Sustainability & Environment (Agri-Food Systems) · Natural Resources & Biodiversity · also Geoinformatics for Natural Resource Management, Earth Observation, Spatial Analysis for Sustainability
Watching land, water, vegetation and carbon from space and analysing it spatially: sensors, indices, classification, change detection and the open datasets used in sustainability work. Precision-agriculture uses of the same tools are in the crop sciences.
In the exams: ICAR AICE subject 71 Unit 3 (RS and GIS for ecosystem analysis); ICAR AIEEA-PG Physical Science Unit IV; UGC NET Unit 9
Remote sensing is the measurement of the Earth's surface from a distance, almost always by recording electromagnetic radiation reflected or emitted by land, water, vegetation and the atmosphere. A geographic information system (GIS) is the set of data models, software and methods used to store, analyse and map anything that has a location. Together they let a forest agency count hectares lost each month, a drought early-warning unit compare this season's vegetation with twenty years of records, a watershed planner delineate catchments from an elevation model, and a carbon project estimate biomass across land no field crew could ever walk.
This manual follows the workflow of an environmental analyst. It starts with the physics of radiation and spectral signatures, describes the sensors and platforms that collect data (optical, thermal, radar, LiDAR and drones), and then covers pre-processing, spectral indices, classification and accuracy assessment, and change detection. It moves to GIS data models, map projections and spatial analysis, digital elevation models and watershed delineation, and the applied themes that dominate sustainability work: deforestation, drought and crop condition monitoring, and biomass and carbon estimation. It closes with the open data archives and cloud platforms that made planetary-scale analysis routine, and with satellite positioning and field data collection, without which no map can be validated.
Precision-agriculture uses of the same tools (variable-rate application, yield mapping) are taught in the crop sciences; forest inventory with LiDAR is developed further in the Forestry track. Here the emphasis is environmental monitoring at landscape to global scale.
Examiners worldwide test this subject in three ways: recall of definitions and sensor characteristics (resolutions, bands, wavelengths), short calculations (NDVI and other indices, Wien's law, ground sampling distance, accuracy from an error matrix, map scale, slope, biomass to carbon dioxide), and interpretation of cases (which sensor for cloudy tropics, which index for open water, why kappa is no longer recommended). Postgraduate entrance and fellowship examinations in environmental science, national research-service examinations, earth-science lecturer-eligibility tests and the environmental-analyst competitions of several national agencies all include this material.
Electromagnetic radiation travels as waves at the speed of light, c ≈ 3 × 10⁸ m s⁻¹. Wavelength (λ) and frequency (ν) are linked by c = λν, so a radar operating at 5.4 GHz has a wavelength of about 5.6 cm. The energy of a photon is E = hν = hc/λ, where h is Planck's constant; short wavelengths carry more energy per photon, which is one reason thermal and microwave sensors need larger detectors or longer dwell times to collect a usable signal. Remote sensing uses several regions of the spectrum: ultraviolet (below 0.4 µm), visible (about 0.4 to 0.7 µm: blue, green, red), near infrared (NIR, about 0.7 to 1.3 µm), shortwave infrared (SWIR, about 1.3 to 3 µm), thermal infrared (about 3 to 14 µm, with the most-used window between 8 and 14 µm) and microwave (about 1 mm to 1 m). Every object above absolute zero emits radiation. A blackbody, an ideal emitter, follows Planck's law; integrating over all wavelengths gives the Stefan–Boltzmann law, M = σT⁴, with σ = 5.67 × 10⁻⁸ W m⁻² K⁻⁴, so doubling absolute temperature raises emitted energy sixteen-fold. Wien's displacement law gives the wavelength of peak emission: λmax = 2898 µm K / T. The Sun, at about 5800 K, peaks near 0.5 µm, in the visible; the Earth's surface, at about 300 K, peaks near 9.7 µm, in the thermal infrared. That is why daytime optical sensors measure reflected sunlight and thermal sensors measure emitted heat. Real surfaces are not blackbodies: emissivity (ε, between 0 and 1) scales their emission, and water and dense vegetation have emissivities close to 0.98 or above, whereas dry bare soils and rocks vary more.
Before radiation reaches a satellite it passes through the atmosphere twice (for reflected sunlight). Gases absorb in particular bands: ozone in the ultraviolet, water vapour and carbon dioxide in parts of the infrared. Regions with little absorption are atmospheric windows, and sensor bands are placed in them. Scattering also matters. Rayleigh scattering by molecules much smaller than the wavelength is strongest at short wavelengths (it varies with λ⁻⁴), which makes the sky blue and the blue band hazy. Mie scattering by aerosols of similar size to the wavelength affects visible and NIR; non-selective scattering by cloud droplets affects all optical wavelengths equally, which is why clouds are white and why optical sensors cannot see through them. Microwaves pass through cloud, which is the main reason radar is used in persistently cloudy regions. A spectral signature is the pattern of reflectance across wavelengths. Healthy green vegetation absorbs strongly in the blue and red because of chlorophyll, reflects a little more in the green (around 0.55 µm), rises sharply at the red edge (about 0.68 to 0.75 µm) and reflects strongly in the NIR because the spongy mesophyll scatters light; in the SWIR, leaf water absorbs, with deep absorption features near 1.45 and 1.94 µm. Stressed or senescent leaves lose chlorophyll, so red reflectance rises and the red edge shifts to shorter wavelengths. Clear water reflects a little in blue and green and almost nothing in NIR and SWIR; turbid water reflects more in the red; algal blooms show a chlorophyll signature. Bare soil reflectance generally rises gradually with wavelength and falls as moisture or organic matter increases. Snow is bright in the visible and dark in the SWIR, which separates it from cloud.
