Satellite observations

In October 1975 a New Zealand Landsat II scene entered a computing world that looked nothing like modern GIS. The DSIR Physics and Engineering Laboratory programme handled satellite data through computer-compatible tapes, IBM mainframe processing, minicomputers and programs written in PL/I. Images had to be rectified, classified and interpreted by specialists who understood both the sensor and the limitations of the available computers. That early work established that satellite observation could contribute to New Zealand mapping and environmental analysis, but there was nothing routine about it. Obtaining a scene was only the beginning. Turning it into useful geographic information was a substantial technical project in its own right.

By the 2020s the same country could be represented by repeated national satellite mosaics, multispectral time series and land-cover products available through ordinary data services. A GIS analyst no longer needed to begin with magnetic tape or build a complete processing chain simply to look at a new observation. Satellite earth observation did not suddenly become useful when a particular sensor was launched, and it did not replace aerial photography. It became routine when repeated observations could be found, processed, compared and incorporated into continuing geographic datasets and operational workflows.

By the mid-1990s remote sensing was already an established specialist method in New Zealand research. Leonard Brown’s 1995 Massey University doctoral work combined remote sensing, digital image processing and GIS to investigate resource-management questions in Westland, including alluvial gold mining and indigenous forest. Other mid-1990s Massey work used digital image analysis and GIS to examine land-use change on the Hautere Plains. These projects were no longer demonstrations that a satellite could see New Zealand. They treated imagery as evidence that could be analysed with other geographic information to answer a specific land-management question.

The distinction between image and information remained important. A satellite records energy reflected or emitted from the Earth’s surface in particular wavelength bands. The resulting pixels do not arrive with labels saying forest, pasture, wetland or urban land. Classification requires rules, reference information and checking. A change between two images may represent a genuine change on the ground, but it may also reflect cloud, season, illumination, sensor differences, image alignment or the way the classification was produced. As satellite data became easier to obtain, those interpretive problems did not disappear. They became easier for more people to encounter.

The Land Cover Database provided a maintained national record of land cover. LCDB1 was completed in 2000 using SPOT imagery acquired during the summer of 1996/97. LCDB2 was released in July 2004 from Landsat 7 ETM+ imagery acquired between September 2001 and March 2002. The second database did more than provide a newer picture. It corrected parts of the earlier classification and created a basis for identifying changes over roughly five years.

LCDB was a classified and checked land-cover database derived from satellite imagery and other evidence. The satellite imagery was interpreted into a thematic database, checked and edited, then maintained as a GIS dataset with consistent classes. The result could be combined with protected areas, land environments, administrative boundaries and other spatial information. By 2007 the Ministry for the Environment was using the first two LCDB epochs as national environmental indicators for land cover, land use and biodiversity reporting. Users needed interpreted data that could answer their questions. It was a repeated national classification built from imagery.

That repeated character became more important as the series continued. Ministry for the Environment material describes LCDB as using satellite imagery to produce national land-cover maps approximately every five years from the 1996/97 baseline. Later editions added 2008 and 2012 epochs, and version 5, launched in 2020, extended the maintained record again. A single satellite scene can show what was visible on one date. A controlled national series can show where classes have changed and where apparent change needs further checking. Once that series exists, earth observation becomes part of environmental information infrastructure rather than a collection of interesting images.

The Land Use and Carbon Analysis System pushed repeated earth observation into another kind of national obligation. LUCAS was developed so New Zealand could measure and report land use, land-use change and forestry for international greenhouse-gas accounting. Ministry for the Environment documentation describes national land-use maps for nominal dates at the end of 1989, 2007, 2012, 2016 and 2020. Satellite imagery and aerial photography were interpreted with other spatial information to establish classes and identify change, particularly where forest gain or loss affected carbon reporting.

This is a useful example because it exposes the work between sensor and statistic. The 2012 LUCAS interpretation guide did not treat imagery as self-explanatory. It defined land-use classes and described how interpreters should identify them from remotely sensed evidence. Later mapping used change detection, manual checking and additional information to decide whether a spectral difference represented an actual land-use change. The 2020 land-use map, commissioned in 2022 and documented in 2024, used Sentinel-2 imagery and more recent analytical methods, including automated and deep-learning steps, but still retained quality-control and review processes. More automation did not turn national land-use mapping into a one-button classification problem.

The value of repeated earth observation depended on keeping the comparison stable enough to mean something. LCDB made that visible. Each later edition required cross-date checking so apparent change could be separated from differences in imagery, classification and interpretation. If an earlier polygon had been misclassified, leaving the mistake untouched would create a false change when the later image was interpreted correctly. LCDB2 therefore included corrections to LCDB1 as well as new mapping. The database became a maintained historical record in which improving the baseline was part of producing a credible change series.

