Forestry & Forest Sciences · Mensuration, Inventory & Geospatial Analysis · also Aerial Photo Interpretation and Remote Sensing, Forest Remote Sensing, Photogrammetry
Reading the forest from above: aerial photographs and satellite images, spectral behaviour of vegetation, image classification, change detection and forest-cover monitoring.
In the exams: IFoS Forestry Paper II Section A (forest mensuration and remote sensing); ICAR AIEEA-PG Group 09 (aerial photo interpretation and remote sensing); ICAR AICE-JRF 59 Unit 4
Remote sensing is the measurement of the forest from a distance, by sensors carried on aircraft, drones and satellites that record the energy reflected or emitted by trees, soil and water. Foresters have used aerial photographs since the 1920s to map forest types, plan roads and estimate tree heights. Today satellite images cover every forest on Earth every few days, many of them free of charge, and they are the basis of national forest-cover reports, deforestation alerts, fire monitoring and the activity data that countries report under the climate convention. Drones carry cameras and sensors over individual stands at centimetre detail, and radar sees through cloud to monitor forests in the wet tropics.
This subject covers the physical basis of remote sensing, the electromagnetic spectrum and the spectral signature of vegetation; aerial photographs, their scale, stereoscopic viewing and the measurement of height from parallax and relief displacement; photo-interpretation keys for forest types; satellite sensors and the four kinds of resolution; vegetation indices; image classification and accuracy assessment; forest cover and change detection; radar; drones; fire detection and burned-area mapping; and national forest-cover monitoring systems. Spatial data handling, GNSS, LiDAR and map production are in forest-gis-lidar, the companion subject in this track. Plot-based inventory, sampling designs and the combination of field plots with remote sensing in estimation are in forest-inventory-sampling. The use of remote sensing to estimate carbon stock change, REDD+ reference levels and measurement, reporting and verification is treated in forests-climate-carbon. The statistics of map accuracy, area estimation from satellite maps and spatial autocorrelation are developed in the Agricultural Statistics track (spatial-statistics-remote-sensing), and the general remote sensing of land and water in the Agricultural Engineering track (remote-sensing-gis). This subject concentrates on what a forester needs: how forests look to sensors and how to turn images into reliable forest information.
Remote sensing is examined worldwide: in civil-service forestry examinations (usually alongside forest mensuration), in postgraduate entrance and doctoral fellowship examinations in forestry, in forest-service and range-officer recruitment examinations, and in professional remote sensing and photogrammetry certification. Questions ask for spectral regions and the behaviour of leaves, photo scale and height calculations, elements of interpretation, sensor resolutions, NDVI and other indices, classification methods, error matrices and their accuracies, the canopy-density classes used in national forest-cover reports, radar wavelengths, drone ground sample distance and burn severity indices. Calculations of scale, height, NDVI, accuracy and ground sample distance are common and are worked through below.
Remote sensing records electromagnetic energy. Passive sensors record sunlight reflected from the surface or heat emitted by it; active sensors, such as radar and LiDAR, send out their own pulses and record what returns. Energy is described by its wavelength. The regions used in forestry are the visible (about 0.4–0.7 µm: blue 0.4–0.5, green 0.5–0.6, red 0.6–0.7), the near infrared (NIR, about 0.7–1.3 µm), the shortwave infrared (SWIR, about 1.3–3.0 µm), the thermal infrared (about 3–14 µm, with the most-used window at 8–14 µm) and the microwave (about 1 mm to 1 m), used by radar. Sunlight peaks in the visible, near 0.5 µm; the Earth, at about 300 K, emits most strongly near 10 µm. Wien's displacement law gives the wavelength of peak emission as about 2898 ÷ T µm, with T in kelvin, which is why fire, at 800–1000 K, is detected best near 3–4 µm.
The atmosphere scatters and absorbs energy. Scattering by molecules (Rayleigh scattering) is strongest at short wavelengths, which is why the sky is blue and why blue bands look hazy. Water vapour, carbon dioxide and ozone absorb in particular bands, leaving atmospheric windows in which sensors are designed to work. Clouds block visible and infrared energy but not microwaves, which is why radar matters in the cloudy tropics. Before images from different dates are compared, they are usually corrected for atmospheric effects to give surface reflectance, the fraction of incoming energy reflected by the surface.