Every sensor is described by four resolutions. Spatial resolution is the ground size of one pixel; spectral resolution is the number and width of bands; radiometric resolution is the number of brightness levels a detector records (8 bits give 2⁸ = 256 levels, 12 bits give 4096); temporal resolution is the revisit interval. Improving one usually costs another: a sensor that images a 2300 km swath every day, such as the Moderate Resolution Imaging Spectroradiometer (MODIS, 250 m to 1 km pixels, 36 bands) cannot also have 10 m pixels. Passive sensors record sunlight or emitted heat; active sensors send their own energy (radar, LiDAR). Multispectral sensors record a few broad bands; hyperspectral sensors record a hundred or more narrow contiguous bands, enough to resolve absorption features of minerals and pigments. The longest-running civilian record is the Landsat series, operating since 1972; its current Operational Land Imager instruments record 30 m multispectral bands, a 15 m panchromatic band and, through a separate thermal instrument, thermal bands acquired at 100 m, with a 16-day revisit per satellite and 8 days when two are flying. The Sentinel-2 mission of the European Copernicus programme carries a multispectral instrument with 13 bands at 10, 20 and 60 m, including three red-edge bands, and a revisit of about 5 days at the equator with two satellites. Satellites in sun-synchronous polar orbits cross the equator at the same local time on every pass, so illumination is comparable between dates; geostationary satellites, about 36,000 km above the equator, stare at one hemisphere and image it every few minutes at coarse resolution, ideal for weather, fire and cloud. Thermal sensors measure brightness temperature, which, after emissivity and atmospheric correction, gives land surface temperature, used for urban heat islands, evapotranspiration and drought stress.
Synthetic aperture radar (SAR) is an active, side-looking microwave system that synthesises a long antenna from the platform's motion to obtain fine resolution. Common bands are X (about 3 cm), C (about 5.6 cm), L (about 23 cm) and P (about 70 cm); longer wavelengths penetrate deeper into canopies and soil. SAR measures backscatter, controlled by surface roughness, geometry and dielectric constant (mostly water content). Smooth water reflects energy away and looks dark; rough surfaces and vegetation volume scatter look brighter; buildings form bright double-bounce returns. Polarisation describes the orientation of the transmitted and received waves: like-polarised (HH, VV) returns are dominated by surface scattering, cross-polarised (HV, VH) by volume scattering in vegetation. SAR images carry speckle, a granular noise reduced by multilooking or filtering, and terrain distortions called foreshortening, layover and radar shadow. Interferometric SAR (InSAR) compares phase between acquisitions to measure elevation or millimetre-to-centimetre ground deformation. The Sentinel-1 constellation provides free C-band SAR worldwide; L-band missions and the P-band BIOMASS mission, launched in 2025, target forests. Light detection and ranging (LiDAR) times laser pulses: range = c × t / 2, because the pulse travels out and back. Discrete-return systems record several returns per pulse; full-waveform systems record the whole returned energy profile. From the cloud of points, the lowest returns give a digital terrain model and the highest a surface model; their difference is canopy height. Airborne LiDAR is the reference for forest structure; spaceborne LiDAR such as the Global Ecosystem Dynamics Investigation (GEDI) on the International Space Station samples canopy height in footprints about 25 m across. Unmanned aerial vehicles (UAVs, drones) carry cameras, multispectral, thermal or LiDAR sensors at centimetre resolution. Ground sampling distance (GSD) = detector pixel size × flying height / focal length. Overlapping photographs processed by structure-from-motion photogrammetry yield orthomosaics and 3D surface models. Flight rules, privacy and payload limits vary by country.
