Agricultural Engineering · Land & Water Resources Engineering · also Remote Sensing and GIS Applications, IT for Land and Water Management
Mapping land and water from satellites and in GIS: images, indices, elevation models and watershed delineation.
In the exams: ICAR AICE subject 40 Unit 5 (RS and GIS in SWC); TNPSC Unit III; NCEES PE GIS; skill modules in the ICAR 6th Deans' Committee B.Tech
Remote sensing is the measurement of a surface from a distance, usually by recording the electromagnetic energy it reflects or emits, from a satellite, an aircraft or a drone. A geographic information system (GIS) is the combination of software, data and method used to store, analyse and map information that has a location. Global navigation satellite systems (GNSS) give that location. Together they let an agricultural engineer see a whole watershed, command area or district at once: which fields are cropped, where the land is eroding, which streams drain which slopes, how much of an irrigation scheme actually received water this season, and where a check dam or farm pond will catch the most runoff.
This manual follows the order in which the subject is taught in most agricultural engineering degrees: the physics of the electromagnetic spectrum and spectral signatures; sensors, platforms and the four kinds of resolution; visual image interpretation and aerial photo geometry; image correction and classification with accuracy assessment; vegetation and water indices; GIS data models, coordinate systems and map projections; spatial analysis; digital elevation models and watershed delineation; GNSS; applications to watershed, irrigation and erosion work; and a tool-neutral workflow for desktop GIS practice. Every worked number in the manual was computed by script, and the formulas are given so the student can repeat them.
In India the subject is examined in the ICAR AICE-JRF/SRF soil and water conservation paper and in state commission papers such as Tamil Nadu's; in the United States the NCEES PE Agricultural and Biological Engineering specification lists GIS among common system applications. Elsewhere it is a taught course and a daily professional tool. Missions, software and datums change; the principles and the arithmetic do not, and those are what examiners test.
Electromagnetic radiation (EMR) is energy that travels as waves of electric and magnetic fields at the speed of light, c, about 3 × 10^8 m/s. Wavelength (λ) and frequency (ν) are linked by c = λν, so a longer wavelength means a lower frequency. The energy of one photon is E = hν = hc/λ, where h is Planck's constant; a photon of blue light at 0.45 µm therefore carries twice the energy of a near-infrared photon at 0.9 µm. The part of the spectrum used in remote sensing runs from the ultraviolet through the visible (about 0.4–0.7 µm: blue 0.4–0.5, green 0.5–0.6, red 0.6–0.7), near-infrared (NIR, about 0.7–1.3 µm), shortwave infrared (SWIR, about 1.3–3 µm), thermal infrared (about 3–14 µm, most used in the 8–14 µm band) and microwave (about 1 mm to 1 m). Band limits are conventions and vary slightly between textbooks. Every object above absolute zero emits radiation. A blackbody is an ideal emitter; the Stefan–Boltzmann law says total emitted energy rises with the fourth power of absolute temperature (M = σT^4), so doubling temperature multiplies emission sixteen-fold. Wien's displacement law gives the wavelength of peak emission, λmax = 2898/T µm with T in kelvin. The Sun, at about 6000 K, peaks near 0.48 µm, in the visible; the Earth's surface, at about 300 K, peaks near 9.7 µm, in the thermal infrared. That single calculation explains why optical sensors record reflected sunlight and thermal sensors record emitted heat. Real surfaces emit less; the ratio is their emissivity (0–1).
Between the Sun, the ground and the sensor, the atmosphere scatters and absorbs energy. Rayleigh scattering by gas molecules much smaller than the wavelength is proportional to λ^-4, so blue light is scattered far more than red; this makes the sky blue and gives blue bands a haze that must be corrected. Mie scattering by particles of about the same size as the wavelength (smoke, dust, water vapour) affects longer visible wavelengths too. Non-selective scattering by particles much larger than the wavelength, such as cloud droplets, scatters all visible wavelengths equally, which is why clouds look white. Absorption is mainly by water vapour, carbon dioxide and ozone; ozone removes most ultraviolet. The wavelength ranges that pass through with little loss are atmospheric windows: the visible and NIR, SWIR windows near 1.6 and 2.2 µm, thermal windows at about 3–5 and 8–14 µm, and the microwave region, which passes through cloud. Sensors place their bands inside these windows. Energy reaching a surface is partly reflected, absorbed or transmitted; for any wavelength the three fractions add up to the incident energy. Reflection may be specular (calm water) or diffuse (rough surfaces). Spectral reflectance is the ratio of reflected to incident energy at a given wavelength, expressed from 0 to 1 or as a percentage.
A spectral signature, or reflectance curve, is how reflectance changes with wavelength for a material. Healthy green vegetation has low reflectance in the blue (near 0.45 µm) and red (near 0.67 µm) because chlorophyll absorbs these for photosynthesis, a small peak in the green near 0.55 µm (why leaves look green), and a steep rise between about 0.68 and 0.75 µm called the red edge. In the NIR, reflectance is high, often 40–50 percent or more, because of multiple scattering at cell-wall and air-space boundaries in the spongy mesophyll. In the SWIR, leaf water absorbs strongly near 1.4, 1.9 and 2.7 µm, so SWIR reflectance falls as leaf water content rises. Stress usually shows first as a fall in NIR reflectance and a shift of the red edge toward shorter wavelengths, then as a rise in red reflectance as chlorophyll declines. Bare soil reflectance usually rises steadily with wavelength; it is lowered by moisture, organic matter and coarse surface roughness, and iron oxides add absorption features in the visible. Clear water reflects little at all wavelengths and absorbs nearly all NIR and SWIR, so water bodies appear very dark in NIR images; turbid water and algae raise visible reflectance. These contrasts are the basis of every vegetation and water index. In the microwave region the response is different: radar backscatter depends on surface roughness relative to wavelength, geometry and the dielectric constant, which rises sharply with moisture, so wet soil returns more energy than dry soil of the same roughness.