A green leaf has a distinctive spectral signature. In the visible, chlorophyll absorbs strongly in the blue (near 0.45 µm) and the red (near 0.67 µm) for photosynthesis and reflects somewhat more in the green (near 0.55 µm), which is why leaves look green; visible reflectance is low, typically under 10–15%. Between about 0.68 and 0.75 µm reflectance rises steeply; this is the red edge, and its position shifts to shorter wavelengths when a plant is stressed or its chlorophyll content falls. In the NIR, reflectance is high, often 40–50% or more, because light is scattered at the boundaries between cell walls and air spaces in the spongy mesophyll; this is a property of leaf structure, not of pigment. In the SWIR, reflectance falls again, with deep water-absorption features near 1.45 and 1.94 µm, so SWIR reflectance increases as leaves dry. Soil and rock show a gently rising curve without the red trough and NIR plateau, and water absorbs almost all NIR, so it appears dark in NIR images.
Canopies differ from single leaves. Leaf area, leaf angle, crown shape and shadows all affect what the sensor sees. Broadleaved canopies usually reflect more NIR than coniferous canopies, whose needles, clumped shoots and deep shadows trap light; this is one basis for separating forest types. Stressed, diseased or defoliated trees show higher red reflectance (less chlorophyll), lower NIR reflectance (damaged mesophyll and less leaf area) and higher SWIR reflectance (less water), changes used to map insect outbreaks and drought damage (see forest-entomology and forest-pathology). Seasonal changes in leaf area and colour, called phenology, help separate deciduous from evergreen forests when images from several dates are used.
Most photographs used for mapping are vertical, taken with the camera axis pointing straight down (within a few degrees); oblique photographs are tilted and give a familiar view but variable scale. Classic mapping cameras used film with a 23 cm × 23 cm frame and a focal length of 152 mm; modern large-format digital cameras record the same kind of geometry. Photographs are taken along flight lines, with each photo overlapping the next by about 60% (forward overlap or endlap) so that every point appears on at least two photos, and adjacent strips overlapping by about 20–30% (side overlap or sidelap).
Scale is the ratio of a distance on the photo to the same distance on the ground. For a vertical photo over flat terrain, scale = f ÷ (H − h), where f is the focal length, H the flying height above sea level (datum) and h the elevation of the ground. A camera with a 152 mm focal length flown 3,040 m above ground gives a scale of 0.152 ÷ 3,040 = 1:20,000. Scale can also be found by measuring the photo distance between two points whose ground distance is known: if a 500 m road segment measures 2.5 cm on the photo, the scale is 0.025 ÷ 500 = 1:20,000. Because terrain varies, scale on a photo varies too: hilltops, being closer to the camera, appear at a larger scale than valleys. Ground distance = photo distance × scale denominator, and ground area = photo area × (scale denominator)²; 1 cm² on a 1:20,000 photo covers 200 m × 200 m = 4 ha. Forestry type mapping traditionally used photos at 1:10,000 to 1:25,000; detailed tree counts and crown measurements need larger scales, such as 1:5,000 or less.
Two overlapping photos viewed so that each eye sees one of them give a three-dimensional impression, the stereomodel, because the two photos were taken from different positions. Pocket (lens) stereoscopes are small and portable; mirror stereoscopes allow whole overlapping photos to be viewed; digital photogrammetric workstations use screens with polarised glasses. Vertical exaggeration makes trees and slopes look steeper than they are, which helps interpretation.
Parallax is the apparent shift of an object between two photos caused by the change in camera position. Higher objects, being closer to the camera, have greater parallax. The difference in parallax between the top and the base of an object (differential parallax, dP) is measured with a parallax bar or stereometer, and gives the object's height: h = H × dP ÷ (P + dP), where H is the flying height above the base of the object and P is the absolute parallax of the base, taken in practice as the average photo base length (the distance between the principal point of one photo and the transferred principal point of its neighbour). For H = 1,500 m, P = 90 mm and dP = 1.5 mm, h = 1,500 × 1.5 ÷ 91.5 = 24.6 m.
On a vertical photo, objects that stand above the ground lean outwards from the principal point (the centre of the photo); this is relief displacement. It gives a second height formula from a single photo: h = d × H ÷ r, where d is the displacement (photo distance from base to top of the object), r is the radial distance from the principal point to the top of the object, and H is the flying height above the base. A tree whose top is 80 mm from the principal point and displaced 2 mm, on a photo flown 1,200 m above ground, is 2 × 1,200 ÷ 80 = 30 m tall. The method needs the base to be visible, so it works for isolated trees near the edges of photos and poorly in closed canopy, where parallax is preferred. Relief displacement is also why photos must be orthorectified before they are used as maps; an orthophoto has been corrected for relief and tilt using a digital elevation model so that it has a uniform scale.
