Photogrammetry, Remote Sensing and Image Processing

1

This section covers the definition of photogrammetry, its principles and types, a brief history, and its scope and applications.

2

Definition and Principle • Photogrammetry (ASPRS) is the art, science and technology of obtaining reliable information about physical objects and the environment through the processes of recording, measuring and interpreting photographic images and patterns of recorded electromagnetic radiant energy.

3

In short, it derives metric (position and shape) and interpretative (identification) information from images. • Basic principle: a photograph is a central (perspective) projection — every image point, the perspective centre (camera lens) and the corresponding object point lie on one straight line (the collinearity condition).

4

A single photograph gives only direction, so two overlapping photographs taken from different positions allow the rays to be intersected, reconstructing the object in three dimensions (the stereoscopic or space-intersection principle). • A map, by contrast, is an orthogonal projection at constant scale; a photograph has varying scale because of relief and tilt, which is why photographs must be rectified/ortho-rectified before they can be used as maps.

5

Basis of classification Types Camera station Terrestrial (close-range) photogrammetry — camera on the ground; aerial photogrammetry — camera in an aircraft or UAV; space/satellite photogrammetry Technology Analogue (optical-mechanical plotters), analytical (computer-assisted, Helava 1957) and digital (softcopy) photogrammetry Purpose Metric photogrammetry (measurements — coordinates, distances, volumes, DEMs, maps) and interpretative photogrammetry (photo interpretation and remote sensing) Platform/sensor Manned aircraft, UAV/drone, satellite; frame camera, pushbroom scanner, laser scanner combined systems Brief History • 1839 — invention of photography (Daguerre, Niépce);

6

1849–1859 — Aimé Laussedat, 'the father of photogrammetry', used terrestrial photographs for mapping ('metrophotography');

7

1858 — first aerial photograph from a balloon (Nadar);

8

1901 — Pulfrich's stereocomparator introduces stereoscopic measurement; the First and Second World Wars drive aerial photography and analogue stereo-plotters (Zeiss, Wild);

9

1957 — Helava's analytical plotter;

10

1970s–80s — analytical aerial triangulation and bundle adjustment;

11

1990s — digital (softcopy) photogrammetry, GPS/IMU direct georeferencing;

12

2000s — airborne LiDAR and high-resolution satellite stereo;

13

2010s — UAV photogrammetry with structure-from-motion (SfM), dense image matching and 3D point clouds.

14

Scope, Applications, Merits and Limitations • Applications: topographic mapping and map revision;

15

DEM/DTM and orthophoto production; cadastral and land-records mapping; engineering surveys — highway and canal alignment, earthwork volumes, quarry and stockpile measurement, as-built and deformation surveys; architectural and heritage documentation (temples and monuments — widely used in Nepal after the 2015 earthquake); forestry (tree heights, canopy), agriculture and land-use/land-cover mapping; geology, glaciology and landslide studies; disaster damage assessment; traffic accident and forensic recording; industrial metrology;

16

3-D city models, virtual reality and games. • Advantages: covers large or inaccessible areas quickly; measurements are made in the office, reducing field work; the photograph is a permanent record of conditions at the moment of exposure, and can be re-measured later for new purposes; gives 3-D information and continuous coverage; economical for large areas. • Limitations: depends on weather, season and daylight; needs ground control and skilled staff; high initial investment in cameras, aircraft/UAV and software; features hidden under tree canopy, in shadow or under buildings cannot be mapped; accuracy is limited by image scale/GSD; permissions and airspace regulations (in Nepal, CAAN and local authority permission for UAV flights).

1

This section covers the types of aerial photographs, the scale of a vertical photograph, the photogrammetric process and flight planning, interior, relative, absolute and exterior orientation, aerial triangulation and block adjustment, the effects of relief and tilt displacement, rectification, oblique photography, photo mosaics and photo maps, pre- and post-pointing, and the properties of an ideal ground control point.

2

Types of Aerial Photographs Type Description Vertical photograph Camera axis truly vertical, or unintentionally tilted by less than about 3°; used for mapping and mosaics Tilted photograph Unintentional tilt greater than ≈ 3° Low oblique Axis intentionally inclined; the horizon does not appear High oblique Axis strongly inclined; the horizon appears in the photograph Convergent / trimetrogon Two or more cameras exposed simultaneously at angles — used for reconnaissance and 3-D city modelling (modern oblique camera systems) • Reference points on a vertical photograph: the principal point (intersection of the optical axis with the photo, located from the fiducial marks), the nadir point (plumb point — relief displacement is radial from this point) and the isocentre (tilt displacement is radial from it).

