distinguish between irradiance () and spectral radiance (), explain why satellite sensors measure radiance, and describe how field spectroradiometer data are integrated into spectral libraries for multispectral sensors. A field spectroradiometer measures over grass and over a reference panel of reflectance ; calculate the target reflectance.

Irradiance vs. radiance & spectral library integration: Irradiance () is total power per unit area arriving from a hemisphere, whereas spectral radiance () is flux per unit projected area and solid angle—invariant with distance in lossless space, making it directly measurable by an imaging sensor's IFOV. Field spectroradiometers record continuous target spectra relative to a white Spectralon panel; to match multispectral bands, continuous library spectra are convolved with each band's Spectral Response Function (SRF).

Computing target reflectance:

Answer:

Radiance is directional and distance-invariant while irradiance is hemispherical; field spectra are convolved with band SRFs for sensor matching. Target reflectance is ().

Question 2

Explain why raw Digital Numbers (DNs) must be calibrated before multi-date analysis, and outline the workflow from DN to TOA radiance and TOA reflectance (plus relative radiometric normalization). If a Landsat 8/9 OLI pixel has with reflectance rescaling factors and at solar zenith angle , calculate the TOA reflectance .

Why calibration and normalization are needed: Raw DNs reflect sensor-specific gain/offset, solar zenith angle, Earth–Sun distance, and detector drift. Converting DN to removes sensor gain/offset, converting to removes solar angle and Earth–Sun distance differences, and Relative Radiometric Normalization (RRN) aligns uncalibrated dates via linear regression over Pseudo-Invariant Features (PIFs).

Computing from OLI rescaling coefficients: Unadjusted planetary reflectance:

Dividing by :

Answer:

Calibration removes sensor gain, solar irradiance geometry, and Earth–Sun distance effects; RRN aligns dates via PIFs. Here ().

Question 3

Compare empirical atmospheric correction methods (Dark Object Subtraction, Flat Field, and IARR) with physically based radiative transfer models (6S and Sen2Cor), highlighting assumptions, required inputs, and limitations.

Empirical methods (scene-based):

  • DOS: Assumes dark targets (deep water, shadows) have , using their minimum value to subtract additive path radiance; simple and needs no external atmospheric data, but ignores spatial aerosol variation and (in DOS1) multiplicative transmittance.
  • Flat Field: Divides by a bright, spectrally flat in-scene target; requires a known homogeneous flat feature.
  • IARR: Divides by the image-wide mean spectrum; needs no field targets, but distorts spectra if vegetation dominates the scene.

Radiative transfer models (6S and Sen2Cor):

  • 6S simulates Rayleigh/Mie scattering, gaseous absorption (), adjacency, and BRDF using atmospheric optical depth inputs and LUTs.
  • Sen2Cor operationally corrects Sentinel-2 by classifying scenes, estimating AOT (DDV) and water vapor ( band), applying DEM terrain correction, and inverting libRadtran LUTs to deliver accurate surface reflectance ().
Answer:

Empirical methods (DOS, Flat Field, IARR) rely on in-scene target assumptions with zero external atmospheric parameters, whereas RTMs (6S, Sen2Cor) physically model absorption, scattering, aerosols, water vapor, and terrain to invert accurate BOA reflectance.

Question 4

Distinguish between systematic and non-systematic geometric errors, explain why orthorectification is required in mountainous regions, and compare Nearest Neighbor, Bilinear, and Cubic Convolution resampling. Three independent check GCPs yield residuals , , and pixels; compute the overall geometric .

Geometric errors, orthorectification, and resampling: Systematic errors (Earth rotation skew, curvature, scan geometry) are corrected from ephemeris metadata, whereas non-systematic errors (attitude drift, terrain relief displacement) require GCPs and a DEM. Polynomial warping cannot fix elevation-dependent relief displacement in mountains, so orthorectification uses a camera model + DEM to place every pixel at true planimetric . For resampling: Nearest Neighbor preserves exact DNs (essential for categorical maps) but looks blocky; Bilinear smooths smoothly (); Cubic Convolution gives crisp edges () with slight DN overshoot.

Computing GCP check RMSE:

Answer:

Orthorectification uses a DEM + sensor model to remove mountain relief displacement; NN preserves DNs, Bilinear smooths, and Cubic sharpens. .

Question 5

List and illustrate the eight elements of visual image interpretation, and contrast a True-Color Composite (TCC) with a standard False-Color Composite (FCC) in terms of band-to-RGB assignment and appearance of vegetation, water, and urban features.

Eight visual interpretation elements:

  1. Tone/Color: Pixel brightness or hue (e.g., dark clear water vs. bright turbid water).
  2. Size: Relative scale-based dimensions (house vs. warehouse).
  3. Shape: Outline geometry (meandering river vs. straight highway).
  4. Texture: Spatial smoothness/roughness (smooth pavement vs. mottled forest).
  5. Pattern: Spatial repetition (urban street grid, terraced slopes).
  6. Shadow: Silhouette revealing height/relief.
  7. Site: Topographic/ecological location (mangroves on coasts).
  8. Association: Co-occurring contextual features (runways + hangars = airport).

TCC vs. Standard FCC:

  • TCC (): Natural human-eye view; vegetation is green, clear water is dark blue, urban areas are gray/brown.
  • Standard FCC (): Healthy vegetation appears bright red/magenta (high NIR reflectance), clear water appears black/very dark blue (strong NIR absorption), and urban/bare soil appears cyan/blue-gray, sharpening vegetation-health and land–water contrast.
Answer:

The eight elements (tone/color, size, shape, texture, pattern, shadow, site, association) guide photo-interpretation. TCC maps R–G–B naturally, whereas standard FCC maps NIR–Red–Green to R–G–B, rendering healthy vegetation red and clear water black.

← Active Remote Sensing and Explorations (part 2: 2.3–2.4)Digital RS Image Processing and Enhancements →