Heights can also be estimated from shadows on flat ground: h = shadow length × tan(sun elevation angle). With a 20 m shadow and a sun elevation of 45°, the tree is 20 m tall. The method needs flat terrain, a clear shadow tip and the sun angle at the time of exposure.
Photo interpretation is the identification of objects and the judgement of their significance. Interpreters use a standard set of elements. Tone or colour is the relative brightness or hue; on colour-infrared photographs, healthy broadleaved vegetation appears bright red because of its high NIR reflectance, conifers a darker red-brown, stressed or dead trees greenish or grey, and water black. Size, absolute and relative, separates trees from shrubs and mature from young stands. Shape, of crowns seen from above and of shadows seen from the side, separates species: conical spruce and fir crowns, rounded oak crowns and star-shaped palm crowns. Texture, the visual roughness made by crowns and their shadows, ranges from fine in young plantations to coarse in old, uneven natural forest. Pattern, the spatial arrangement of objects, shows plantations in rows, orchards in grids, shifting cultivation in a patchwork and natural forest without regular arrangement. Shadow reveals profile and height. Site (position in the landscape: ridges, valleys, wetlands, aspect) and association (things that occur together, such as mangroves with tidal creeks or sal with well-drained alluvial plains) let the interpreter use ecological knowledge. Height, measured stereoscopically, completes the list.
An interpretation key is a set of reference material that helps interpreters identify features quickly and consistently: written descriptions, example photographs or image chips, and drawings. Selective keys present examples of each class, and the interpreter chooses the one that matches. Elimination keys, also called dichotomous keys, lead the interpreter through a series of choices from general to specific, such as broadleaved or coniferous, then crown shape, then texture, until only one class remains, in the same way as a botanical key (see forest-botany-taxonomy). Keys are prepared for a particular region, season, image type and scale, because the appearance of a forest type changes with all of these. They are developed and checked with field visits, and a good key states the probability of confusion between similar classes.
In forest mapping, the interpreter outlines stands, areas of uniform type, density and height, on the stereomodel and labels each with a code, for example forest type, crown density class and height class. Crown density (crown closure) is estimated with a crown density scale or dot grid, the proportion of the ground covered by tree crowns. Classes are often in steps of 10–30%. Height classes come from parallax measurement or from a few measured trees. Typical errors arise from shadows on north-facing slopes (in the northern hemisphere), similar-looking species, seasonal differences between photos, and scale differences between hilltops and valleys. Photo-interpreted stand maps form the base of the stratification used in forest inventory, where field plots are then distributed by stratum (see forest-inventory-sampling), and the stand outlines become the compartment polygons of the GIS (see forest-gis-lidar). Although satellite and LiDAR data now dominate, photo interpretation remains the reference method for checking automated maps and for the visual assessment of sample points in national monitoring.
Most Earth-observation satellites travel in near-polar, sun-synchronous orbits at about 500–800 km altitude, crossing the equator at the same local time on every pass so that illumination is comparable from date to date. Geostationary satellites, at about 36,000 km above the equator, see the same hemisphere continuously and are used for weather and for frequent fire observation, but at coarse resolution. Optical sensors are of two main designs: whiskbroom scanners, which sweep a mirror across the track, and pushbroom sensors, which use a line of detectors and image a whole line at once; pushbroom designs give better signal and are used in most modern sensors. The swath is the width of ground imaged on one pass: about 185 km for Landsat and 290 km for Sentinel-2.
Spatial resolution is the size of the smallest ground area resolved, usually given as pixel size: 250–1,000 m for MODIS, 30 m for Landsat multispectral bands, 10–20 m for Sentinel-2, 5.8 m for the LISS-IV sensor on the Indian Resourcesat satellites and 0.3–1 m for commercial very-high-resolution satellites. Spectral resolution is the number and width of bands: a panchromatic band records one broad band; multispectral sensors record 4–15 bands; hyperspectral sensors record hundreds of narrow, contiguous bands that can separate species and chemistry. Temporal resolution is the revisit interval: daily for MODIS, 16 days for each Landsat (8 days with two satellites), and about 5 days at the equator with two Sentinel-2 satellites. Radiometric resolution is the number of brightness levels a sensor can record: an 8-bit sensor records 2⁸ = 256 levels, an 11-bit sensor 2,048 and a 12-bit sensor 4,096. There are trade-offs: finer spatial resolution usually means a narrower swath, less frequent coverage of a given place and larger data volumes, and narrow bands collect less energy each. The choice of sensor depends on the size of the features to be mapped, the area, the season and frequency needed and the budget.