That creates an unusual information-management problem. A normal corporate database is often expected to replace an old value with a better one. A time-series land-cover database has to preserve the historical state while also ensuring that differences between epochs represent changes on the ground rather than changes in interpretation. Classification definitions, polygon boundaries, imagery dates and editing rules were needed to interpret the results. The GIS layer may look simple when displayed, but its usefulness depends on the discipline applied between each observation date.

LUCAS developed the same problem under the pressure of national greenhouse-gas reporting. A land-use change had consequences beyond cartography because forest establishment, harvesting or conversion affected emissions and removals reported internationally. The mapping process therefore combined satellite imagery with aerial photography and other spatial information, and interpretation guidance defined how land-use classes should be recognised. The nominal map dates, such as the end of 1989 or 2007, should not be confused with a single image captured on New Year's Eve. A national map assembled for a reporting date may draw on imagery from a window around that date, with interpretation used to establish the best defensible state.

By the 2020 map, the technical assistance available to that process had expanded considerably. Sentinel-2 supplied repeated multispectral coverage, and the documented methodology included automated and deep-learning steps alongside manual interpretation and quality assurance. That is a useful measure of how far the workflow had moved from the 1970s. The computer was no longer struggling simply to turn tape into a rectified image. It could help identify likely change across the country. Human review still remained because the consequence of classifying the wrong patch of scrub, forest or grassland was not fixed by giving the algorithm more confidence.

For much of the remote-sensing era, imagery access itself constrained what could be attempted. A project might purchase selected scenes, work with the dates that could be afforded and store only what was needed for the immediate task. In 2008 the United States Geological Survey changed that calculation for Landsat. From 1 October, all Landsat 7 ETM+ scenes in the USGS EROS archive were made available at no charge. The wider programme then opened the Landsat 1-5 archive for free electronic access through the remainder of that year.

Remote sensing still required processing, storage and skilled interpretation. Someone still had to find the right scenes, download them, manage large raster files, deal with cloud and sensor problems, process them and understand what comparisons were valid. It did, however, remove the need to treat every additional historical scene as another purchase decision. A researcher could assemble a longer sequence, revisit an earlier year or test several dates before deciding which observations were usable. For New Zealand, where cloud can make optical imagery an exercise in patience, the ability to look across more acquisitions was particularly valuable.

The archive also changed historical analysis. Landsat’s long record meant a contemporary project could reach back into earlier decades without having collected the data itself at the time. The satellite had been repeatedly observing the country whether or not a particular New Zealand agency had an active project underway. Once those observations were easily accessible, the archive became a time machine with some fairly obvious caveats: different sensors, different seasons, cloud, missing data and changing preprocessing all had to be handled properly. Users gained more frequent opportunities to obtain suitable observations. It was the existence of a long, repeated record that could be interrogated.

The next shift was frequency and spectral availability. Sentinel-1A launched in April 2014 with synthetic-aperture radar, providing an all-weather, day-and-night observing capability. Sentinel-2A followed in June 2015 with a 13-band multispectral instrument designed for land monitoring. These European programme milestones expanded the open satellite data available to New Zealand practitioners. Sentinel-2 in particular provided 10 metre optical bands useful for regional vegetation, land-cover and coastal work, with repeated coverage that made it practical to look for suitable cloud-free observations across a season rather than rely on a single acquisition.

Resolution still needs some discipline. A 10 metre satellite pixel does not compete directly with a 10 centimetre council orthophoto. The orthophoto can show individual structures and small surface features that the satellite cannot resolve. Sentinel-2 contributes something different: repeated national coverage, consistent spectral bands and manageable data volumes. For mapping changes in vegetation, land cover or broad agricultural patterns, those qualities may be more useful than centimetre detail. The better dataset depends on the question, not the smallest number printed beside the word resolution.

Spectral information also changes what counts as an image. Ordinary red, green and blue display is only one representation of a multispectral sensor. Near-infrared and short-wave infrared bands respond differently to vegetation, water and moisture. Indices such as NDVI or NDRE combine selected bands to highlight particular patterns. They can be very useful, but they are not direct readouts of plant health, crop yield or biomass. Their interpretation depends on the vegetation, season, atmosphere, sensor and purpose of the analysis. Routine access makes it easy to calculate an index. It does not make the meaning of that index automatic.

Users selected suitable imagery from the more frequent coverage. New Zealand’s cloud, mountains, coast and strong seasonal differences make date selection part of the analysis. A nominal five-day or ten-day revisit does not mean a useful optical image appears every five or ten days. A pass may be obscured by cloud, affected by haze or captured at a time of year that makes comparison with an earlier scene misleading. The archive can therefore contain many observations of a place while only a smaller subset is suitable for a particular question.