Raw images store digital numbers (DN), not physical quantities. Radiometric calibration converts DN to at-sensor spectral radiance with a linear gain and offset supplied in the image metadata: L = gain × DN + offset. Radiance depends on the sun's position, so it is converted to top-of-atmosphere (TOA) reflectance: ρTOA = π L d² / (ESUN cos θs), where d is the Earth–Sun distance in astronomical units, ESUN is the mean solar exoatmospheric irradiance in that band and θs the solar zenith angle. TOA reflectance still includes the atmosphere's contribution: path radiance from scattering adds brightness, especially in the blue, while absorption and attenuation reduce it. Atmospheric correction estimates surface reflectance. The simplest method, dark-object subtraction, assumes the darkest pixels in a scene (deep clear water, dense shadow) should have near-zero reflectance and subtracts their value from the band. Physically based methods use radiative-transfer models fed with aerosol optical depth, water vapour and elevation. Most space agencies now distribute analysis-ready surface-reflectance products (Level-2), so many analysts never perform this step themselves, but they must know which level they are using: mixing TOA and surface reflectance in one time series creates false trends. Clouds, cloud shadows, haze and snow must be masked, usually with the quality band shipped with each product. Thermal data are converted from radiance to brightness temperature by inverting Planck's law and then corrected for emissivity to give land surface temperature. Radar data are calibrated to backscatter coefficient (sigma-nought or gamma-nought, often expressed in decibels), filtered for speckle and terrain-flattened. Sensor-to-sensor differences in band width and position mean that an index from one sensor is not identical to the same index from another; harmonised products adjust for this.
Geometric errors come from platform motion, Earth rotation and curvature, sensor view angle and relief displacement. Georeferencing ties image coordinates to map coordinates using ground control points (GCPs): features identifiable in both the image and a reference such as a map, orthophoto or survey. A polynomial transformation is fitted to the GCPs; a first-order (affine) transformation needs at least three points, a second-order needs at least six. The fit is reported as root-mean-square error (RMSE); a common working target is an RMSE below half a pixel, though the acceptable value depends on the purpose. In rugged terrain a polynomial is not enough, and orthorectification uses a digital elevation model and the sensor model to remove relief displacement. After the transformation, output pixel values are assigned by resampling: nearest neighbour copies the nearest input value and preserves original values (best before classification), bilinear interpolation averages four neighbours, and cubic convolution uses sixteen, producing smoother images but altering values. Co-registration aligns two images of the same place; misregistration of even one pixel creates spurious change along edges. Topographic correction reduces the brightness difference between sunlit and shaded slopes, with methods such as the cosine and C-correction. Enhancement improves visual interpretation without changing the data used for measurement: contrast stretching, histogram equalisation, false-colour composites (the standard false-colour composite shows NIR as red, red as green and green as blue, so vegetation appears red), band ratios, principal component analysis to compress correlated bands, and pan-sharpening, which fuses a fine panchromatic band with coarser multispectral bands. Filtering with moving windows smooths noise (low-pass) or sharpens edges (high-pass).
A spectral index combines bands to emphasise one property and suppress others such as illumination. The Normalized Difference Vegetation Index, NDVI = (NIR − Red) / (NIR + Red), exploits the contrast between chlorophyll absorption in the red and mesophyll scattering in the NIR. It ranges from −1 to +1: water and snow are usually negative, bare soil and rock are near 0 to 0.2, sparse vegetation lies roughly between 0.2 and 0.5 and dense green canopies above about 0.6, although the exact thresholds depend on the sensor, the soil and whether surface or TOA reflectance was used. NDVI is related to the fraction of absorbed photosynthetically active radiation, leaf area and green biomass, but the relation saturates over dense canopies with high leaf area index, and NDVI is sensitive to soil background and atmospheric effects. The Enhanced Vegetation Index, EVI = 2.5 × (NIR − Red) / (NIR + 6 × Red − 7.5 × Blue + 1), uses the blue band to correct for aerosols and a canopy background term to reduce soil effects; it saturates less in dense forests. The Soil-Adjusted Vegetation Index, SAVI = (1 + L)(NIR − Red) / (NIR + Red + L), adds a soil brightness factor L, commonly 0.5 for intermediate cover, to reduce the soil influence where cover is sparse, as in drylands. Red-edge indices, possible with sensors that have red-edge bands, respond to chlorophyll content in dense canopies. All indices from reflectance must use the same reflectance level and comparable band definitions for comparison across dates or sensors. A time series of NDVI or EVI traces phenology: green-up, peak, senescence and length of season, and integrated NDVI over a season is a common proxy for productivity.
Two different indices carry the name NDWI, and confusing them is a classic error. McFeeters (1996) defined NDWI = (Green − NIR) / (Green + NIR) to delineate open water: water reflects more green than NIR, so its values are positive, while vegetation and soil are negative. Gao (1996) defined NDWI = (NIR − SWIR) / (NIR + SWIR) to measure liquid water in vegetation canopies; the SWIR band is absorbed by leaf water, so wetter canopies have higher values. To avoid confusion, Gao's index is often called the Normalized Difference Moisture Index (NDMI) when computed with Landsat SWIR1. The Modified NDWI, MNDWI = (Green − SWIR) / (Green + SWIR), proposed by Xu (2006), replaces NIR with SWIR to separate water from built-up land, which can show positive McFeeters NDWI. For fire, the Normalized Burn Ratio, NBR = (NIR − SWIR2) / (NIR + SWIR2), falls sharply after fire because burning removes green leaves (lower NIR) and exposes dry, charred surfaces (higher SWIR). Burn severity is mapped with the difference dNBR = NBR before − NBR after; larger positive values indicate more severe burns, and field-calibrated thresholds are needed because the relation varies by ecosystem. Other indices include the Normalized Difference Snow Index (Green − SWIR)/(Green + SWIR) and built-up indices that use SWIR and NIR. Water indices are applied with a threshold, often zero, adjusted locally by inspecting histograms or using automatic thresholding such as Otsu's method. Shadows, dark asphalt and turbid water confuse simple thresholds, so analysts combine indices, slope masks and radar where possible.