A passive sensor records energy from a natural source: reflected sunlight in the optical bands, or energy emitted by the surface in the thermal and passive microwave bands. An active sensor sends out its own energy and measures the return, so it works by night and, in the microwave, through cloud. Radar (radio detection and ranging) is the main active microwave system; synthetic aperture radar (SAR) uses the platform's motion to simulate a long antenna and gives fine resolution from orbit. Common radar bands are X (about 3 cm), C (about 5.6 cm) and L (about 23 cm); longer wavelengths penetrate further into vegetation canopies and dry soil. LiDAR (light detection and ranging) is an active laser system that times pulses to the ground and back and is the standard source of high-accuracy elevation models. Optical sensors are built as framing cameras, which capture a whole scene at once (aerial and drone cameras), or scanners. An across-track (whiskbroom) scanner sweeps a mirror from side to side and builds the image pixel by pixel with few detectors. An along-track (pushbroom) scanner uses a linear array of detectors, one per pixel across the swath, and builds the image line by line as the platform moves; with longer dwell time per pixel it gives better signal and has no moving mirror. A multispectral sensor records a few broad bands; a hyperspectral sensor records a hundred or more narrow contiguous bands, enough to draw a near-continuous spectrum for each pixel.
Platforms range from ground-based spectroradiometers on a tripod, through drones (unmanned aerial vehicles) and manned aircraft, to satellites. A geostationary orbit is about 35,786 km above the equator; the satellite's period equals one Earth rotation, so it stays over the same point and can image the same hemisphere every few minutes, which suits weather monitoring but gives coarse ground detail. A near-polar sun-synchronous orbit, typically about 600–900 km high, is inclined slightly beyond 90 degrees so that the orbital plane turns with the Sun and the satellite crosses the equator at the same local solar time on every pass, commonly mid-morning. Constant illumination makes images from different dates comparable, which is essential for crop monitoring. Well-known public earth-observation programmes illustrate the range. The Landsat series (United States) has provided 30 m multispectral images with a 16-day repeat and a 185 km swath since the 1980s. The Sentinel-2 mission (European Union) provides 10, 20 and 60 m bands with a 290 km swath; its two satellites together revisit about every 5 days at the equator. India's Resourcesat satellites carry the LISS-III (about 23.5 m), LISS-IV (about 5.8 m) and AWiFS (about 56 m) sensors. MODIS on two US satellites gives near-daily global coverage at 250–1000 m. Specifications change as missions are replaced, so check current documentation. Drones give centimetre pixels on demand over small areas, subject to national aviation rules.
Resolution has four meanings, and exam questions deliberately mix them. Spatial resolution is the smallest ground area a sensor resolves, usually quoted as pixel size or ground sampling distance. For a nadir-looking scanner the ground resolution cell is D = H × β, where H is the altitude and β the instantaneous field of view (IFOV) in radians: at 800 km with an IFOV of 0.1 milliradian, D = 800,000 × 0.0001 = 80 m. For a digital camera the ground sampling distance is GSD = H × p / f, where p is the detector pixel pitch and f the focal length: a drone at 120 m with 4.4 µm pixels and an 8.8 mm lens gives 6 cm. Each 30 m pixel covers 900 m², or 0.09 ha; a 10 m pixel covers 0.01 ha. Spectral resolution is the number and width of bands; narrower, more numerous bands mean finer spectral resolution. Radiometric resolution is the number of brightness levels a sensor records, 2^n for n bits: 8 bits give 256 levels (0–255), 11 bits 2048, 12 bits 4096. Temporal resolution is the revisit interval over the same place; two identical satellites spaced half an orbit apart halve a 16-day repeat to 8 days. The resolutions trade off against one another: a finer pixel collects less energy, so bands must be wider or dwell longer; a wider swath gives more frequent coverage but usually coarser pixels. Field-level crop mapping needs 10–30 m or finer; district drought monitoring can use daily 250 m data.
Image interpretation is identifying objects and judging their significance from an image. It remains the way an analyst checks any automated result. Textbooks list the elements of interpretation, roughly in order of complexity. Tone or colour is the relative brightness or hue of an object. Size is judged against known objects and the image scale. Shape is the outline: canals and roads are long and regular, natural streams meander, orchards and fields have geometric boundaries. Texture is the frequency of tonal change, fine for a smooth paddy nursery and coarse for a forest canopy. Pattern is the spatial arrangement of objects, such as the regular rows of an orchard or the dendritic pattern of drainage on uniform rock. Shadow reveals height and profile but hides what lies within it. Site and association use context: a building with a large circular structure beside a canal is more likely a pumping station, and salt-affected patches are associated with low-lying, poorly drained ground. Multispectral images are viewed as colour composites by assigning three bands to the red, green and blue guns of a display. A true-colour composite puts red, green and blue bands on red, green and blue. The standard false colour composite (FCC) puts NIR on red, red on green and green on blue; because healthy vegetation reflects strongly in the NIR, it appears in shades of red, with vigorous crops bright red, while clear water appears dark blue to black, turbid water lighter blue, bare soil grey to brown or cyan, and settlements bluish grey. Doubtful areas are confirmed by field visits (ground truth).