This changed the practitioner’s job from ordering the one affordable scene to screening a sequence. The user could inspect cloud cover, acquisition date and sensor characteristics, then choose several observations rather than treating the first available image as definitive. That is one reason free archives and repeated Sentinel coverage reinforced each other. The archive reduced the cost of looking again. The sensor programme increased the number of opportunities to find a useful view.

Mosaics introduced another layer of judgement. A national satellite mosaic may appear continuous on screen even though neighbouring areas were captured on different dates. Cloud removal can require selecting pixels or tiles from several acquisitions. Vegetation may be at different stages of growth, snow may differ between mountain ranges, river colour may reflect different flow conditions and coastal water may have been recorded under different tides. The seamless visual result is a composite rather than a photograph of the whole country at one moment.

The LINZ 2017/18 Sentinel-2 product makes that distinction unusually clear. Its national 10 metre layer was assembled from acquisitions spanning December 2016 to December 2017. The purpose was to create a consistent cloud-free visual layer, not to preserve one simultaneous national observation. That approach is entirely sensible for a basemap. It would be inappropriate to use the same mosaic uncritically for an analysis requiring every location to represent the same season or event date. Routine earth observation therefore increased the supply of imagery while making metadata about time even more important.

By the 2010s New Zealand earth-observation work increasingly treated images as inputs to repeatable data production. LUCAS provides a national example. Coastal research provides a more specialised one. Nam-Thang Ha’s 2021 University of Waikato doctoral work examined seagrass dynamics and above-ground biomass in Tauranga Harbour using Landsat and Sentinel imagery, GIS and machine-learning methods. The research combined observations through time rather than treating a satellite image as a static background map.

The mature workflow is worth distinguishing from the early idea of simply viewing an image. Raw sensor measurements may first be corrected and geometrically aligned. Bands can then be displayed as natural or false-colour composites, transformed into indices or passed into classification and modelling processes. The resulting output might be a categorical land-cover layer, an estimate, a probability surface or a change map. By the time the output appears in GIS, the original satellite measurement may be several analytical steps behind the product the user actually works with.

That chain introduces opportunities for error as well as capability. A land-cover polygon is not true merely because it was derived from a satellite. A machine-learning classification can be wrong in consistent and convincing ways. A time-series trend can reflect inconsistent cloud masking or seasonal timing. A national dataset therefore needs metadata explaining the image dates, methods, classes and accuracy, and users still need to know whether the product is fit for their question. Earth observation became routine partly because these processing chains became easier to repeat. It became dependable only where those chains remained controlled.

In 2017/18 satellite imagery crossed another institutional boundary. LINZ reported that, for the first time, it had released satellite earth-observation data through the LINZ Data Service. Period presentation material identifies the pilot as a seamless cloud-free 10 metre imagery layer covering mainland New Zealand and offshore islands. It was assembled from Sentinel-2A and Sentinel-2B acquisitions between December 2016 and December 2017 and published as 450 orthorectified RGB GeoTIFF tiles in NZTM.

The wording needs care. “Cloud-free” described a mosaic assembled from multiple acquisitions rather than a satellite capture of all New Zealand on one miraculous afternoon. The publicly distributed 8-bit RGB version was primarily a visual product, and LINZ documentation noted that original values had been modified for visualisation. Higher-bit RGB and near-infrared imagery could be requested separately. Those details separate a useful national basemap from an analytical multispectral product.

New Zealand organisations had used satellite imagery before LINZ’s later services. New Zealand scientists had been processing Landsat more than forty years earlier. The change is institutional. A national land-information agency was now distributing a satellite-derived national layer through the same public data infrastructure used for topography, cadastre, elevation and aerial imagery. Earth observation had moved close to the centre of ordinary geospatial distribution. The user could encounter a national Sentinel mosaic as another available dataset rather than as the result of a bespoke remote-sensing procurement.

The same trajectory is visible in the Ministry for the Environment’s later national satellite-data search service. Its maintained archive now points users across a sequence that includes Landsat 4 imagery from 1990, Landsat 7, SPOT and DMC datasets through the 2000s and early 2010s, Landsat 8, and annual Sentinel-2 national datasets from 2016 onward. That service should not be read as proof that all those datasets were originally published in the year they were captured. It demonstrates something else: by the 2020s, national satellite observations from several decades were being organised as an accessible continuing resource.

Optical imagery still carries one stubborn New Zealand limitation. Cloud is not impressed by a project deadline. Repeated coverage increases the chance of finding useful optical observations, but a wet season or mountainous region can still frustrate attempts to build a clean time series. Synthetic-aperture radar offers a different observing method. Sentinel-1 actively transmits microwave energy and measures the return, allowing it to acquire data at night and through cloud that blocks ordinary optical imagery.