3

On a truly vertical photo all three coincide.

4

Scale of a Vertical Photograph • Scale S = f/(H − h), where f = focal length, H = flying height above datum and h = ground elevation of the point; equivalently S = photo distance ÷ ground distance.

5

Because h varies, the scale of a single photograph varies from point to point — higher ground is at a larger scale. • Average scale uses the mean ground height havg; datum scale uses h = 0.

6

Example: f = 152 mm, H = 2 000 m above MSL, ground at 500 m → S = 0.152/1 500 = 1:9 868. • Ground coverage of one photo = photo format × scale denominator (a 23 cm × 23 cm photo at 1:10 000 covers 2.3 km × 2.3 km).

7

For digital cameras the equivalent measure is the ground sample distance, GSD = pixel size × H/f.

8

The Photogrammetric Process and Flight Planning • Process: project planning → ground control survey → flight planning and photography → image processing → interior/relative/absolute orientation (or bundle adjustment) → aerial triangulation → stereo compilation, DTM generation and feature extraction → orthophoto and map production → field completion and quality check. • Flight planning fixes the photo scale (from the map scale and accuracy needed), hence the flying height H = f × scale denominator above mean ground level, and then the camera, overlaps, flight lines, number of photos and the exposure interval. • Overlaps: forward (end) overlap ≈ 60% (55–65%) so that every point appears on at least two photos and stereo coverage is continuous; side lap ≈ 20–30% between adjacent strips.

9

Air base B = (1 − forward overlap) × ground coverage along the flight line; the distance between flight lines = (1 − side lap) × ground coverage across the line; exposure interval t = B/V (V = ground speed); number of photos per strip = (length of strip ÷ B) + 1 (+ 2 extra at each end). • Other considerations: flight lines usually run in the direction of the longer dimension of the area and, in hilly country, along the valleys/contours; sun elevation above ≈ 30° to limit shadows; cloud-free, haze-free weather; leaf-off season for topographic detail; allowance for crab (rotation of the photo about the vertical axis relative to the flight line) and drift (departure from the planned line due to wind); and for large height differences the scale variation may require different flying heights.

10

Orientations Orientation What it does / elements Interior (inner) orientation Reconstructs the internal geometry of the camera at exposure: focal length, position of the principal point relative to the fiducial marks, and lens distortion — taken from the camera calibration certificate; in digital systems it is done automatically by measuring the fiducials or from the sensor geometry Relative orientation Restores the relative position and attitude of the two photos of a stereo pair so that corresponding rays intersect — i.e. all y-parallax is removed; it has 5 elements (by, bz, ω, φ, κ) and is checked at the six standard Gruber points; the result is a stereo model at an arbitrary scale and datum Absolute orientation Scales and levels the model and shifts it onto the ground coordinate system using ground control points; it has 7 elements (3 translations, 3 rotations and 1 scale) — a spatial similarity transformation Exterior orientation The position and attitude of each photo in object space:

11

X0, Y0, Z0, ω, φ, κ — 6 elements; nowadays measured directly by GNSS/IMU (direct georeferencing) or computed in the bundle adjustment Aerial Triangulation and Block Adjustment • Aerial triangulation (phototriangulation) extends control from a few ground control points to every photograph of a block, drastically reducing field survey.

12

It uses tie points (between photos of a strip and between strips) and pass points measured in the overlaps. • Methods: strip formation and polynomial strip/block adjustment (older), independent model adjustment, and the rigorous bundle block adjustment — a least-squares solution of the collinearity equations that determines all exterior orientation elements and the tie-point coordinates at once, and can include GNSS/IMU observations and self-calibration parameters. • Requirements: well-distributed GCPs (typically around the perimeter of the block and some in the interior, plus check points), good tie-point geometry and reliable image measurements; results are judged by the residuals at check points and the σ0 of the adjustment.

13

Relief and Tilt Displacement;

14

Rectification • Relief (height) displacement: a point above (or below) the datum is imaged away from (or towards) the nadir point along the radial line: d = r·h/H, where r = radial distance of the image from the nadir point, h = height of the point above datum and H = flying height above datum.