An aerial photograph is a central (perspective) projection, not a map, so its scale changes with ground height. For a truly vertical photograph the scale is S = f/(H − h), where f is the camera focal length, H the flying height above a datum and h the ground elevation above the same datum. A camera with f = 152 mm flying at 3500 m above mean sea level over ground at 460 m gives S = 0.152/3040 = 1:20,000. Forgetting to subtract ground elevation, a common slip, gives about 1:23,000. Scale can also be found from a known distance: if a 1 km ground distance measures 5 cm on the photo, the scale is 0.05/1000 = 1:20,000. Because of central projection, objects above or below the datum are displaced radially from the principal point. Relief displacement is d = r h/H, where r is the radial distance of the displaced image point from the principal point, h the object height and H the flying height above the object's base. A tower 150 m high, imaged 80 mm from the principal point on a photo flown 3000 m above its base, is displaced 80 × 150/3000 = 4 mm outward. Displacement is zero at the principal point and grows toward the edges, which is why photo edges are not used for measurement. Tilt adds further displacement about the isocentre. An orthophoto (orthoimage) is an image from which relief and tilt displacement have been removed using a digital elevation model, so it has uniform scale and can be measured like a map.
For mapping, photographs are taken along flight strips with about 60 percent forward overlap (endlap) between successive photos and about 20–30 percent sidelap between strips, so every point appears on at least two photos. Two overlapping photos form a stereopair; viewed through a stereoscope, each eye sees one photo and the brain fuses them into a three-dimensional model. The difference in position of a point on the two photos is its parallax, and parallax differences give heights. The air base is the ground distance between successive exposure stations: a 23 cm format photo at 1:10,000 covers 2.3 km on the ground, and with 60 percent overlap the air base is 0.4 × 2300 = 920 m. Photogrammetry is the science of making measurements from photographs. Modern digital photogrammetry, including structure-from-motion software used with drones, matches thousands of points across many overlapping images to build point clouds, surface models and orthomosaics. Accurate results need ground control points (GCPs) surveyed by GNSS or total station, or an onboard high-accuracy GNSS; without them positions may be good relative to one another but shifted as a whole.
A raw image stores digital numbers (DNs), not physical quantities. Radiometric calibration converts DN to at-sensor radiance with the gain and offset in the image metadata, and then to top-of-atmosphere reflectance using the sun angle and Earth–Sun distance. Atmospheric correction removes the scattering and absorption of the atmosphere to give surface reflectance. The simplest method is dark-object subtraction: the darkest pixels in a band (deep clear water or dense shadow) should be near zero, so their value is treated as haze and subtracted. Physically based radiative transfer models do better, and many agencies now distribute surface-reflectance products ready to use. Correction matters whenever images from different dates or sensors are compared or when indices are related to field measurements. Geometric correction removes distortions from Earth rotation, curvature, platform attitude, terrain relief and sensor geometry, and places the image in a map coordinate system. In image-to-map rectification, ground control points (GCPs) identifiable on both image and map, such as road junctions and canal crossings, are used to fit a polynomial transformation. The fit is judged by the root mean square error (RMSE) of the GCP residuals; if a point's residual is 0.3 pixel in x and 0.4 pixel in y, its error is √(0.3² + 0.4²) = 0.5 pixel. A total RMSE below about half a pixel is a commonly quoted target, though project specifications vary. The image is then resampled onto the new grid. Nearest-neighbour resampling keeps original values and suits classified maps; bilinear interpolation (four neighbours) and cubic convolution (sixteen) give smoother images but alter values.
Enhancement improves how an image looks without adding information: contrast stretching spreads a band's narrow range of DNs over the full display range, filters smooth or sharpen, and principal component analysis compresses correlated bands. Classification assigns each pixel, or each object, to a thematic class such as paddy, sugarcane, fallow, water or settlement. Unsupervised classification lets an algorithm group pixels into spectral clusters without prior training; the analyst then labels each cluster. K-means assigns each pixel to the nearest of k cluster means and recalculates the means until they stabilise; ISODATA does the same but also splits clusters that are too spread out and merges those that are too close. Unsupervised methods need little prior knowledge and reveal natural groupings, but clusters do not always match the classes the user wants: one crop may split into several clusters and two different crops may fall into one.
In supervised classification the analyst first delineates training sites, areas of known class identified by field visits or reliable reference imagery, and the algorithm learns each class's spectral statistics from them. The parallelepiped classifier draws a box around each class's range in every band; it is fast but leaves pixels unclassified or ambiguous where boxes miss or overlap. The minimum-distance-to-means classifier assigns a pixel to the class whose mean is nearest in spectral space; it ignores class variability. The maximum likelihood classifier assumes each class is normally distributed, uses the mean vector and covariance matrix, and assigns each pixel to the class with the highest probability; it was long the standard and needs enough training pixels to estimate the covariance (a common rule of thumb is at least ten times the number of bands per class). Machine-learning classifiers such as random forest and support vector machines make no normality assumption and now dominate practice; they still depend on representative training data. Object-based image analysis (OBIA) first segments the image into objects, groups of adjacent similar pixels such as a field, and then classifies objects using spectral, shape, texture and context attributes. It suits high-resolution images where single pixels are smaller than the features of interest and pixel classification gives a noisy salt-and-pepper result. Multi-date classification is usually more accurate than single-date for crops, because crops that look alike on one date differ in their seasonal curves.