Classification assigns each pixel or object to a thematic class such as forest, cropland, water or settlement. In unsupervised classification the algorithm groups pixels into spectral clusters without training data, and the analyst then labels the clusters; k-means and ISODATA (which can split and merge clusters) are the standard examples. In supervised classification the analyst provides training samples of known classes and the algorithm learns their signatures. Classical parametric classifiers include minimum-distance-to-means, the parallelepiped classifier and maximum likelihood, which assumes each class has a normal distribution in feature space and assigns a pixel to the class with highest probability. Machine-learning classifiers such as random forests, support vector machines and, increasingly, deep neural networks make no distributional assumption, handle many input layers (multi-date bands, indices, radar, terrain) and are now the norm. Random forests also report variable importance and an out-of-bag error, which is not a substitute for independent accuracy assessment. Object-based image analysis (OBIA) first segments the image into homogeneous objects and then classifies objects using spectral, shape, texture and context information; it suits very high resolution imagery where single pixels are smaller than the features of interest. Good training data matter more than the algorithm: samples should represent the variability of each class, be spread across the image, be checked against high-resolution imagery or field data, and be kept separate from validation samples. Class definitions must be mutually exclusive and exhaustive, ideally following a standard legend such as the FAO Land Cover Classification System, so that maps from different producers can be compared. Mixed pixels, where one pixel contains several covers, are handled by spectral unmixing or by mapping continuous fractions instead of discrete classes.
A map is not finished until its accuracy is assessed against reference data of higher quality, collected with a probability sampling design. The error (confusion) matrix cross-tabulates map classes against reference classes. Overall accuracy is the sum of the diagonal divided by the total number of samples. For each class, producer's accuracy is the number correctly classified divided by the reference total for that class; its complement is omission error (reference pixels of the class that the map left out). User's accuracy is the number correctly classified divided by the map total for that class; its complement is commission error (pixels the map labelled as the class that are actually something else). Producer's accuracy tells the map maker how well a class was captured; user's accuracy tells a map user how reliable a label is on the ground. The kappa coefficient, once standard, compares agreement with chance agreement; authoritative good-practice guidance (Olofsson and colleagues, 2014) and critiques (Pontius and Millones, 2011) now advise against relying on it because it is redundant with overall accuracy and its chance correction is not meaningful for map comparison. Good practice instead reports the full error matrix expressed as estimated area proportions, user's, producer's and overall accuracies with confidence intervals, and area estimates adjusted for the map errors. Because rare classes such as forest loss occupy a small share of a landscape, stratified random sampling by map class with enough samples in each stratum is used, and the stratified estimator corrects area totals. Simply counting mapped pixels (pixel counting) gives biased area estimates whenever omission and commission errors are unequal. Reference data should be interpreted blind to the map label, from imagery of the same date or field visits, with a documented response design.
Land cover is the physical material at the surface (trees, grass, water, impervious surface); land use is what people do with the land (forestry, grazing, cropping, conservation). A single cover can host several uses, and the same use can sit on different covers, so land use often needs ancillary data beyond images. Change detection identifies differences in the state of an object or phenomenon by observing it at different times. Its success depends on images of the same season (to avoid confusing phenology with change), the same or harmonised sensors, accurate co-registration and consistent radiometric processing. Bitemporal methods compare two dates. Image differencing subtracts one date's band or index from another's and thresholds the result, often at a chosen number of standard deviations from the mean difference. Image ratioing divides one date by another. Change vector analysis treats each pixel as a vector in multi-band space and computes the magnitude (how much change) and direction (what kind of change) between dates. Principal component analysis of stacked multi-date bands places stable variance in the first components and change in later ones. These pixel-based methods say where change happened but not what it changed from and to. Post-classification comparison classifies each date independently and cross-tabulates the maps, giving a from–to transition matrix; it is intuitive and robust to radiometric differences, but its errors compound: if two maps are each 85 percent accurate and their errors are independent, the expected accuracy of the change map is roughly 0.85 × 0.85, about 72 percent. Direct classification of the stacked two-date image avoids part of this error.
Dense archives now allow every available clear observation to be used rather than two chosen dates. Time-series algorithms fit a model of seasonality and trend to each pixel and flag breaks when new observations depart from the model. Well-known families include algorithms that decompose a series into trend, season and remainder and detect breakpoints; temporal segmentation of annual composites to capture both abrupt disturbance and gradual recovery; and continuous change detection that updates harmonic models with each new image. They separate abrupt events (clearing, fire, flood) from gradual trends (degradation, greening, recovery) and date them. Remote sensing of land change also requires attention to what is being measured: forest loss in a tree-cover product is the removal of canopy, which includes harvest in plantations and natural disturbance, and is not the same as deforestation in the land-use sense, which is conversion of forest to another use. Annual change matrices summarise gross gains and losses for each class; net change hides large gross flows, so both are reported. Rates are often expressed as an annual percentage. The compound annual rate of change between areas A1 and A2 over t years is r = (1/t) ln(A2/A1) × 100 (Puyravaud's formula), which is preferred to the simple average because it accounts for compounding. Results feed land-use models, greenhouse-gas inventories (activity data for land-use, land-use change and forestry), urban planning and conservation reporting. Accuracy assessment of change maps follows the same good-practice principles as single-date maps, with special care because change classes are rare.