15

It is zero at the nadir point and increases outwards; it is the reason a photograph is not a map, but also allows heights to be computed: h = d·H/r. • Tilt displacement is radial from the isocentre: points on the upper (higher) side of the tilted photo are displaced inwards and those on the lower side outwards, and the photo scale varies across the photograph. • Rectification removes tilt to produce an equivalent vertical photograph (optical, analytical or digital).

16

Differential rectification (ortho-rectification) removes both tilt and relief displacement using a DEM and the exterior orientation, producing an orthophoto — an image with the geometry of a map, on which distances, angles and areas can be measured directly.

17

A true orthophoto additionally corrects building lean using a DSM.

18

Oblique Photography, Mosaics and Photo Maps • Oblique photographs cover a larger area and are easier for the layman to interpret, but have strongly varying scale; high obliques (with the horizon) suit reconnaissance and illustration, and modern multi-head oblique camera systems are used for 3-D city models and façade mapping. • Photo mosaic — several photographs joined to show a continuous area: uncontrolled (photos matched by detail only — quick, least accurate), semi-controlled and controlled (rectified/ortho photos fitted to ground control — accurate, used as a photo map).

19

Digital mosaics need seamline selection, colour and radiometric balancing and feathering. • Photo map / orthophoto map: an orthophoto mosaic with a grid, names, contours and marginal information — combining the detail of a photograph with the geometry of a map.

20

Ground Control, Pre-pointing and Post-pointing • Pre-pointing (pre-marking/signalising) — the control points are targeted on the ground before the flight (white or painted crosses/discs of a size matched to the photo scale), giving the sharpest and most reliable image points; it costs more and the targets may be disturbed before the flight.

21

Post-pointing — natural, well-defined points (corners, path junctions, small isolated features) are identified on the photographs after the flight and surveyed afterwards; cheaper but less precise and sometimes ambiguous. • Properties of an ideal ground control point: sharply and unambiguously defined and identifiable on all the photographs on which it appears; a small, well-contrasted feature at or near ground level (so that it has no relief displacement of its own); permanent and stable, and accessible for survey; free of shadow, tree cover and reflective surfaces; well distributed — around the perimeter of the block, in the corners and at intervals in the interior; not too close to the edge of a photograph; and surveyed to an accuracy better than the required map accuracy (usually by GNSS).

1

This section covers the human eye and its characteristics, stereoscopic, pseudoscopic and anaglyph vision, the application of stereo vision and parallax in photogrammetry, the digital photogrammetric process, and area-based and feature-based image matching.

2

The Human Eye and Depth Perception • The eye acts like a camera: the cornea and lens focus light through the pupil (aperture controlled by the iris) onto the retina, whose cones (colour, concentrated at the fovea) and rods (low light) convert it to nerve signals.

3

Accommodation is the change of focal length by the ciliary muscles; convergence is the inward rotation of the two eyes when viewing a near object.

4

The eye base (interpupillary distance) averages about 65 mm and the visual acuity of a normal eye is about 1 minute of arc. • Monocular depth clues: relative size, linear perspective, overlapping (occlusion), shadows, texture gradient and motion parallax.

5

Binocular (stereoscopic) perception comes from the different parallactic angle subtended at the two eyes by objects at different distances — the brain fuses the two slightly different images into one three-dimensional impression.

6

Depth perception by the eyes alone is effective up to roughly 500–600 m.

7

Stereoscopic, Pseudoscopic and Anaglyph Vision • Stereoscopic vision in photogrammetry is obtained by viewing the two overlapping photographs of a stereo pair so that the left eye sees only the left photo and the right eye only the right photo; the brain then perceives a three-dimensional stereo model.

8

Viewing aids: pocket (lens) stereoscope, mirror stereoscope, zoom stereoscopes, and in digital systems anaglyph (red-cyan) glasses, polarised glasses, active shutter glasses and autostereoscopic (glasses-free) screens. • Vertical exaggeration: the stereo model usually appears higher than reality (roughly the ratio of the air base to the flying height divided by the ratio of the eye base to the viewing distance) — helpful for interpretation, but it must be remembered when judging slopes. • Pseudoscopic vision is the reversed relief seen when the photographs are interchanged (left photo to the right eye) or rotated through 180°: hills look like valleys and valleys like hills.