The Landsat series, beginning in 1972, provides the longest continuous record of the Earth's forests, and since 2008 the whole archive has been free. Landsat 8 (2013) and Landsat 9 (2021) carry the Operational Land Imager, with 30 m bands from coastal blue to SWIR and a 15 m panchromatic band, and a thermal sensor at 100 m. The European Copernicus programme's Sentinel-2 satellites carry a 13-band multispectral instrument with bands at 10, 20 and 60 m, including three red-edge bands useful for vegetation; its data are free. MODIS on NASA's Terra and Aqua satellites, and its successor VIIRS, give daily global coverage at coarse resolution, used for phenology, fire and continental mapping. National missions add to these. The Indian Resourcesat satellites carry three sensors, AWiFS (about 56 m, wide swath), LISS-III (23.5 m) and LISS-IV (5.8 m), and LISS-III data are used in that country's biennial forest-cover assessment; the China–Brazil Earth Resources Satellites (CBERS) carry medium-resolution cameras used in South American and Chinese land monitoring; and several countries fly high-resolution panchromatic and stereo imagers. Commercial constellations provide sub-metre images, and small-satellite constellations image the whole land surface daily at 3–5 m. Radar satellites are covered in a later chapter. Hyperspectral missions, such as the Italian PRISMA and the German EnMAP, provide research data for species and canopy chemistry.
A vegetation index combines reflectance in two or more bands into a single number related to the amount or vigour of green vegetation. Indices exploit the contrast between low red and high NIR reflectance of green leaves. The simple ratio is NIR ÷ red. The normalised difference vegetation index is NDVI = (NIR − red) ÷ (NIR + red), which ranges from −1 to +1. Water gives negative values, bare soil and rock values near 0.1–0.2, sparse vegetation 0.2–0.5, and dense green forest 0.6–0.9. With NIR reflectance of 0.45 and red reflectance of 0.05, NDVI = 0.40 ÷ 0.50 = 0.80. Because it is a ratio, NDVI partly cancels differences in illumination caused by slope and sun angle. It is widely used to map vegetation, track phenology and detect loss of green cover, and long NDVI time series from AVHRR and MODIS have been used to study greening and browning of the world's vegetation.
NDVI has limitations. It saturates in dense canopies: once leaf area index exceeds about 3–4, NDVI hardly increases, so it cannot separate a moderately dense from a very dense forest. It is affected by soil background in sparse vegetation, by atmospheric haze and by the sensor's band positions, so values from different sensors are not directly comparable without adjustment. NDVI is not biomass: two forests with very different wood volume can have the same NDVI.
Several indices address these limits. The soil-adjusted vegetation index, SAVI = (NIR − red) × (1 + L) ÷ (NIR + red + L), adds a soil adjustment factor L, usually 0.5, to reduce soil effects in sparse vegetation. 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 is less prone to saturation in dense forest; it is a standard MODIS product. Red-edge indices, using the red-edge bands of Sentinel-2, are sensitive to chlorophyll content and early stress. Moisture indices use NIR and SWIR: the normalised difference moisture index, NDMI = (NIR − SWIR1) ÷ (NIR + SWIR1), falls as canopies lose water, which makes it useful for drought and insect damage. The normalised burn ratio, NBR = (NIR − SWIR2) ÷ (NIR + SWIR2), uses the longer SWIR band near 2.2 µm and is described with fire mapping below. In each case the formula, the bands and whether the input is surface reflectance or raw digital numbers should be stated, because the same name is sometimes used for slightly different formulas.
Indices are related to leaf area index, the fraction of absorbed photosynthetically active radiation (FAPAR) and gross primary productivity, and these relationships underpin global productivity products. They are also used as predictor variables, alongside individual bands, texture measures and LiDAR or radar metrics, in models that predict stand attributes such as basal area or biomass from field plots (see forest-inventory-sampling). In such models the relationship is fitted on plots measured close in time to the image, and its prediction error must be carried into the estimate, a subject treated in agri-statistics' spatial-statistics-remote-sensing.
Before classification, images are geometrically corrected so that each pixel sits in the right place on a map projection, radiometrically calibrated, and corrected for the atmosphere to surface reflectance. Clouds and their shadows are masked. In mountains, topographic correction reduces the difference in brightness between sunlit and shaded slopes, which otherwise causes the same forest to look like two classes. Images from different dates may be composited, for example by taking the median of all cloud-free observations in a season, a common approach on cloud platforms (see forest-gis-lidar).