A classified map is worthless until its accuracy is measured against independent reference data, points not used in training. Results are summarised in an error (confusion) matrix, with map classes in rows and reference classes in columns in the most common convention; always check which way a given text arranges it. Overall accuracy is the sum of the diagonal divided by the total number of samples. Producer's accuracy for a class is its diagonal count divided by the reference (column) total; its complement is the omission error, reference pixels of that class the map missed. User's accuracy is the diagonal count divided by the map (row) total; its complement is the commission error, pixels labelled that class that are really something else. The producer asks how well the reference was mapped; the user asks how often a pixel labelled paddy on the map is really paddy. The kappa coefficient compares observed agreement with the agreement expected by chance from the row and column totals: κ = (po − pe)/(1 − pe), where po is overall accuracy and pe = Σ(row total × column total)/N². For a two-class matrix with 80 and 90 correct out of 200, rows 100 and 100 and columns 90 and 110, po = 0.85, pe = (100 × 90 + 100 × 110)/200² = 0.5, and κ = 0.35/0.5 = 0.70. Kappa is still widely taught and examined, but several researchers argue it adds little to overall, user's and producer's accuracies; current good-practice guidance favours reporting the full matrix, error-adjusted class areas and confidence intervals.
A vegetation index combines two or more bands into one number that tracks the amount or vigour of green vegetation while reducing the effect of illumination, soil and atmosphere. All the classic indices exploit the same contrast: green leaves absorb red light and scatter NIR strongly, while soil and water show much less contrast between the two. The simple ratio, SR = NIR/Red, was among the first. The normalised difference vegetation index, NDVI = (NIR − Red)/(NIR + Red), introduced in the 1970s, is by far the most widely used. It ranges from −1 to +1: water and snow are usually negative, bare soil and rock near zero to about 0.2, sparse vegetation roughly 0.2–0.4 and dense green canopies about 0.6–0.9; the exact thresholds depend on the sensor, the atmosphere and the crop and should not be treated as fixed. Computed from surface reflectance, a pixel with NIR 0.50 and red 0.08 has NDVI = 0.42/0.58 = 0.72. Reversing the numerator gives −0.72; using the simple ratio gives 6.25; both are common exam slips. NDVI should be computed from reflectance, not raw DN, if values are to be compared between dates or sensors. Its weaknesses are well known. It saturates over dense canopies, rising little once the leaf area index exceeds about 3 (the exact point varies by crop), so it poorly separates good from very good crops. It is sensitive to soil background at low cover, especially dark or wet soil, and to residual atmospheric effects.
Several indices address these weaknesses. The soil-adjusted vegetation index, SAVI = (1 + L)(NIR − Red)/(NIR + Red + L), adds a soil-brightness factor L; L = 0.5 is the usual value for intermediate cover, L = 0 turns SAVI back into NDVI. For NIR 0.30 and red 0.10, NDVI = 0.5 while SAVI = 1.5 × 0.20/0.90 = 0.33, showing how SAVI damps the response over partly bare ground. The enhanced vegetation index, EVI = G(NIR − Red)/(NIR + C1·Red − C2·Blue + L), uses the blue band to correct for aerosols and has less saturation over dense canopies; the coefficients adopted for the MODIS product are G = 2.5, C1 = 6, C2 = 7.5 and L = 1. For NIR 0.40, red 0.05 and blue 0.03, EVI = 2.5 × 0.35/(0.40 + 0.30 − 0.225 + 1) = 0.59. Red-edge indices, which use bands near 0.70–0.75 µm, are more sensitive to chlorophyll content in dense canopies and are used for nitrogen status where the sensor has red-edge bands. Vegetation indices are related empirically to leaf area index, green biomass, fraction of absorbed radiation, crop coefficients and yield. Such relationships are crop-, stage-, sensor- and region-specific; a regression fitted in one district must be recalibrated before it is used in another.
Two different indices are both called NDWI, and exam questions sometimes exploit the confusion. McFeeters' normalised difference water index, (Green − NIR)/(Green + NIR), maps open water: water reflects more green than NIR and gives positive values, while vegetation and soil are usually negative. For green 0.08 and NIR 0.02, NDWI = 0.6; for a crop with green 0.10 and NIR 0.30 it is −0.5. Gao's NDWI, (NIR − SWIR)/(NIR + SWIR), measures vegetation water content, because leaf water absorbs SWIR; it is often called the normalised difference moisture index (NDMI) to avoid confusion. A modified version of McFeeters' index replaces NIR with SWIR to suppress built-up areas. For drought monitoring, the vegetation condition index compares current NDVI with the historical range for the same place and time of year: VCI = (NDVI − NDVImin)/(NDVImax − NDVImin) × 100. With current NDVI 0.45, historical minimum 0.20 and maximum 0.70, VCI = 50 percent. Low VCI values indicate poor conditions relative to that pixel's own history, which removes much of the effect of different land cover and climate. Thermal bands add land surface temperature; a crop short of water closes its stomata, transpires less and warms, so the combination of temperature and vegetation index is used in drought indices and in surface energy balance models that estimate actual evapotranspiration.
A geographic information system stores two kinds of information about each feature: where it is (spatial data) and what it is (attribute data). A GIS has hardware, software, data, methods and people, and its value lies in combining layers that were collected separately, such as soils, slope, land use and rainfall, to answer questions none of them answers alone. There are two main ways to represent the world. The vector data model represents discrete features as points (a well, a gauging station), lines, also called arcs, made of connected vertices (a canal, a stream, a road) and polygons, closed areas (a field, a soil mapping unit, a watershed). Each feature is linked to a row in an attribute table in a relational database. In the simple, non-topological or spaghetti model, each feature is stored independently, so a boundary shared by two fields is stored twice and may not match exactly, leaving slivers and gaps. The topological model stores relationships explicitly: connectivity (which arcs meet at a node, needed for networks such as canals), adjacency (which polygons lie left and right of an arc) and containment (which features lie inside a polygon). Topology lets a GIS detect digitising errors such as undershoots, overshoots and unclosed polygons, and supports network analysis. Many modern formats store simple features and compute topology on demand.