9

It is avoided by orienting the photos correctly, with the shadows falling towards the observer. • Conditions for good stereoscopic viewing: the photos must have sufficient overlap, be of the same scale (within ≈ 15%), be correctly oriented along the flight line, and be free of y-parallax.

10

Parallax and Height Determination • Stereoscopic (x-) parallax of a point is the difference in its position on the two photographs of a pair, measured parallel to the flight line: p = x − x′.

11

Parallax is greater for higher points (which are nearer the camera). • For a truly vertical pair of equal flying height, with air base B and focal length f: h = H − (B·f)/p, and for the difference in height between two points, Δh = H·Δp/(pa + Δp), where pa is the parallax of the lower (datum) point and Δp the parallax difference.

12

Parallax differences are measured with a parallax bar (stereometer) under a mirror stereoscope, or digitally by image matching. • y-parallax (difference perpendicular to the flight line) prevents comfortable stereo fusion and indicates faulty relative orientation or unequal tilts; removing it is precisely the purpose of relative orientation (2.2).

13

Digital (Softcopy) Photogrammetry • Digital photogrammetry works on digital images — either scanned film or images captured directly by digital aerial, satellite or UAV cameras — on a softcopy workstation with stereo display and 3-D cursor, or through automatic processing. • Typical workflow: import images and camera calibration → build image pyramids → automatic interior orientation (fiducial or sensor model) → automatic extraction of tie points by image matching → bundle block adjustment with GCPs (and GNSS/IMU data) → dense image matching to produce a DSM/point cloud → editing to a DTM → orthophoto generation and mosaicking → stereo digitising of features → quality control and delivery.

14

The UAV equivalent is the SfM-MVS pipeline (structure from motion + multi-view stereo). • Advantages: automation and speed, no film handling, easy integration with GIS, ability to reprocess and to produce dense point clouds; requirements: large storage and computing power, good image texture for matching, and careful quality control (matching fails on water, shadow, snow and uniform surfaces).

15

Image Matching Method Principle, merits and limitations Area-based matching (ABM) A small window (template) in the left image is compared with candidate windows in the right image; similarity is measured by cross-correlation or by least-squares matching (which also solves for shape and radiometric parameters, giving sub-pixel accuracy).

16

Merits: very high accuracy on well-textured surfaces, dense results.

17

Limits: needs texture and similar geometry/radiometry — fails on repetitive patterns, steep slopes, occlusions, shadows, water and uniform areas Feature-based matching (FBM) Distinct features (interest points, edges, regions) are first extracted in each image by operators such as Förstner, Harris, SIFT or SURF, described by a descriptor and then matched between images; a robust estimator (e.g., RANSAC) removes wrong matches.

18

Merits: robust to geometric and radiometric differences, scale and rotation; works with wide baselines — the basis of automatic tie-point extraction and SfM.

19

Limits: results are sparse and must be interpolated or densified Relational / hybrid and semi-global matching Combine features and areas, or add smoothness constraints along many directions (SGM) — the standard in modern dense matching • Constraints that help matching: the epipolar constraint (the match must lie on the corresponding epipolar line, reducing the search to one dimension), approximate orientation, height range limits, image pyramids (coarse-to-fine strategy) and consistency checks (left-right matching).

1

This section covers the introduction and brief history of remote sensing, types of sensors and platforms, spatial, spectral, radiometric and temporal resolution, the electromagnetic radiation spectrum, spectral reflectance curves, and the interaction of EMR with the atmosphere and the earth's surface.

2

Definition, Components and History • Remote sensing is the science of obtaining information about an object, area or phenomenon without being in physical contact with it, by measuring the electromagnetic radiation reflected or emitted by it. • Components of the process:

3

(1) energy source or illumination, (2) radiation and the atmosphere, (3) interaction with the target, (4) recording by the sensor, (5) transmission, reception and processing, (6) interpretation and analysis, (7) application. • History:

4

1858 first aerial photograph from a balloon; aerial photo-reconnaissance in the World Wars;

5

1960 TIROS-1 (first weather satellite);

6

1972 Landsat-1 (ERTS-1) — the start of civilian earth-resources satellite imaging;

7

SPOT-1 (1986, stereo and pushbroom);

8

IRS series (India, 1988 onward);

9

IKONOS (1999) — first sub-metre commercial imagery;

10

RADARSAT and ERS radar satellites;

11

Sentinel series (ESA Copernicus, free data, Sentinel-2 from 2015); and today constellations of small satellites/CubeSats with daily coverage.

