Environmental Data & GIS: Reading the Planet from Orbit

On most web maps, Greenland looks about as big as Africa. It is not. Africa is about fourteen times larger: roughly 30.4 million square kilometres against 2.17 million. Nobody lied; the map made a choice that is excellent for navigation and poor for comparing sizes, and the choice is invisible unless you know to ask.

Environmental data is full of choices like that. A satellite picture of a field becomes a green-brown map through a formula. A "twenty-year trend" depends on which years you pick. A forest-loss figure comes from a map that is itself an estimate. A graduate of this course does not distrust every map; they ask four questions of any one: what was measured, at what pixel size, in which projection, and with what error?

Remote sensing: reading light from orbit

Every surface reflects some fraction of the sunlight that hits it, and the fraction changes with the colour (wavelength) of the light. That pattern is its spectral signature. Chlorophyll in a leaf absorbs red and blue light for photosynthesis, so a healthy leaf reflects little red. The spongy cells inside the leaf scatter near-infrared light (just past what our eyes see), so a healthy leaf reflects a lot of it. Water absorbs nearly everything beyond red. Snow is bright in visible light and dark in the short-wave infrared. A sensor that measures a handful of bands can therefore tell crops, water, soil and snow apart even where they look alike to the eye.

The long data record is public. Landsat 1 launched on 23 July 1972, and Landsat 9 on 27 September 2021, carrying on a record with 30 m multispectral pixels. The European Sentinel-2 pair launched on 23 June 2015 and 7 March 2017 (a third, 2C, on 5 September 2024), with 13 bands and pixels down to 10 m, covering a 290 km swath.

Without JavaScript: a healthy crop reflects about 0.04 in red and about 0.50 in near-infrared; water reflects almost nothing in near-infrared; bare soil rises slowly from red to infrared.

NDVI: one number for how green

The simplest and most used trick is the normalised difference vegetation index. Take the near-infrared (NIR) and red reflectances, subtract, and divide by the sum. Healthy plants are dark in red and bright in NIR, so the answer is high; bare soil is nearly equal in both, so it sits near zero; water is darker in NIR than red, so it is negative. Rouse and colleagues at Texas A&M reported it from Landsat data in 1973–74, and Compton Tucker of NASA published early uses in 1979.

Without JavaScript: healthy crops near 0.76 fall to about 0.39 under severe stress; forest is about 0.86, town about 0.1 and water below zero.

NDVI is a proxy. It saturates in dense canopy (a rainforest and a lush field can both read near 0.9), wet soil darkens, haze and clouds shift it, and different sensors define "red" and "NIR" slightly differently. It tracks greenness, not yield or health. Its great strength is that it is computed the same way across fifty years of images.

Resolution: four ways a pixel can be too coarse

A pixel records one average for the whole square, so anything smaller than the square is blended into its surroundings. Remote-sensing courses distinguish four kinds of resolution. Spatial: the size of the square (10 m for Sentinel-2's visible bands, 30 m for Landsat, 250 m to 1 km for the MODIS instruments that watch the whole planet daily). Spectral: how many bands, and how narrow. Temporal: how often the same spot is revisited (about 5 days for Sentinel-2 with two satellites; 16 days for one Landsat). Radiometric: how finely brightness is divided, the bit depth. You rarely get all four at once; fine pixels cost a narrow swath and fewer visits.

Without JavaScript: at 10 m all five copses and the hedgerow are visible; at 20 m the hedgerow is gone and the forest area is about 14% low; at 40 m only two of the five copses remain.

This is the mixed-pixel problem. A 30 m Landsat pixel straddling a road and a field reads as neither. It also explains why a deforestation map from 30 m data cannot see selective logging of a few trees, and why the same landscape gives different area figures at different resolutions.

Map projections: flattening a sphere

A globe cannot be flattened without tearing or stretching; Gauss proved this in 1827. Every flat map makes a trade. The Mercator projection (Gerardus Mercator, 1569) keeps angles right, so a straight line on the map is a steady compass course, which is why sailors used it. The price is that scale grows with latitude: at 60° everything is drawn twice as wide and twice as tall, four times the area. Web maps use a variant, Web Mercator (EPSG:3857, introduced by Google Maps in 2005), which cuts off at about 85°. An equal-area projection keeps sizes right and bends shapes; the Gall–Peters map is one, and the Equal Earth projection (Šavrič, Patterson and Jenny, 2018) is a newer, rounder one.

The practical rule for data work: the coordinate system is part of the data. Latitude and longitude (WGS 84, EPSG:4326) are angles, not distances: a degree of longitude is 111.32 km at the equator and half that at 60°. To measure area or distance, use a projection suited to the region, such as a UTM zone (the world is cut into 60 zones, each 6° wide) or an equal-area projection for areas. Mixing two coordinate systems quietly shifts features by metres to kilometres.

Without JavaScript: at 60° Mercator draws everything 2 times too wide and 2 times too tall, so 4 times the area; the equal-area projection keeps area but draws a circle as a flat ellipse. Greenland is truly 0.07 of Africa and is drawn at about 0.75 of it in Mercator.