The raster data model divides space into a regular grid of cells, each holding one value: a class code, an elevation, a reflectance or a rainfall amount. Satellite images, DEMs and continuous surfaces are naturally raster. The cell size sets the resolution: a 1 m grid needs a million cells per square kilometre, a 10 m grid ten thousand, so halving the cell size quadruples storage. Raster analysis is simple and fast because every layer shares the same grid, and overlay is cell-by-cell arithmetic. Its drawbacks are blocky boundaries, large files and loss of detail smaller than a cell. Storage can be reduced by run-length encoding, which stores a run of identical values as one value and a count, and by quadtrees, which subdivide only those quadrants that are not uniform. Vector suits discrete features with precise boundaries and networks, gives compact storage and good cartography, but overlay is computationally heavier. The choice depends on the data and the analysis, and most projects use both, converting between them by rasterisation and vectorisation. A triangulated irregular network (TIN) represents a surface by non-overlapping triangles joining irregularly spaced points, often formed by Delaunay triangulation; it places more points where terrain changes quickly and fewer on flat ground.
Locations are defined against a model of the Earth. The geoid is the equipotential surface of gravity that approximates mean sea level; it is irregular. An ellipsoid is a smooth mathematical surface fitted to the geoid; a geodetic datum fixes an ellipsoid and its position relative to the Earth. WGS 84 is the global datum used by GPS; many countries also maintain national datums, and coordinates of the same point differ between datums by metres to hundreds of metres, so a datum mismatch is a classic cause of layers not lining up. Geographic coordinates are latitude and longitude in degrees; one degree of latitude is about 111 km and one arc-second about 31 m. A map projection transforms the curved surface to a plane and always distorts something: shape, area, distance or direction. Conformal projections, such as Mercator and Lambert conformal conic, preserve local angles and shapes; equal-area projections, such as Albers equal-area conic, preserve area and suit area statistics; equidistant projections preserve distance along some lines. No projection is both conformal and equal-area. On the Mercator projection the scale factor at latitude φ is sec φ, and area is exaggerated by sec² φ: at 60 degrees, distances are doubled and areas quadrupled. The Universal Transverse Mercator (UTM) system divides the world into 60 zones, each 6 degrees of longitude wide, numbered eastward from 180°W; the zone number is floor((longitude + 180)/6) + 1, so 77°E lies in zone 43 with central meridian 75°E. The scale factor on the central meridian is 0.9996, the false easting 500,000 m, and the false northing 0 in the northern hemisphere and 10,000,000 m in the southern. Map scale is expressed as a representative fraction; 1:50,000 means 1 cm represents 500 m, so 1 cm² represents 25 ha. A large-scale map, such as 1:5000, shows a small area in detail; a small-scale map, such as 1:1,000,000, shows a large area generally.
Spatial analysis turns layers into answers. The simplest operations are queries: attribute queries select features by their values (all fields where crop = sugarcane), and spatial queries select by location (all wells within a village boundary, all fields that touch a canal). A buffer is a zone of specified distance around a feature: a 100 m buffer around a well is a circle of π × 100² = 31,416 m², or 3.14 ha; buffers around streams define riparian strips, and buffers around canals define zones for seepage studies. Overlay combines two or more layers. In vector overlay, intersect keeps only areas common to both inputs with attributes from both; union keeps all areas of both; clip cuts one layer to the outline of another without adding the second layer's attributes; erase removes the area of one layer from another. Overlay often creates sliver polygons where boundaries that should coincide were digitised slightly differently; these are removed by setting a snapping or cluster tolerance or by eliminating polygons below a size threshold. Dissolve merges adjacent polygons that share an attribute value, for example soil units into hydrologic soil groups.
In raster analysis, map algebra applies operations cell by cell and is classified by the neighbourhood involved. Local operations use only the cell's own values in one or more layers, such as adding layers or computing NDVI from two bands. Focal (neighbourhood) operations use a moving window around each cell, such as a 3 × 3 mean filter or slope calculation. Zonal operations summarise one layer within zones defined by another, such as the mean NDVI of each field or the mean slope of each sub-watershed; this is called zonal statistics. Global operations use the whole raster, such as distance from every cell to the nearest canal. Reclassification replaces ranges of values with new codes, for example slopes of 0–3, 3–8 and above 8 percent into classes 1, 2 and 3. Site suitability is usually done by weighted overlay: each criterion layer (slope, soil, land use, distance to stream) is reclassified to a common suitability score, each is given a weight, and the weighted sum ranks every cell. Weights are often set by the analytic hierarchy process (AHP), in which experts compare criteria pairwise and the weights come from the principal eigenvector of the comparison matrix, with a consistency ratio conventionally required to be below 0.1. The result must be checked in the field.
Field data are collected at points (rain gauges, soil samples, piezometers) but decisions need surfaces. Interpolation estimates values at unsampled locations. Thiessen polygons assign each location the value of the nearest sample; in hydrology they give the Thiessen weighted mean rainfall of a catchment, the sum of each gauge's rainfall times the fraction of catchment area in its polygon. Inverse distance weighting (IDW) estimates a value as the weighted mean of nearby samples, with weights 1/d^p; with power p = 2, a point 1 km from a sample of 10 and 2 km from a sample of 20 gets (10/1 + 20/4)/(1/1 + 1/4) = 12. IDW is exact at sample points and never predicts beyond the range of the data, but produces bull's-eye patterns around isolated samples. Splines fit a smooth surface through the points. Kriging is a geostatistical method that models spatial autocorrelation with a semivariogram, the average squared difference between pairs of samples as a function of their separation. The semivariogram's nugget represents measurement error and variation at scales shorter than the sampling interval, the sill is the plateau variance, and the range is the distance beyond which samples are no longer correlated. Kriging gives the best linear unbiased estimate under its assumptions and, uniquely among these methods, a map of estimation variance. Spatial autocorrelation, the tendency of nearby values to be alike, is measured globally by Moran's I, which is positive for clustering and near its expected value (slightly below zero) for a random pattern.