12

The Electromagnetic Spectrum and Radiation Laws Region Wavelength Remote-sensing use Gamma and X-ray < 0.03 μm / 0.03–0.3 μm Absorbed by the atmosphere; airborne gamma-ray survey of minerals Ultraviolet 0.3–0.4 μm Fluorescence, oil-slick detection; strongly scattered Visible 0.4–0.7 μm Blue 0.45–0.52, green 0.52–0.60, red 0.63–0.69 μm — photography, colour composites Near infrared (NIR) 0.7–1.3 μm Vegetation vigour, biomass, water-body delineation (water is very dark) Short-wave infrared (SWIR) 1.3–3 μm Soil and vegetation moisture, minerals, burnt areas, snow/cloud discrimination Thermal infrared (TIR) 3–5 and 8–14 μm Emitted heat — surface temperature, urban heat, forest fires, geothermal Microwave 1 mm – 1 m (bands X ≈ 3 cm, C ≈ 5.6 cm, L ≈ 24 cm, P) Radar (active) and passive microwave — all-weather, day and night; penetrates cloud and, at long wavelengths, vegetation and dry soil • Radiation laws: all objects above 0 K emit radiation (Stefan-Boltzmann: total emitted energy ∝ T⁴);

13

Wien's displacement law gives the wavelength of peak emission λmax = 2 898/T μm — so the sun (≈ 6 000 K) peaks at ≈ 0.48 μm in the visible and the earth (≈ 300 K) at ≈ 9.7 μm in the thermal infrared.

14

Interaction with the Atmosphere and the Surface • Scattering:

15

Rayleigh scattering by molecules, ∝ λ−4 — dominant for short wavelengths, it causes the blue sky and the haze that reduces contrast in blue bands;

16

Mie scattering by aerosols, dust and smoke (particles ≈ the wavelength); non-selective scattering by cloud and fog droplets (all wavelengths equally — clouds appear white).

17

Absorption by water vapour, carbon dioxide and ozone blocks certain wavelengths, leaving the atmospheric windows in which remote sensing is possible.

18

Scattered light reaching the sensor without touching the ground is path radiance, which must be removed in atmospheric correction. • At the surface the incident energy is reflected, absorbed or transmitted (EI = ER + EA + ET).

19

Reflection is specular from smooth surfaces (calm water) and diffuse (Lambertian) from rough surfaces; roughness is judged relative to the wavelength. • Spectral reflectance curves (the 'spectral signature'): healthy vegetation — low reflectance in blue and red (chlorophyll absorption), a small green peak (hence green colour), a sharp red edge and very high reflectance in the NIR (leaf structure), with water-absorption dips near 1.4 and 1.9 μm; stressed or dry vegetation loses the NIR plateau.

20

Water — moderate in blue-green, and almost zero in the NIR and SWIR, making water bodies very dark and easy to delineate; turbidity and chlorophyll raise visible reflectance.

21

Bare soil — a gradual increase with wavelength, affected by moisture (darker), organic matter, iron oxide and texture.

22

Snow — very high in the visible, low in SWIR (which distinguishes snow from cloud). • Vegetation indices exploit these differences:

23

NDVI = (NIR − Red)/(NIR + Red), ranging from −1 to +1; healthy vegetation gives high positive values (≈ 0.3–0.9), bare soil ≈ 0.1–0.2 and water negative.

24

Sensors, Platforms and Resolution • Sensors: passive (use natural solar or emitted energy — cameras, multispectral and hyperspectral scanners, thermal sensors) and active (provide their own energy — radar/SAR, LiDAR, scatterometers).

25

By construction: frame cameras, whiskbroom (across-track scanning) and pushbroom (along-track linear array) scanners. • Platforms: ground-based (field spectrometers, towers), airborne (aircraft and UAV — flexible, high resolution, small area) and spaceborne.

26

Orbits: geostationary (≈ 36 000 km above the equator, fixed over one point — meteorological satellites such as INSAT, GOES, Himawari) and sun-synchronous near-polar (≈ 700–900 km, crossing the equator at the same local solar time each pass — Landsat at ≈ 705 km with a 16-day revisit, Sentinel-2, SPOT, IRS).