A GIS represents the world in two main data models. The vector model uses points (wells, sample plots), lines (roads, streams) and polygons (fields, protected areas), each with attributes in a table; topology records spatial relationships such as adjacency and connectivity and allows checks for gaps and overlaps. The raster model divides space into a grid of cells, each holding a value; it suits continuous surfaces such as elevation, temperature and reflectance, and makes overlay arithmetic simple. A triangulated irregular network (TIN) represents surfaces with triangles whose density follows terrain complexity. Data are stored with metadata describing lineage, accuracy and coordinate reference system (CRS). A geographic CRS locates points by latitude and longitude on an ellipsoid tied to a datum; the World Geodetic System 1984 (WGS 84) is the global standard used by GNSS. A projected CRS transforms the curved surface to a plane, and every projection distorts at least one of area, shape, distance or direction. Conformal projections, such as Mercator and transverse Mercator, preserve local angles and shapes; equal-area projections, such as Albers conic and Lambert azimuthal equal-area, preserve area and are required when measuring hectares over large regions; equidistant projections preserve distance from one or two points. The Universal Transverse Mercator (UTM) system divides the world into 60 zones, each 6° of longitude wide, uses a scale factor of 0.9996 on the central meridian, assigns a false easting of 500,000 m and, in the southern hemisphere, a false northing of 10,000,000 m. One degree of latitude spans about 111 km everywhere; one degree of longitude spans about 111 km × cos(latitude), shrinking to zero at the poles, so areas computed in degrees are meaningless. Many layers fail to overlay simply because they carry different datums or projections.
Spatial analysis derives new information from location. Vector operations include buffering (zones at a set distance around features, such as riparian strips), overlay (intersect keeps areas common to both layers, union keeps everything, clip cuts one layer to another's boundary, erase removes it), dissolve (merging polygons with the same attribute) and spatial joins that transfer attributes by location. Raster operations follow map algebra: local operations combine cell values at the same location (adding layers, reclassifying, multi-criteria weighting), focal operations use a moving neighbourhood (slope, smoothing), zonal operations summarise one raster within zones of another (mean NDVI per district), and global operations use the whole grid (distance or cost surfaces). Site-suitability analysis combines reclassified criteria with weights, often derived by the analytic hierarchy process. Interpolation estimates values between sample points: inverse distance weighting gives nearer points more weight, splines fit smooth surfaces, and kriging uses the semivariogram to provide estimates and their variance. Network analysis finds shortest paths and service areas; hydrological analysis is covered in the next chapter. Scale is expressed as a representative fraction: at 1:50,000, one centimetre on the map is 50,000 cm, or 500 m, on the ground. A large-scale map (1:5,000) shows a small area in detail; a small-scale map (1:1,000,000) shows a large area with little detail. The modifiable areal unit problem warns that results change with the size and shape of aggregation units, and ecological fallacy warns against inferring individual behaviour from areal averages. Data quality has positional, attribute, temporal and logical-consistency components and must be documented in metadata.
A digital elevation model (DEM) is a raster of heights. A digital terrain model (DTM) represents the bare ground; a digital surface model (DSM) includes vegetation and buildings. Elevation comes from photogrammetric stereo pairs, radar interferometry, LiDAR and ground survey. Global DEMs include the Shuttle Radar Topography Mission (SRTM) data collected in 2000 by C-band interferometry between about 60°N and 56°S at about 30 m posting; a stereo-optical global DEM from the ASTER instrument; and the Copernicus DEM derived from X-band interferometry, distributed at 30 m globally. Radar and photogrammetric global DEMs are surface models in forests, so canopy height biases them upward. Vertical accuracy is reported as RMSE against reference points and varies with slope and cover. Terrain derivatives are computed with focal operations. Slope is the maximum rate of change of elevation: in degrees, slope = arctan(√((dz/dx)² + (dz/dy)²)); in percent, it is rise / run × 100, so a 45° slope is 100 percent. Aspect is the compass direction a slope faces and controls solar radiation, snowmelt and moisture; in the northern hemisphere south-facing slopes are warmer and drier, and in the southern hemisphere north-facing slopes are. Curvature separates convex ridges from concave hollows. Hillshade simulates illumination for display. Combined indices, such as the topographic wetness index ln(a / tan β), where a is the upslope contributing area per unit contour width and β the slope angle, predict where soils will be saturated. DEM resolution strongly affects derivatives: slopes computed from coarse DEMs are systematically gentler.