A digital elevation model (DEM) is a raster of ground heights. Strictly, a digital surface model (DSM) records the top of whatever is there, trees and buildings included, while a digital terrain model (DTM) records bare ground; in everyday use DEM often means either, so always check. Heights come from contour digitisation of topographic maps, ground survey, stereo photogrammetry from aerial or satellite images, radar interferometry, and LiDAR. The Shuttle Radar Topography Mission produced a near-global DEM at 1 arc-second (about 30 m) spacing; other free global DEMs come from stereo optical satellites and later radar missions. Their vertical errors are several metres and vary with terrain and cover, so they are too coarse for field-level land grading or a farm drainage design, where survey or LiDAR is needed, but well suited to delineating watersheds of a few square kilometres and larger. A radar or photogrammetric DEM over dense forest or tall crops is closer to a surface model than to bare ground. DEM quality is assessed by comparing with independent check points, as RMSE of height differences. The grid spacing must suit the landforms of interest: a narrow gully or a field bund is invisible in a 30 m DEM.
Most terrain analysis starts with slope and aspect, computed by a focal operation from each cell's 3 × 3 neighbourhood. Slope is the rate of change of height, given as degrees or percent: a rise of 12 m over 200 m horizontal is 6 percent or 3.43 degrees. Remember that percent slope is rise over run times 100, not the angle, and that 45 degrees is 100 percent. Aspect is the direction the slope faces, measured clockwise from north. Curvature separates convex from concave slopes; profile curvature affects flow acceleration and erosion, plan curvature affects flow convergence. From slope and flow accumulation come the topographic wetness index, TWI = ln(a/tan β), where a is the upslope contributing area per unit contour width and β the slope; high values mark flat, convergent sites likely to be wet, and the slope length and steepness factor (LS) of the universal soil loss equation can be computed from the same two layers. Contours can be generated from a DEM, and from a contour map the slope between two lines is the contour interval divided by their horizontal distance: two 10 m contours 2 cm apart on a 1:25,000 map are 500 m apart on the ground, a slope of 2 percent.
Automated watershed delineation follows a standard sequence. First, sinks (pits) are filled: a sink is a cell or group of cells lower than all neighbours, usually an artefact of DEM errors, which would otherwise trap flow. Real depressions such as quarries or closed basins may need to be retained deliberately. Second, flow direction is computed for each cell. The D8 method sends all flow to the one of eight neighbours with the steepest downward slope, where slope is the height drop divided by the distance, and diagonal neighbours are √2 times further away than cardinal ones. For a 30 m cell at 100 m, with its east neighbour at 97 m and its south-east neighbour at 95.5 m, the east drop is 3/30 = 0.100 and the south-east drop is 4.5/42.43 = 0.106, so flow goes south-east even though the difference between the drops is small; forgetting the √2 still gives south-east here, but it reverses the answer in other cases. A widely used coding stores direction as powers of two: east 1, south-east 2, south 4, south-west 8, west 16, north-west 32, north 64 and north-east 128. Multiple-flow-direction methods such as D-infinity share flow between neighbours and represent divergent hillslopes better. Third, flow accumulation counts the number of upstream cells draining through each cell; multiplied by the cell area it gives contributing area, so 5000 cells of 30 m drain 4.5 km². Fourth, a threshold on accumulation defines the stream network; the threshold controls how dense the network is and should be checked against mapped streams or imagery. Streams are ordered, commonly by the Strahler method, in which two streams of order n join to form order n + 1. Fifth, an outlet or pour point is placed on the stream, snapped to the highest-accumulation cell nearby, and all cells draining to it form the watershed. In very flat areas and irrigated plains, where canals and embankments control flow, DEM delineation fails and needs burned-in drainage lines or field survey.
GNSS is the general name for satellite positioning systems. The United States' GPS, Russia's GLONASS, the European Union's Galileo and China's BeiDou are global; India's NavIC (IRNSS) and Japan's QZSS are regional. Each system has three segments: the space segment of satellites, the control segment of ground stations that track the satellites and upload orbit and clock corrections, and the user segment of receivers. GPS satellites orbit at about 20,200 km in six orbital planes, with a period of about 12 hours (half a sidereal day); the baseline constellation is 24 satellites, and more are normally operating. The GPS signals include L1 at 1575.42 MHz, L2 at 1227.60 MHz and L5 at 1176.45 MHz; the L1 wavelength is c/f ≈ 19 cm. A receiver measures the travel time of each signal and multiplies by the speed of light to get a range. Because the receiver clock is not synchronised with the satellite clocks, each range includes a common clock error and is called a pseudorange. With four unknowns (x, y, z and receiver clock bias) at least four satellites are needed for a three-dimensional fix. The sensitivity is striking: a 1 microsecond timing error is about 300 m of range, and 10 nanoseconds about 3 m. Positions are computed on the WGS 84 ellipsoid (or the system's own closely aligned frame) and give ellipsoidal height, not height above mean sea level.