27

Resolution Meaning and examples Spatial Size of the smallest resolvable ground element (pixel/GSD):

28

Landsat 8/9 multispectral 30 m (pan 15 m), Sentinel-2 10/20/60 m, MODIS 250–1 000 m, WorldView/Pléiades < 0.5 m, UAV a few centimetres Spectral Number and width of the wavebands recorded: panchromatic (1 broad band), multispectral (a few bands), hyperspectral (hundreds of narrow contiguous bands — Hyperion, PRISMA) Radiometric Number of brightness levels recorded — the bit depth:

29

8-bit = 256 levels, 12-bit = 4 096 levels (Landsat 8 is 12-bit); higher radiometric resolution shows finer differences in brightness Temporal Revisit interval:

30

Landsat 16 days, Sentinel-2 ≈ 5 days (two satellites), geostationary every 10–30 minutes, agile commercial constellations daily • There are trade-offs: a higher spatial resolution usually means a narrower swath, a longer revisit time and a larger data volume; a finer spectral resolution means less energy per band and therefore a coarser spatial resolution for the same signal-to-noise ratio.

31

The sensor is selected from the scale, detail, spectral discrimination and frequency the application demands.

1

This section covers image enhancement and the histogram, filtering, radiometric and geometric distortions and their correction, supervised and unsupervised image classification, accuracy assessment and the process of image interpretation.

2

The Digital Image and Its Histogram • A digital image is a raster of pixels, each holding a digital number (DN) per band, which represents the radiance recorded.

3

The histogram shows how many pixels have each DN — its shape reveals contrast (a narrow histogram = low contrast), brightness, saturation, bimodality (e.g., land and water) and noise, and it is the basis of most enhancement operations. • Processing is normally organised as pre-processing (restoration/correction) → enhancement → transformation → classification → accuracy assessment → presentation.

4

Radiometric Distortion and Correction • Sources: sensor response and calibration drift, striping/banding (detectors out of calibration in whiskbroom sensors), line dropout (detector failure), random noise; atmospheric effects — scattering adds path radiance (especially in blue bands) and absorption reduces the signal; variable sun elevation and topographic shading; sensor viewing geometry (BRDF). • Corrections: conversion of DN to at-sensor radiance and then to top-of-atmosphere reflectance using the sensor's calibration coefficients; atmospheric correction — simple dark-object subtraction (histogram minimum adjustment), empirical line method with field spectra, or radiative-transfer models (6S, FLAASH, Sen2Cor); sun-angle and Earth-sun distance normalisation; topographic normalisation (cosine, Minnaert or C-correction — important in the steep terrain of Nepal); destriping by histogram matching and de-noising/line replacement.

5

Geometric Distortion and Correction • Systematic (predictable) distortions: earth rotation skew, scan skew and mirror velocity variation, panoramic distortion (increasing pixel size off-nadir), map projection and earth curvature effects — corrected by the sensor model at the ground station.

6

Non-systematic distortions: variation in platform altitude, attitude (roll, pitch, yaw) and velocity, and relief displacement. • Geometric correction (rectification/georeferencing): identify well-distributed ground control points on the image and on a map/reference image → fit a polynomial transformation (first order = affine for small, flat scenes; higher orders for more complex distortion) → check the RMSE at the GCPs (usually required below 0.5–1 pixel) → resample the image into the new grid. • Resampling methods: nearest neighbour (takes the value of the closest pixel — keeps the original DN values, so it must be used before classification, but gives a blocky appearance), bilinear interpolation (weighted average of 4 pixels — smoother, alters DNs) and cubic convolution (16 pixels — sharpest image, alters DNs most). • Ortho-rectification additionally removes relief displacement using a DEM and the sensor model, and is required for imagery of hilly terrain and for high-resolution imagery.