Watershed delineation from a DEM follows a standard sequence. First, sinks (cells surrounded by higher neighbours, mostly artefacts) are filled or breached so water can flow off the surface. Second, flow direction is computed for each cell. The D8 method sends all flow to the one of eight neighbours with the steepest downhill drop; multiple-flow-direction and D-infinity methods split flow between neighbours and represent divergent hillslopes better. Third, flow accumulation counts, for each cell, the number of upstream cells draining through it; multiplying by cell area gives contributing area. Fourth, a stream network is defined by choosing an accumulation threshold: cells above the threshold are channels, and lower thresholds produce denser networks, so the threshold is checked against mapped streams. Streams are then ordered, commonly by the Strahler method, in which two streams of order n join to form order n + 1 and a lower-order tributary does not raise the order. Fifth, outlets (pour points) are placed on the stream network, snapped to the cell of highest accumulation nearby, and the watershed is every cell that drains to that outlet. Flat areas, roads, culverts, reservoirs and urban drainage confuse DEM-based delineation, and burning mapped streams into the DEM is one common fix. Delineated catchments feed morphometric analysis (drainage density = total stream length / basin area, bifurcation ratio, relief ratio, elongation), hydrological and erosion models, priority setting for watershed programmes and the aggregation of other data, such as land cover or rainfall, to the catchment as the planning unit.
Satellite monitoring has turned forest change from a periodic estimate into an operational service. Global tree-cover loss maps derived from the Landsat archive at 30 m resolution report annual canopy removal since 2000 for every country, and near-real-time alert systems based on Landsat, Sentinel-2 and Sentinel-1 radar flag new disturbances within days to weeks, with radar alerts continuing through cloudy seasons. Several tropical countries operate national systems: Brazil's annual clear-cut mapping and rapid alert programmes for the Amazon are the longest-running examples, and national forest monitoring systems built with FAO support in countries across Africa, Asia and Latin America combine satellite land monitoring with national forest inventories to report under the UN climate convention. Degradation, the loss of carbon or function within forest that remains forest, is harder: selective logging, fire and fuelwood removal leave a canopy that still looks green at 30 m. Methods that help include spectral mixture analysis to estimate fractions of green vegetation, non-photosynthetic vegetation and soil, canopy-gap detection with fine-resolution imagery, radar texture, LiDAR and dense time series that catch short-lived signals. Interpretation needs the land-use context: plantation harvest followed by replanting is not deforestation, while forest converted to pasture or cropland is. Accuracy assessment with stratified samples and area-adjusted estimates is standard practice for reporting, including results-based payments for reduced emissions, where independent verification checks the activity data. Commodity regulations that require deforestation-free supply chains rely on geolocated plots overlaid on forest maps with a cut-off date, which has made polygon data quality and map accuracy a commercial issue.
Drought monitoring combines meteorological, agricultural and hydrological indicators. Satellite rainfall estimates merged with gauges provide precipitation anomalies and the Standardized Precipitation Index (SPI). Vegetation-based indices compare the current state with the record for the same place and time of year. The Vegetation Condition Index, VCI = (NDVI − NDVImin) / (NDVImax − NDVImin) × 100, scales the current NDVI between the historical minimum and maximum for that pixel and period; values below about 35 to 40 are commonly read as drought stress, though thresholds are calibrated locally. The Temperature Condition Index, TCI = (BTmax − BT) / (BTmax − BTmin) × 100, uses thermal brightness temperature so that hotter-than-usual conditions give low values. The Vegetation Health Index combines them, commonly VHI = 0.5 × VCI + 0.5 × TCI. Anomalies of NDVI from the long-term mean, land surface temperature, evapotranspiration estimated by energy-balance models, and soil moisture from passive L-band microwave missions and active radar extend the picture. Gravity-measuring satellites detect changes in total water storage, including groundwater, over large basins. Crop condition monitoring compares the current season's index profile with the average, previous years and the best year at the same date, by crop mask where available; early-season anomalies trigger field checks and feed food-security early-warning bulletins and international crop-monitor reports for major producing and food-insecure regions. Crop-type mapping uses multi-date optical and radar data, with radar valuable for rice paddies where flooding produces a distinctive backscatter pattern. Yield forecasting links seasonal indices or crop models to statistics but requires calibration for each region and crop. Interpretation must separate drought from other causes of low greenness: late planting, a switch in crops, flooding, pests or conflict.