GNSS errors come from the satellites (orbit or ephemeris error, satellite clock error), the signal path (ionospheric delay, the largest single error for single-frequency receivers, which dual-frequency receivers largely remove because the delay depends on frequency; tropospheric delay, from water vapour and dry air, modelled but not removed by dual frequency), and the receiver environment (multipath, where signals reflect off buildings, water or vehicles before arriving; receiver noise). The combined range error is the user equivalent range error (UERE). Geometry multiplies it. Dilution of precision (DOP) describes how satellite geometry amplifies range errors: satellites spread across the sky give low DOP, satellites clustered in one part of the sky give high DOP. Position error is roughly DOP × UERE, so a PDOP of 2.5 with a 4 m UERE gives about 10 m. PDOP covers three-dimensional position, HDOP horizontal, VDOP vertical, TDOP time and GDOP all together. Vertical accuracy is usually worse than horizontal because all satellites are above the horizon. Planning tools predict DOP so precise work can avoid poor windows.
Differential GNSS (DGNSS) uses a reference receiver at a known point to compute corrections that are sent to nearby rovers, removing most of the errors common to both, which gives sub-metre to metre accuracy with code measurements. Satellite-based augmentation systems (SBAS) broadcast wide-area corrections from geostationary satellites; examples are WAAS (United States), EGNOS (Europe), GAGAN (India) and MSAS (Japan). Real-time kinematic (RTK) positioning uses carrier-phase measurements from a base and rover, resolves the integer number of wavelengths, and gives centimetre accuracy over baselines typically up to a few tens of kilometres; corrections may come from the user's own base or from a network of continuously operating reference stations (CORS), often streamed over the internet. RTK underpins auto-steer tractors, laser-free land grading control, drain layout and precise topographic survey. Precise point positioning (PPP) uses precise satellite orbits and clocks from global services with a single receiver and converges to decimetre or centimetre accuracy after a period that can be many minutes. Heights need care. GNSS gives ellipsoidal height h. Orthometric height H, the height above the geoid used on maps, is H = h − N, where N is the geoid undulation (geoid height above the ellipsoid) from a geoid model. If h = 250 m and N = −60 m, H = 310 m; adding N instead gives the wrong answer of 190 m. Geoid undulations range over about ±100 m worldwide, so ignoring N is never acceptable for water work, where flow depends on true height differences.
Remote sensing and GIS are now the standard way to characterise and prioritise watersheds for soil and water conservation programmes. A DEM gives the boundary, drainage network, stream orders, slope classes and relief; imagery gives land use and land cover; soil maps and field surveys give soil texture, depth and hydrologic soil group. Linear, areal and relief morphometric parameters are computed from the network. Drainage density is total stream length divided by area: 84 km of streams in 42 km² is 2.0 km/km²; high values indicate impermeable, sparsely vegetated or steep terrain and quick runoff. Stream frequency is the number of stream segments per unit area. The bifurcation ratio is the number of streams of one order divided by the number of the next higher order, for example 24 first-order and 6 second-order streams give 4; values of about 3 to 5 are typical where geology does not strongly control drainage. Shape indices use area A, basin length Lb and perimeter P: form factor A/Lb², circularity ratio 4πA/P² and elongation ratio (2/Lb)√(A/π). A 42 km² basin 10 km long with a 30 km perimeter has form factor 0.42, circularity 0.59 and elongation 0.73; more circular basins tend to produce sharper, higher peak flows than elongated ones of the same area. Sub-watersheds are then ranked by combining morphometric, erosion and land-use indicators. The parameters and weights used vary between national guidelines.
The universal soil loss equation and its revised form (USLE/RUSLE), A = R × K × LS × C × P, fit naturally into a raster GIS because each factor can be mapped as a layer and multiplied cell by cell. R, the rainfall erosivity, is interpolated from rain-gauge records; K, the soil erodibility, comes from soil maps and laboratory data; LS is computed from the DEM using flow accumulation and slope; C, the cover-management factor, is assigned from a land-use classification or estimated from NDVI; P, the support practice factor, comes from field knowledge of contour farming, strip cropping and terracing. For a cell with R = 500, K = 0.3, LS = 1.2, C = 0.4 and P = 1, A = 72 in the units implied by the factors, commonly t/ha/yr in SI. Two cautions matter. First, empirical formulas that relate C to NDVI are region-specific; a published formula should be checked against tabulated C values for local crops. Second, USLE estimates average annual sheet and rill erosion; it does not predict gully erosion, deposition or sediment yield at the outlet, which need a sediment delivery ratio or a process model. The result is best used to rank areas and target treatments, not as an exact figure.
In irrigation schemes, satellite images map the area actually cropped and irrigated each season, which can be compared with the design command area and with water released to show where water is not reaching. Crop-type classification by season gives the crop mix needed for water demand. Surface energy balance models driven by thermal and optical bands estimate actual evapotranspiration pixel by pixel; combined with yield, they give water productivity maps and identify under-performing blocks. Relations between vegetation indices and crop coefficients let crop water demand be estimated from images, again with local calibration. Salt-affected and waterlogged lands are mapped from salt crusts, persistent surface water and stressed vegetation, with field sampling of soil electrical conductivity. Open water bodies are mapped with NIR or water indices; radar maps floods and paddy fields through monsoon clouds, since calm water is a specular reflector and appears dark. Runoff estimation combines land use from imagery with hydrologic soil groups in a GIS to assign a curve number to each cell or sub-catchment; with P = 100 mm and CN = 80, S = 25400/80 − 254 = 63.5 mm, initial abstraction 0.2S = 12.7 mm, and runoff Q = (P − 0.2S)²/(P + 0.8S) = 50.5 mm. Site selection for farm ponds, check dams and percolation tanks uses weighted overlay of slope, stream order, soil, land use and distance criteria, followed by field verification. Groundwater prospect maps combine geology, geomorphology, lineaments and drainage density, but they indicate likelihood, not a guarantee; only drilling or geophysical survey confirms an aquifer.