Classification assigns each pixel or object to a class. In unsupervised classification the computer groups pixels into spectral clusters with algorithms such as k-means or ISODATA (which can split and merge clusters), and the analyst then labels each cluster with a land-cover class; it needs little prior knowledge but the clusters may not match the classes needed. In supervised classification the analyst first selects training areas of known class, and the algorithm learns their spectral characteristics and assigns every other pixel. Classic algorithms are the parallelepiped (box) classifier, minimum distance to means and maximum likelihood, which assumes the classes are normally distributed and uses their means and covariances. Machine-learning classifiers, random forests, support vector machines and, increasingly, deep neural networks, are now standard; random forests are robust, handle many input variables and give a measure of variable importance. Object-based image analysis (OBIA) first segments the image into groups of similar adjacent pixels, which often correspond to tree crowns or stands in high-resolution images, and then classifies the objects using spectral, shape and texture properties. Good training data are the main determinant of a good map: they must represent the full variation within each class, be spread across the area and be independent of the data used for validation.
A map is not complete until its accuracy has been assessed with reference data of higher quality than the map, collected at sample locations chosen by a probability design such as stratified random sampling. The results are summarised in an error (confusion) matrix, with the map classes in rows and the reference classes in columns, or the reverse; the convention must be stated. Overall accuracy is the number of correctly classified samples (the diagonal) divided by the total. Producer's accuracy for a class is the correct count divided by the reference total for that class; its complement is the omission error, the share of true forest that the map missed. User's accuracy is the correct count divided by the map total for the class; its complement is the commission error, the share of mapped forest that is not forest on the ground. For example, if 100 reference forest points include 90 mapped as forest, producer's accuracy is 90%; if the map shows 95 forest points of which those 90 are truly forest, user's accuracy is 90 ÷ 95 = 94.7%.
The kappa coefficient, once standard, compares accuracy with that expected by chance; current good practice (Olofsson and colleagues, 2014) recommends reporting overall, user's and producer's accuracies with confidence intervals instead, and estimating class areas from the reference sample, adjusted for the map's errors, rather than counting map pixels. Pixel counting is biased whenever omission and commission errors are unequal, which is usual for rare classes such as deforestation. The statistics are developed fully in agri-statistics' spatial-statistics-remote-sensing.
A map of forest needs a definition. FAO's Forest Resources Assessment defines forest as land spanning more than 0.5 ha with trees higher than 5 m and canopy cover of more than 10%, or trees able to reach these thresholds in situ, and not predominantly under agricultural or urban land use. This is a land-use definition: a recently clear-felled area that will regenerate remains forest. Remote sensing sees land cover, the physical presence of tree canopy, so a satellite tree-cover map is not the same thing as a forest-area statistic. Countries reporting under the climate convention set their own thresholds within ranges: minimum area 0.05–1 ha, crown cover 10–30% and height 2–5 m. National cover assessments add their own rules. India's biennial assessment, for example, counts as forest cover all lands of more than 1 ha with tree canopy density above 10%, irrespective of ownership, legal status and land use, so it includes orchards and tree plantations outside recorded forest areas. The difference between forest cover and recorded forest area, the area legally designated as forest, is a frequent source of confusion and a frequent examination question.
Change detection identifies differences between images of the same place taken at different times. In post-classification comparison, each date is classified separately and the maps are overlaid; it gives from-to information (forest to cropland) but the errors of both maps combine. Image differencing or ratioing subtracts or divides band values or indices between dates, and a threshold separates change from no change; it is simple but needs well-matched, atmospherically corrected images from similar seasons. Change vector analysis treats the change in several bands as a vector with a magnitude and a direction, the direction indicating the kind of change. Time-series methods use all available images rather than two dates: algorithms such as LandTrendr fit segments to each pixel's Landsat history to find abrupt disturbances and gradual recovery, BFAST separates trend, season and breaks, and CCDC models each pixel continuously and flags departures from its expected behaviour. Time-series approaches reduce false change caused by phenology and clouds and detect degradation as well as clearance.
Deforestation, the conversion of forest to other land use, is comparatively easy to see, as it removes canopy and exposes soil. Forest degradation, a loss of biomass or quality without conversion, such as selective logging, fuelwood removal, understorey fires or grazing, is much harder: the canopy may close again within a year or two, so detection needs frequent images, high resolution, SWIR and moisture indices, radar or LiDAR. Global products include the University of Maryland Global Forest Change data set (Hansen and colleagues, 2013, updated annually), which maps tree cover in 2000 and annual tree-cover loss at 30 m from Landsat; it measures loss of tree cover from any cause, including harvest in plantations and fire, and is not a deforestation map in the land-use sense. Others include the Joint Research Centre's Tropical Moist Forest data set, which separates deforestation from degradation, and near-real-time alert systems based on Landsat, Sentinel-2 and Sentinel-1 radar, distributed through Global Forest Watch. Users must check what each product measures, its definition of forest and its accuracy before using it to report or enforce.