7

Image Enhancement • Contrast enhancement (point operations): linear stretch between the minimum and maximum DN, percentage/saturation stretch, histogram equalisation (assigns more display levels to the most frequent DNs), piecewise linear stretch and density slicing (converting ranges of DN into colours). • Spatial filtering (neighbourhood operations): low-pass (smoothing) filters — mean, median (good against salt-and-pepper noise) — reduce noise and detail; high-pass filters sharpen and emphasise edges and linear features; edge detectors (Sobel, Prewitt, Laplacian) and directional filters aid lineament and road extraction; filtering may also be done in the frequency domain with the Fourier transform (useful for removing periodic striping). • Spectral transformations: band ratios (suppress topographic shading and enhance differences, e.g., NDVI), principal component analysis (removes correlation between bands and compresses information into a few components), tasselled-cap transformation (brightness, greenness, wetness), and image fusion/pan-sharpening (IHS, Brovey, PCA, wavelet) which combines a high-resolution panchromatic band with lower-resolution multispectral bands. • Colour composites: a standard false-colour composite (FCC) displays NIR as red, red as green and green as blue — healthy vegetation appears bright red, water black to dark blue and urban areas cyan/grey.

8

Image Classification Approach Description Unsupervised classification The computer groups pixels into spectral clusters first (K-means, ISODATA), and the analyst labels the clusters afterwards using field knowledge.

9

Merits: little prior knowledge needed, objective, reveals natural groupings.

10

Limits: clusters may not correspond to useful classes, and labelling can be difficult Supervised classification The analyst first defines training areas for each known class → spectral signatures are computed → every pixel is allocated by a decision rule: minimum distance to mean, parallelepiped (box), Mahalanobis distance or the widely used maximum likelihood classifier (probabilistic, assumes normally distributed classes).

11

Merits: classes are those the user wants.

12

Limits: needs good, representative and separable training data Machine learning and object-based methods Support vector machines, random forests, neural networks and deep learning (CNN) for complex non-normal data; object-based image analysis (OBIA) first segments the image into homogeneous objects and then classifies them using spectral, textural, shape and context attributes — essential for high-resolution imagery Hard vs soft classification Hard classification gives one class per pixel; soft/fuzzy or sub-pixel methods give class proportions (spectral unmixing), useful for mixed pixels Accuracy Assessment • Classification accuracy is assessed against independent reference (ground truth) data collected by field visit, GPS survey or higher-resolution imagery, using a statistically sound sample (commonly at least ≈ 50 points per class, by random or stratified random sampling). • The results are tabulated in an error (confusion) matrix: overall accuracy = correctly classified pixels ÷ total; producer's accuracy = correct pixels of a class ÷ the number of reference pixels of that class (its complement is the error of omission); user's accuracy = correct pixels of a class ÷ the number classified as that class (its complement is the error of commission). • The kappa coefficient κ = (po − pe)/(1 − pe) compares the observed agreement with the agreement expected by chance: κ > 0.8 is usually called strong agreement, 0.4–0.8 moderate, and below 0.4 poor.

13

Positional accuracy of the product is reported separately as an RMSE.

14

Visual Image Interpretation • Elements of interpretation: tone/colour, texture (smooth, coarse, mottled), shape, size, pattern (arrangement — e.g., orchards, drainage patterns), shadow (reveals profile and height, but hides detail), site (topographic position), association (related features nearby) and height/depth from stereo viewing.

15

Higher-order elements (pattern, association) require more knowledge and give more reliable identification. • Process: detection → recognition and identification → analysis and delineation → deduction and classification → accuracy check, with interpretation keys and the principle of convergence of evidence (using several elements together) supporting the interpreter, aided by collateral data and field checks.

1

This section covers the methods of terrain model generation — UAV, LiDAR, stereo imagery, high-resolution satellite imagery and the basics of microwave remote sensing, SAR and InSAR — the products DTM, DEM and DSM, and 2-D and 3-D products.

2

DEM, DTM and DSM • DEM (digital elevation model) — a general term for a digital representation of ground elevation, usually a regular grid of heights of the bare earth.

3

DTM (digital terrain model) — the bare-earth surface together with terrain features such as break lines, ridges, stream lines and spot heights.

4

DSM (digital surface model) — the top surface, including buildings, trees and other objects.

5

The difference nDSM = DSM − DTM gives building and canopy heights. • Data structures: regular grid (raster) — simple and easy to process but wasteful in flat areas;

6

TIN (triangulated irregular network) — adapts to terrain complexity and honours break lines; contours and point clouds.

7

Quality depends on point density, distribution, interpolation method and the accuracy of the source data.