Open-source and commercial desktop GIS share the same ideas, so the workflow here applies to both. Start by creating a project folder with separate sub-folders for raw data, processed data and outputs, and never overwrite raw data. Choose one projected coordinate reference system for analysis, normally the local UTM zone or a national grid, because distances and areas cannot be measured correctly in latitude and longitude degrees. Most desktop GIS reproject layers on the fly for display, which hides mismatches; for analysis, reproject each layer explicitly to the project system and confirm that the datum transformation used is the right one. Check every layer's metadata: source, date, scale or resolution, datum and units. Load a basemap or recent imagery and confirm that layers line up; an offset of a few hundred metres usually means a datum or projection error, while a shift of exactly one zone or a reflected layer points to wrong zone numbers or swapped x and y. Georeference scanned maps with at least the minimum ground control points for the chosen transformation (three for first-order affine), preferably many more spread over the whole sheet, and record the RMSE.
A typical practical exercise is to delineate a small watershed and prepare its planning maps. Download a DEM covering the area with a margin beyond the likely divide, clip it, and reproject it to the project UTM zone, using bilinear resampling for continuous elevation. Fill sinks, compute D8 flow direction and flow accumulation, and extract streams using a threshold chosen by comparing the result with imagery; record the threshold. Place the outlet at the proposed structure, snap it to the highest-accumulation cell within a small distance, delineate the watershed and convert it to a polygon. Compute area, perimeter, basin length, stream orders and morphometric ratios. Derive slope in percent, reclassify it into the slope classes used by the local land capability or conservation guidelines, and use zonal statistics to report the area of each class. Classify a recent image into land-use classes, or digitise them over high-resolution imagery, and check accuracy against field points. Overlay land use with hydrologic soil groups to assign curve numbers and compute an area-weighted curve number. Build the RUSLE factor layers and multiply them to map erosion classes. Finish with maps that carry title, legend, scale bar, north arrow, grid, source, date and projection.
Measuring areas in a geographic coordinate system gives meaningless square degrees. Mixing datums shifts layers by tens or hundreds of metres. Resampling a classified map with bilinear or cubic interpolation creates class codes that do not exist; use nearest neighbour. Running an overlay on layers of very different scales produces spurious precision, since a 1:250,000 soil map cannot tell anything about individual fields however finely it is rasterised. Reporting areas to the square metre from 30 m pixels overstates accuracy. Accepting a default stream threshold without checking it produces networks that are too dense or too sparse. Good practice is to keep a processing log or model, so every step and parameter can be repeated; to use scripts or graphical models for repeated tasks; to state accuracy and limitations on every map; and to verify key results in the field. National rules on restricted maps and drone flights differ and must be respected.
A student uses this subject in two ways: to pass examinations that ask for definitions, the four resolutions, index formulas, error-matrix accuracies, projection facts and GNSS principles, and to complete practical assignments in a desktop GIS. The most useful habits to build early are checking coordinate systems before any measurement, computing indices from reflectance rather than raw numbers, keeping raw data untouched, and recording every processing step. A student who can delineate a small watershed from a free DEM, classify a season's image into land-use classes, and report accuracy honestly has the core skill that employers in soil conservation, irrigation and agricultural consultancy look for.
An engineer five years into practice in a soil conservation, watershed or irrigation department uses GIS weekly: delineating catchments for check dams and farm ponds, estimating runoff with curve numbers, preparing erosion-risk maps to target treatments, and verifying scheme works against high-resolution imagery. RTK GNSS is used to survey structure sites and drainage lines. The key decisions are which data are good enough for the job (a 30 m DEM for siting, a field survey for design), which stream threshold and outlet to use, and how to present results so a sanctioning officer can trust them. Field checking of every proposed site remains non-negotiable.
A senior engineer with twenty-five years of experience rarely runs the software but decides what the organisation measures and trusts. That means setting data standards (projection, datum, metadata and accuracy thresholds), choosing between satellite, drone and ground survey for a programme, approving prioritisation criteria and weights for watershed or command-area investments, and using seasonal irrigated-area and evapotranspiration maps to question scheme performance. The senior engineer also judges when a map is being over-interpreted, for example district-scale soil data used for field decisions or an NDVI trend taken as a yield figure, and ensures staff are trained and that maps meet legal and security rules.
A researcher with a PhD and long experience works at the method frontier: calibrating index–crop-coefficient and index–yield relationships for local crops, validating surface energy balance evapotranspiration against flux or lysimeter data, combining radar and optical time series for crop and soil-moisture mapping, testing DEM sources for small-watershed hydrology, and designing statistically sound accuracy assessments with proper area estimation. Such a researcher publishes uncertainty along with maps, questions default parameters, contributes to open datasets and code, and advises governments on which satellite products are fit for which policy decision.
This manual does not yet cover: radar interferometry and polarimetry beyond the basics; hyperspectral analysis methods; the algorithms of specific surface energy balance models; geostatistics beyond the semivariogram and kriging concepts; spatial databases and web mapping services; programming for GIS (scripting languages and cloud processing platforms); drone flight planning and aviation law, which differ by country; and the survey and restricted-map rules of individual countries. Mission specifications quoted (pixel sizes, swaths, repeat cycles) were correct for the missions named at the time of writing but change as satellites are replaced. Numerical thresholds for NDVI classes, C-factor formulas and index–crop-coefficient relations are deliberately not given as fixed values because they vary by sensor, crop and region.
Original practice questions written to the syllabus and pattern of the exams named in each question's examTags. They are not past papers and do not reproduce any board's questions.
These notes are free to read here, with no account.
What the Laisha app adds:
Free for ten days, then a subscription. We would rather write that here than have you find it later.