Radar is an active microwave sensor: it transmits pulses and records the energy scattered back (backscatter) and the time it takes to return. Because microwaves pass through cloud, haze and smoke, and because the sensor provides its own illumination, radar images can be acquired day or night in any weather, which makes radar the most reliable source of regular images over the cloudy humid tropics. Imaging radars on satellites are synthetic aperture radars (SAR): they use the motion of the platform to simulate a very long antenna and so achieve fine resolution, typically 5–30 m. Radar looks sideways, which causes geometric effects in hilly terrain, foreshortening, layover and radar shadow, and the images contain speckle, a grainy noise from the interference of many scatterers within a pixel, reduced by averaging (multi-looking) or filtering.
Radar bands are named by letters. X-band (about 3 cm) and C-band (about 5.6 cm) are scattered mainly by leaves and small twigs in the upper canopy. L-band (about 23 cm) penetrates the canopy and is scattered by branches and stems. P-band (about 70 cm) penetrates further and interacts with large branches and trunks. Longer wavelengths therefore have a closer relationship with woody biomass. Polarisation describes the orientation of the transmitted and received waves: HH means horizontal transmit and horizontal receive, VV vertical and vertical, and HV or VH the cross-polarised combinations. Cross-polarised backscatter comes mainly from volume scattering in the tangle of branches and leaves and is the most sensitive to forest biomass; like-polarised HH backscatter at L- or P-band includes a strong double-bounce component between trunks and the ground, which is enhanced over flooded forest and helps map mangroves and swamp forests.
Backscatter increases with biomass but then saturates: C-band loses sensitivity at low biomass, roughly 20–50 t/ha, and L-band at roughly 100–150 t/ha, depending on forest type and conditions; P-band saturates later. Above saturation, radar cannot separate a medium from a dense forest by backscatter alone. Soil and canopy moisture, rain and freezing also change backscatter, so dates must be chosen with care.
The Copernicus Sentinel-1 satellites carry C-band SAR and provide free images every few days to two weeks over most of the globe, widely used for forest-loss alerts in the tropics. Japan's ALOS-2 PALSAR-2 provides L-band data, and annual global L-band mosaics have been released for forest/non-forest mapping. The NASA–ISRO NISAR mission, launched in 2025, carries L-band and S-band radar, and the European Space Agency's BIOMASS mission, also launched in 2025, is the first P-band SAR in orbit, designed to map above-ground biomass in dense tropical forests. Interferometric SAR (InSAR) uses the phase difference between two radar images taken from slightly different positions to measure surface height; the German TanDEM-X mission produced a global elevation model this way, and the difference between interferometric heights and a terrain model gives forest canopy height. Polarimetric interferometry (Pol-InSAR) and tomography use multiple polarisations and passes to separate the vertical structure of the forest. Radar biomass maps need field plots or LiDAR heights for calibration; see forests-climate-carbon for their use in carbon accounting.
Unmanned aerial vehicles, also called drones or remotely piloted aircraft systems, carry cameras and sensors over areas from a few hectares to a few square kilometres. Multirotor drones take off vertically, hover and fly slowly, suiting small sites, steep terrain and close inspection, but their batteries usually last only 20–40 minutes. Fixed-wing drones fly longer and cover larger areas but need space to launch and land; hybrid vertical-take-off fixed-wing designs combine the two. Sensors include ordinary RGB cameras, multispectral cameras with red-edge and NIR bands, thermal cameras, hyperspectral cameras and LiDAR (covered in forest-gis-lidar). Drones are used to map plantations and count seedlings, measure tree heights and crowns, detect pest and disease damage, assess storm and fire damage, monitor harvesting and restoration, and survey wildlife.