Above-ground biomass (AGB) is the dry mass of living vegetation above the soil, usually in tonnes per hectare. No sensor measures biomass directly; it is estimated by linking remote sensing variables to field plots in which biomass is computed with allometric equations from tree diameter at breast height, height and wood density. Pan-tropical allometric models, such as those of Chave and colleagues, express AGB as a power function of the product of wood density, squared diameter and height. Errors arise at every step: choice of allometry, plot size and location error, and the remote sensing model, so uncertainty must be propagated. Three families of sensors are used. Optical indices relate to biomass in young or open vegetation but saturate once the canopy closes. Radar backscatter increases with biomass up to a saturation level that occurs at lower biomass for shorter wavelengths; L-band saturates later than C-band, and P-band, the longest, is the most sensitive to forest biomass, which is why a P-band mission was launched for global forest biomass. LiDAR measures canopy height and vertical structure directly and gives the strongest relations with biomass across all biomass levels. Spaceborne LiDAR samples are combined with wall-to-wall optical and radar data to produce continuous maps. Plot data used for calibration must match the spatial unit of the sensor: plots much smaller than pixels or footprints, or poorly geolocated, weaken the relation. Below-ground biomass is not observed; it is estimated with root-to-shoot ratios that vary with ecological zone and above-ground biomass, as tabulated in the IPCC guidelines. Dead wood, litter and soil carbon need other methods.
Biomass is converted to carbon with a carbon fraction; the IPCC default for wood is 0.47, and values from about 0.43 to 0.55 have been measured, so species-specific values are preferred where available. Carbon is converted to carbon dioxide by the ratio of molecular masses, 44/12, about 3.67. For example, a forest with 200 t ha⁻¹ of AGB holds about 200 × 0.47 = 94 t C ha⁻¹ above ground, equal to about 345 t CO₂ ha⁻¹. Greenhouse-gas inventories under the IPCC guidelines estimate emissions as activity data multiplied by emission factors: remote sensing supplies activity data (hectares of each land-use change), and field inventories supply emission factors (carbon stock differences per hectare). Two approaches are recognised: the gain–loss method, which adds growth and subtracts removals and disturbance losses, and the stock-difference method, which subtracts carbon stocks measured at two times and divides by the interval. Biomass maps can also provide spatially explicit emission factors, and are useful for checking national estimates, but the IPCC good-practice principles of transparency, accuracy, completeness, consistency and comparability apply: map bias must be assessed against independent field data, and estimates reported with uncertainty, often as a 90 or 95 percent confidence interval. Results-based programmes for reducing emissions from deforestation and forest degradation (REDD+) rely on these methods for forest reference levels and monitoring, reporting and verification. Carbon projects similarly use remote sensing to establish baselines and detect leakage and reversals. Large disagreements remain between published global biomass maps in the tropics, mainly because of calibration data and saturation, so users should check maps against local plots before relying on them.
For most of the history of remote sensing, scenes were sold one by one. The opening of the full Landsat archive at no cost in 2008 changed the field: downloads multiplied, and analyses of every available image across whole continents became possible. The European Copernicus programme adopted a full, free and open data policy for its Sentinel missions, which include Sentinel-1 radar, Sentinel-2 multispectral imagery, Sentinel-3 ocean and land colour and surface temperature, and Sentinel-5P atmospheric composition, the last of which maps gases such as nitrogen dioxide and methane. Other freely available records include MODIS and its successor instruments for daily global coverage, rainfall estimates merged with gauge data, global DEMs, night-time lights, spaceborne LiDAR samples and many derived products: land cover, surface water occurrence, tree cover, burned area, fire detections and evapotranspiration. Space agencies across several continents operate land-observation satellites, and data access policies differ by mission and resolution. Commercial operators provide very high resolution optical imagery, daily small-satellite constellations and commercial SAR, and some programmes make such data available free for tropical forest monitoring or humanitarian response. Analysis-ready data (ARD), processed to a consistent standard such as the CEOS Analysis Ready Data specification, deliver surface reflectance or backscatter with quality masks so that users can analyse time series without repeating corrections. Derived global products carry their own accuracy statements and definitions, which must be read before use; a global land-cover map may have high overall accuracy yet be unreliable for a particular class or region.
Petabyte archives cannot be downloaded to a laptop, so analysis has moved to where the data sit. Several commercial and public cloud platforms host the major archives alongside parallel computing, allowing a user to filter a collection by date, cloud cover and location, compute a cloud-free median composite of a whole country, train a classifier and export results through a script. Open-source infrastructure provides the same capability for national programmes: the Open Data Cube organises ARD into time-indexed arrays and underpins continental services such as those for Africa and Australia. Standards make data interoperable: Cloud-Optimized GeoTIFF files allow reading only the needed part of an image over the internet, and SpatioTemporal Asset Catalog (STAC) metadata let software search collections from different providers in a uniform way. Open-source desktop GIS and programming libraries in languages such as Python and R handle raster and vector processing, and Jupyter-style notebooks combine code, results and narrative. Advantages are scale, speed and reproducibility; risks include dependence on one provider's terms of service and continuity, opaque internal algorithms, costs that grow with use, and limited connectivity in some regions. Good practice is to version code, record data collection identifiers and processing levels, store training and validation data, report accuracy, and keep analyses portable between platforms. Image compositing choices, such as median, greenest pixel or best-available-pixel rules, change results and must be documented. File size is easy to estimate: rows × columns × bands × bytes per value; a 10,000 × 10,000 pixel image with 10 bands at 16 bits (2 bytes) is 2 × 10⁹ bytes, about 2 GB before compression.