8

Method of generation Characteristics Ground survey (total station, GNSS-RTK) Highest accuracy, selective points and break lines, but slow and costly — used for small areas and engineering projects Photogrammetry — stereo imagery Stereo pairs from aircraft or satellite, measured by automatic dense image matching; gives a DSM that must be filtered/edited to a DTM; needs texture and clear weather UAV photogrammetry (SfM) Low-cost, very high resolution (GSD of a few cm) over small areas; accuracy roughly 2–3 times the GSD with good GCPs, or with RTK/PPK onboard positioning; limited flight endurance, regulated airspace Airborne LiDAR (laser scanning) Active sensor: laser ranging + GNSS + IMU; records multiple returns so that pulses reaching the ground through gaps in the canopy allow the bare earth to be extracted (first return → DSM, last/ground return → DTM); vertical accuracy ≈ 5–15 cm, works day or night, high point density; expensive, and it does not see through solid roofs or dense cloud HRSI (high-resolution satellite imagery) Stereo or tri-stereo from Cartosat, WorldView, Pléiades, SPOT etc.; covers large and inaccessible areas (valuable in the Himalaya), with DEM accuracy of about 1–5 m InSAR / radar DEM from the phase difference of two SAR images; all-weather, large area — SRTM (2000) produced near-global 30 m/90 m DEMs; also ALOS PALSAR, TanDEM-X Digitised contours DEM interpolated from existing contour maps — cheap but limited by the accuracy and age of the source map Basics of Microwave Remote Sensing, SAR and InSAR • Microwave (radar) remote sensing is active — the sensor transmits its own pulses and measures the returned backscatter — so it works day and night and through cloud, haze and rain, a decisive advantage in monsoon and mountain regions.

9

X (≈ 3 cm), C (≈ 5.6 cm), L (≈ 24 cm) and P; longer wavelengths penetrate vegetation and dry soil more deeply.

10

Polarisation (HH, VV, HV, VH) carries information about the scattering mechanism. • Imaging geometry: radar is side-looking (SLAR/SAR), measuring in the slant range, which is converted to ground range; synthetic aperture radar (SAR) uses the platform's motion to synthesise a very long antenna, giving a fine azimuth resolution independent of range.

11

Range resolution depends on the pulse length/bandwidth. • Backscatter increases with surface roughness (relative to the wavelength), with the dielectric constant (hence soil and vegetation moisture — water gives a high dielectric constant), and with favourable geometry (corner reflectors in urban areas are very bright, while smooth water is very dark). • Radar-specific distortions: foreshortening (slopes facing the sensor appear compressed), layover (steep slopes imaged before their base) and radar shadow behind steep terrain — all severe in mountainous country; plus speckle, the grainy noise inherent in coherent imaging, reduced by multi-looking and speckle filters (Lee, Frost). • InSAR (interferometric SAR) combines two SAR images acquired from slightly different positions; the phase difference between them gives the terrain height, from which a DEM is produced (SRTM).

12

DInSAR (differential InSAR) compares images from different times to measure ground deformation to centimetre or millimetre level — co-seismic displacement (widely used after the 2015 Gorkha earthquake), landslide and glacier movement, land subsidence in the Kathmandu valley, and infrastructure monitoring;

13

PS-InSAR uses persistent scatterers for long time series.

14

Its limitations are temporal and geometric decorrelation (vegetation, snow) and atmospheric phase delay.

15

Ortho Products, 2-D and 3-D Products • Orthophoto/orthoimage — an image from which tilt and relief displacement have been removed with a DEM, so that it has uniform scale and map geometry; a true orthophoto also corrects building lean using a DSM, and orthomosaics are formed with seamlines and colour balancing.

16

Orthoimages are the standard base layer for GIS, map revision and cadastral work. • Derived 2-D products: contours, slope, aspect and hillshade maps, viewshed/line-of-sight and intervisibility maps, drainage networks and catchments, cut-and-fill and volume computations, flood-inundation and landslide-susceptibility maps, profiles and cross-sections. • 3-D products: point clouds, TIN surfaces, textured 3-D city models (with levels of detail, LoD 1–4), building footprints with heights, 3-D cadastre, tree canopy models, virtual flythroughs and digital twins — increasingly the standard deliverable of UAV and LiDAR surveys. • Quality is reported by the vertical RMSE of the DEM at independent check points, the horizontal RMSE of the orthoimage, point density and the completeness of filtering; accuracy standards relate these to the intended map scale and contour interval.