The ground sample distance (GSD) is the ground size of one image pixel: GSD = sensor pixel size × flying height ÷ focal length, in consistent units. A camera with a 2.4 µm pixel and an 8.8 mm lens flown at 100 m above ground gives GSD = 0.0000024 × 100 ÷ 0.0088 = 0.027 m, about 2.7 cm. GSD halves if the flying height is halved. Flights are planned as parallel lines, with high overlap, typically 75–85% forward and 65–80% side, much more than for classic aerial photography, because forest canopies are complex, move in wind and contain few features that match between images. Flights near solar noon reduce shadows, and calm, uniformly lit (all sunny or all overcast) conditions improve results. Ground control points, marked targets surveyed with high-accuracy GNSS, tie the products to map coordinates; drones with real-time or post-processed kinematic GNSS reduce the number needed. Multispectral cameras need a calibration panel or a sunlight sensor to convert images to reflectance.
Drone images are processed with structure from motion (SfM) photogrammetry, which finds matching features in many overlapping images, solves for camera positions and builds a dense three-dimensional point cloud, from which an orthomosaic and a digital surface model are produced. Subtracting a digital terrain model from the surface model gives a canopy height model, from which tree tops can be detected and heights measured. The main limitation in forests is that photogrammetry sees only surfaces visible from above: it reconstructs the top of the canopy well but cannot see the ground under closed canopy, so the terrain model must come from LiDAR or from ground visible in gaps. LiDAR drones overcome this because laser pulses pass through gaps to the ground. Drone flights are regulated nationally, and the rules share a common pattern: registration of the aircraft and pilot, a height ceiling of roughly 120 m above ground for routine flights, flight within visual line of sight, and airspace zones near airports where permission is needed. For example, the United States Part 107 rules set 400 feet (about 122 m) above ground, the European Union's open category sets 120 m, and India's Drone Rules, 2021 publish permitted, controlled and restricted zones on an online airspace map. Check the current national rules before every survey. Pilots must check current national rules, permissions, privacy and wildlife-disturbance obligations before flying.
Active fires are detected as thermal anomalies: pixels much hotter than their surroundings in the middle-infrared band near 4 µm, where emission from flaming fires peaks, compared with the thermal band near 11 µm. Because a fire covering a small fraction of a pixel raises the 4 µm signal strongly, sensors can detect fires far smaller than their pixels. MODIS detects active fires at 1 km resolution about four times a day from Terra and Aqua, and VIIRS on the Suomi NPP, NOAA-20 and NOAA-21 satellites at 375 m, detecting smaller fires. NASA's Fire Information for Resource Management System (FIRMS) distributes these detections within a few hours. Geostationary satellites detect fires every 10–15 minutes at coarser resolution. Several national agencies relay these detections to field staff: Brazil's space research institute publishes daily fire hotspots for its biomes, and the Forest Survey of India sends satellite fire detections inside forest areas to forest officers by text message and email for ground verification. Fire detections are points or pixels with a time and a confidence; they indicate fire activity, not the area burned, and they miss fires under cloud, under dense canopy and between overpasses. Fire radiative power, the energy radiated by a fire, is related to the rate of fuel combustion and emissions.
Burned areas are mapped from the change in reflectance after fire: burning removes green vegetation, lowering NIR reflectance, and exposes dry soil and ash, raising SWIR reflectance. The normalised burn ratio, NBR = (NIR − SWIR2) ÷ (NIR + SWIR2), is high in healthy vegetation and low in burned areas. The difference between pre-fire and post-fire values, dNBR = NBR(pre) − NBR(post), measures burn severity: with a pre-fire NBR of 0.60 and a post-fire NBR of 0.10, dNBR = 0.50. Classes proposed by Key and Benson are, in round figures, unburned below about 0.1, low severity 0.1–0.27, moderate-low 0.27–0.44, moderate-high 0.44–0.66 and high above 0.66; these thresholds were developed in North American conifer forests and should be calibrated with field observations, for example the Composite Burn Index, elsewhere. A relative version, RdNBR, divides by a function of the pre-fire NBR so that severity is comparable between sparse and dense vegetation. Global burned-area products, such as MODIS MCD64A1 at 500 m, are used for emissions estimates; Landsat and Sentinel-2 give finer maps for management.
Remote sensing supports each stage of fire management (see forest-fire-science). Before fires, it maps fuel types, fuel moisture and fire risk; during fires, active-fire detections and infrared aircraft or drone imagery show the fire perimeter and hot spots, with thermal cameras seeing through smoke; after fires, burned-area and severity maps guide salvage, erosion control and restoration, and long-term recovery is tracked with vegetation indices. Burned-area and fire-emissions data also feed national greenhouse-gas inventories, discussed in forests-climate-carbon.