Layers and overlay: what is near what

A geographic information system stores the world as layers. In the vector model, a feature is a point (a well), a line (a river) or a polygon (a district), each with a table of attributes. In the raster model, a feature is a grid of cells with a value in each, like the NDVI scene above. The questions a GIS answers are almost always spatial ones: what is inside this boundary, what is within 2 km of that river, which polygons overlap, and what is the population of the overlap? The method is old. In 1854 John Snow plotted cholera deaths in Soho against water pumps; Roger Tomlinson led the Canada Geographic Information System in the 1960s, often called the first true GIS.

Without JavaScript: within 2 km of the river there are 2 of 16 villages (about a third of the people); within 4 km, 10 villages and about 81% of the people.

A caution that shows up here: the answer jumps. Moving the buffer from 3 to 4 km adds seven villages. Results from overlay depend on the distance, the boundaries and the units chosen, a family of effects geographers call the modifiable areal unit problem: redraw the district lines and the statistics change, though the people have not moved.

Time series and trends: signal, season and noise

One image is a snapshot; a series of images is evidence of change. The record of NDVI from the AVHRR instruments begins in 1981. A study by Zhu and colleagues (2016) using satellite leaf-area data for 1982–2009 found greening over 25% to 50% of the world's vegetated land and browning on under 4%, and a model-based attribution of about 70% of it to rising CO₂; others stress land use and water. Regional stories are just as real: much of the Sahel has become greener since the mid-1980s drought, with scientists debating how much is rainfall recovery and how much is land management.

Three things hide inside any such series: a trend, a seasonal cycle (NDVI rises and falls every year, so compare like months, or use annual means) and noise (weather, clouds, sensor drift). The usual tool is a straight line fitted by least squares, with an interval around the slope from Probability & Statistics (34/05). Two traps are specific to time series. The window matters: a short window ending in a drought can slope the other way from the long-term line. And neighbouring years are correlated (a wet year is often followed by a wet year), so there are fewer independent points than it seems and the simple interval is too narrow.

Without JavaScript: all 20 years give a trend of +0.032 per decade ± 0.016. Years 13–18, ending in the dry year, give −0.099 ± 0.184, and years 14–20 give −0.007 ± 0.166: both are consistent with no trend and neither refutes the long one. Years 1–10 give +0.036 ± 0.025, widening to ± 0.038 once autocorrelation is allowed for.

Uncertainty: a map is an estimate

A land-cover map is a classification, and classifications are wrong somewhere. The standard check is an accuracy assessment: pick sample locations, find out what is really there (from higher-resolution images or the field), and tabulate map class against reference class in a confusion matrix. User's accuracy says how many pixels mapped as "cleared" really were. Producer's accuracy says how much of the really cleared land was mapped. A single headline, overall accuracy, hides both.

The trap, set out in the good-practice guide by Olofsson and colleagues (2014), is that counting mapped pixels gives a biased area. If a rare class is mapped from a huge area of stable land, even a tiny leak of true change into the "stable" class adds up. The remedy is to estimate area from the sample, stratified by map class, with an interval.

Without JavaScript: with 85% user's accuracy and 0.8% missed clearing in the stable class, overall accuracy is 99.1%, yet the sample-based estimate of cleared land is about 1,642 ha, give or take about 870 ha, against 1,000 ha mapped.

The same thinking applies to everything a graduate meets: a weather station network that is dense in cities, a model that fills gaps between stations, a published figure for global forest loss, a population grid built by splitting census totals over a land-cover map. Each hides a method and an error. Ask for the confusion matrix, the interval and the reference data.

What a graduate just sees

• A pixel is an average, so anything smaller than the pixel is blended into its neighbours.
• Healthy vegetation is dark in red and bright in near-infrared; NDVI compresses that into one number, which saturates in dense canopy.
• Fine pixels, wide coverage and frequent visits trade off against each other.
• Every flat map distorts something: Mercator keeps angles and inflates area toward the poles.
• Latitude and longitude are angles; measure areas and distances in a projected coordinate system.
• "Near" in a GIS means a number you chose, and the answer can jump when you change it.
• A trend depends on its window, and neighbouring years are not independent.
• A map with 99% overall accuracy can still miss most of a rare class.

Where it shows up next

• Conservation & Applied Ecology (36/15): protected-area boundaries, habitat maps and deforestation alerts.
• Soils, Agriculture & Food Systems (36/18): crop monitoring from NDVI, drought watches and yield forecasts.
• Climate Science (36/14) and Hydraulics & Water Resources (34/67): gridded temperature and rainfall, snow cover, floodplain maps.
• In the world: the flood-risk map on a property listing, a wildfire perimeter, a ship-routing app, a city's tree-canopy audit, and how an insurer or a government checks that a "protected" forest still stands.

If you want to go deeper

• 📘 Textbooks. Thomas Lillesand, Ralph Kiefer and Jonathan Chipman, Remote Sensing and Image Interpretation (Wiley, 7th edition, 2015), and Paul Longley, Michael Goodchild, David Maguire and David Rhind, Geographic Information Science and Systems (Wiley, 4th edition, 2015). Check for a newer edition.
• 📘 Gentler. Mark Monmonier, How to Lie with Maps (University of Chicago Press, 3rd edition, 2018), a short and entertaining tour of projections, scale and persuasion.
• 🌐 Free training. NASA's Applied Remote Sensing Training (ARSET) offers courses on satellite data for environmental applications.
• 🌐 Free hands-on. The QGIS Training Manual works through vector and raster analysis in the free QGIS software, in 20 modules.