Global navigation satellite systems (GNSS) include the United States' GPS, Russia's GLONASS, the European Union's Galileo and China's BeiDou, complemented by regional systems in Japan and India. Each satellite broadcasts its position and a precise time signal. A receiver measures the travel time of signals, converts it to a pseudorange and solves for its position by trilateration. Because the receiver's clock is not synchronised with the satellite atomic clocks, there are four unknowns, three coordinates and the clock error, so at least four satellites are needed for a three-dimensional fix. Errors come from satellite clock and orbit errors, delay in the ionosphere and troposphere, multipath (signals reflected from buildings, rocks or tree trunks before reaching the antenna), receiver noise and satellite geometry. Geometry is summarised by dilution of precision (DOP); positional DOP (PDOP) below about 4 to 6 is usually acceptable, and satellites spread widely across the sky give low DOP. A typical single-frequency handheld receiver in open sky is accurate to a few metres; under dense forest canopy errors can rise to tens of metres because of signal blockage and multipath. Differential GNSS uses a base station at a known point to correct errors shared by a nearby rover, reaching sub-metre accuracy; real-time kinematic (RTK) positioning uses carrier-phase measurements and a nearby base or correction network to reach centimetre accuracy; precise point positioning (PPP) uses precise orbit and clock products without a local base, converging to decimetre or better accuracy after a period of observation. Multi-constellation, dual-frequency receivers improve availability and accuracy, especially under canopy and in valleys. Averaging many fixes at a point reduces random error but not bias such as multipath.
Field data, also called ground truth or reference data, train classifiers, calibrate biomass and other models and validate maps. Their value depends on design. Sampling for validation should be probability-based (simple random, stratified random or systematic), so that accuracy and area estimates are unbiased; convenience samples along roads give biased estimates. Each plot should be large enough and homogeneous enough to match the pixel or object being assessed, with consideration of GNSS error: a plot centred within one 10 m pixel with a 5 m position error may describe the neighbouring pixel. Timing should match image acquisition, especially for crops and other fast-changing covers. Data collection protocols record date, time, coordinates and their accuracy, coordinate reference system, class according to a defined legend, photographs in standard directions, and notes on mixed or ambiguous sites. Mobile data collection apps based on open-source form standards replace paper forms and record coordinates, photographs and validation rules directly. Where fieldwork is not possible, interpretation of very high resolution imagery by trained analysts, preferably two interpreters with a consensus process, provides reference data; crowdsourcing and citizen science extend coverage but need quality checks. Ethical and legal points include permission to access land, privacy of identifiable people and property in photographs and drone imagery, informed participation of local and indigenous communities and data-sharing agreements that respect their rights. Metadata and data management plans ensure that field data can be reused for future maps and that maps can be audited, a requirement in results-based carbon and deforestation-free supply-chain schemes.
A student uses this subject to read any remote sensing paper and to answer the standard exam items: which band, which index, which resolution, which classifier, which projection. The practical skills to build are computing indices from surface reflectance, producing a supervised classification with a separate validation sample and an error matrix, delineating a watershed from a free DEM, and checking a layer's coordinate reference system before overlaying it. Common mistakes are confusing the two NDWIs, reading producer's accuracy as user's accuracy, measuring areas in degrees, and comparing NDVI from different processing levels.
A practitioner five years into an environmental agency, consultancy or NGO produces maps that others act on: forest-loss reports for a jurisdiction, drought bulletins, flood extents from radar, land-cover baselines for impact assessments and carbon projects. Daily decisions include choosing Sentinel-1 when the monsoon or wet season blocks optical images, setting a stratified validation sample that gives enough points in rare change classes, reporting area-adjusted estimates with confidence intervals, and documenting code and data on a cloud platform so that a colleague or auditor can reproduce the result.
A senior professional with twenty-five years of experience designs national or corporate monitoring systems: which sensors and products to adopt, how to combine them with forest inventories and field networks, how to meet reporting requirements for climate conventions, results-based payments and deforestation-free supply-chain rules, and how to staff and fund a team. Judgements include when a global product is good enough and when a national map is needed, how to handle disagreements between maps, how to avoid lock-in to one cloud provider, and how to communicate uncertainty to ministers, buyers and communities.
A researcher with a PhD and long experience advances methods and sets standards: biomass retrieval from P-band and L-band radar and spaceborne LiDAR, time-series algorithms for degradation, deep learning with limited labels, data fusion across sensors, uncertainty propagation from plots to national totals, and validation protocols adopted by international bodies. Such researchers sit on mission science teams and guideline committees, supervise doctoral students, review the accuracy claims of global products and examine postgraduate and fellowship candidates, so they must also know where the field's methods remain unsettled.
This manual does not teach the mathematics of radiative transfer, SAR image formation, interferometric processing or deep-learning architectures, and it gives sensor specifications only for the most widely used missions; mission details (launch status, revisit, band sets) change and should be checked with the operating agency. Hyperspectral, ocean-colour and atmospheric-composition remote sensing are only named. Threshold values quoted for indices are indicative ranges that must be calibrated locally. Precision agriculture and forest inventory with LiDAR are covered in other subjects.
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.