Many countries map their whole forest area, wall to wall, at regular intervals from medium-resolution satellite imagery, interpreting each image into forest and non-forest and often into canopy-density classes. The design choices are always the same: the imagery and its resolution, the mapping scale, the minimum mappable unit, the class thresholds, and how the map is checked on the ground. India's biennial assessment, run since 1987, is a widely cited example: it interprets Resourcesat LISS-III imagery at 23.5 m digitally at a scale of 1:50,000 with a minimum mappable unit of 1 ha, checked by ground truthing. Its forest cover is classified by canopy density into very dense forest (canopy density of 70% and above), moderately dense forest (40% to below 70%) and open forest (10% to below 40%); scrub, degraded land with canopy density below 10%, is not counted as forest cover. Tree cover, meaning patches of trees outside recorded forest smaller than 1 ha, is estimated separately by sampling. Australia maps forest extent and change across the continent from the Landsat archive for its national greenhouse-gas inventory, and the European Union and others use Copernicus high-resolution layers of tree-cover density. Whatever the system, methods, satellite data and interpretation improve over time, so changes between successive reports reflect both real change and methodological change; users should read the methodology before quoting change figures.
Brazil's National Institute for Space Research (INPE) runs PRODES, which has mapped annual clear-cut deforestation in the Brazilian Amazon since 1988 using Landsat-class images, with a minimum mapping area of 6.25 ha, and DETER, a near-real-time system using coarser, more frequent images to send alerts to the environmental enforcement agency IBAMA. The combination of annual measurement for statistics and frequent alerts for enforcement is a model widely copied. Many tropical countries have built national forest monitoring systems with support from FAO, the Global Forest Observations Initiative (GFOI) and others, often using free tools such as Google Earth Engine, FAO's SEPAL platform and Collect Earth for visual interpretation of sample plots on high-resolution imagery. Temperate countries combine satellite mapping with their national forest inventories; for example, the United States Forest Inventory and Analysis programme uses remote sensing to stratify and improve the precision of plot-based estimates (see forest-inventory-sampling).
Countries seeking results-based payments under REDD+ need a national forest monitoring system under the Warsaw Framework of the United Nations Framework Convention on Climate Change. Remote sensing supplies activity data, the area of each kind of change such as deforestation, degradation or afforestation, which is multiplied by emission factors, the carbon change per hectare from field inventory, to give emissions and removals (see forests-climate-carbon). The IPCC and the GFOI Methods and Guidance Documentation recommend that activity data be estimated from a probability sample of reference observations, with map-based stratification and stated uncertainty, rather than by counting map pixels, for the reasons given in the accuracy chapter. Transparency, consistency over time, and the ability of others to reproduce the estimates are as important as technical sophistication. Independent global data sets can be used to check national figures, but definitions differ, so differences must be explained rather than simply resolved in favour of one source.
Learn the spectral curve of a leaf and why NIR is high, photo scale and both height formulas, the elements of interpretation, the four resolutions with example sensors, NDVI and its limits, the error matrix and the difference between producer's and user's accuracy, change-detection methods, radar wavelengths, drone GSD, NBR and dNBR, and the FSI canopy density classes. Practise scale, height, NDVI, accuracy and GSD calculations.
A forester five years in uses satellite and drone images to update stand maps, check plantation survival, investigate encroachment and illegal felling reported by alerts, respond to fire alerts and map burned areas, and supply ground-truth points to national assessments. Practical skills include downloading and viewing free imagery, calculating indices, planning a drone survey within the rules, and reading accuracy statements critically.
A senior forester commissions and uses national and state forest-cover assessments, sets definitions and thresholds, decides which alerts trigger field action, uses remote sensing evidence in enforcement and in reporting to the climate convention, and explains differences between national and global data sets to the public and the press.
A senior researcher develops methods to map species, degradation and biomass from multispectral, hyperspectral, radar and LiDAR data; builds time-series and deep-learning change detection; designs sampling-based area estimation with uncertainty; and works with space agencies on new missions such as BIOMASS and NISAR.
Sensor specifications, mission launches and product versions change frequently; check the space agency's current documentation before relying on band positions, revisit times or availability. Radar saturation levels and burn-severity thresholds are approximate and vary with forest type, so they are given as rough ranges. Methodological details of national forest-cover reports (sensor, scale, classes) are as described in recent reports and change over time; check the current report of the country concerned. Drone regulations are summarised only in outline and must be checked against current national rules. The statistics of accuracy assessment and area estimation are developed in agri-statistics' spatial-statistics-remote-sensing; GIS, GNSS, LiDAR and cartography are in forest-gis-lidar; inventory designs using remote sensing are in forest-inventory-sampling; carbon applications in forests-climate-carbon.
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.