Radar observations support surface-change, moisture, flood, maritime and deformation analysis. By the Sentinel era, New Zealand practitioners could obtain both optical multispectral and radar observations. The cases described here used optical imagery.

A later agricultural example shows how normal sensor fusion had become. Simon Tenbusch’s 2025 Lincoln University precision-agriculture dissertation examined management-zone creation for maize in the Whakatāne area using Sentinel-2 NDRE, one metre LiDAR elevation information, yield data and GIS/statistical modelling. Satellite imagery was one input among several. The analysis did not ask whether satellite imagery or LiDAR was the better technology. It combined observations at different scales to support a specific management question.

This is a useful endpoint because it reverses the early Landsat problem. In the 1970s the challenge was getting satellite measurements through specialist computers and turning them into something map-like. In the mature workflow, satellite data arrive as one component in a larger analytical stack. The practitioner chooses bands, dates and derived measures alongside terrain and production information. The sensor becomes less visible because the analysis is organised around the agricultural problem rather than around the novelty of obtaining an image from space.

By the 2020s the accumulation itself had become useful infrastructure. The Ministry for the Environment’s satellite-data search service brings together national image series that span Landsat 4 material from 1990, Landsat 7, SPOT and DMC coverage through the 2000s and early 2010s, Landsat 8 and annual Sentinel-2 datasets from 2016 onward. The service is a modern access point to imagery acquired under different programmes and at different times. Those historical datasets originally arrived through several distribution systems; the modern service makes them accessible as a sequence.

That changes how a project can begin. An analyst investigating coastal change, land-cover conversion or vegetation response does not necessarily start by asking which new image should be purchased. The first question can be what observations already exist, which dates are comparable and whether the archive is adequate before new acquisition is considered. Earlier projects often had to build an image collection around the immediate budget. Later projects could begin with decades of accumulated observation and then decide where the record was incomplete.

The growing archive also made older decisions easier to revisit. A land-cover boundary mapped in one decade could be compared with earlier and later observations. A classification method could be rerun on an older image using newer tools, provided differences in sensor and preprocessing were handled honestly. That possibility does not make history perfectly measurable from space. It does mean that New Zealand’s landscape has an increasingly dense digital observational memory, much of it created before anyone knew exactly which future questions would be asked of it.

A longer archive also increases the temptation to treat every sensor as though it were measuring the landscape in the same way. Landsat, SPOT and Sentinel differ in pixel size, spectral bands, calibration, revisit pattern and processing history. A vegetation boundary visible in one system may be represented differently in another even when nothing has changed on the ground. Long-term analysis therefore depends on harmonising the observations or choosing methods that acknowledge those differences. The archive makes comparison possible; it does not make comparison automatic. That distinction is another reason maintained products such as LCDB and LUCAS became valuable. They carried forward an interpretation framework as well as a collection of images.

The most durable contribution of satellite earth observation is repetition. Aerial photography remains indispensable when a project needs very fine detail, while LiDAR and drones provide other ways to measure the surface. Satellites contribute a different rhythm. The same country is observed again and again, allowing analysts to look for change in land cover, forest extent, vegetation condition, coastal environments and other broad surface patterns. The archive keeps growing even when no particular local project is asking a question at that moment.

That repeated record changed the questions GIS could ask. Instead of treating imagery only as a current backdrop, analysts could compare years, seasons and events. National reporting systems such as LCDB and LUCAS formalised that idea by converting repeated observations into maintained datasets with defined classes and dates. Research projects extended it through multispectral analysis, machine learning and sensor fusion. National services then made both rawer imagery products and derived information easier to discover. GIS users could analyse repeated observations as a time series.

Routine access still required users to understand the imagery and its limitations. A modern GIS package can display Sentinel-2, calculate vegetation indices and run classification tools with far less friction than the DSIR teams faced in the 1970s. That accessibility is useful, but it can disguise the number of assumptions built into the result. Optical imagery still needs cloud and atmospheric handling. Sensors differ. A summer-to-winter comparison can be technically precise and analytically useless. Training data and validation determine whether a classification deserves confidence.

Remote-sensing specialists continued to select, process and interpret observations. It is the widening of the boundary between specialist remote sensing and everyday GIS. More GIS practitioners can use satellite-derived products directly, while specialists build and validate the methods that make those products reliable. National agencies can publish mosaics and maintained classifications so every user does not need to rebuild the same processing chain. The difficult science increasingly sits behind datasets that look ordinary when opened in GIS.

That is why earth observation became routine. Repeated observations and accessible data services changed how imagery could be used. The country had satellite images in the 1970s and high-quality aerial photography long before that. The change came when repeated observation, open archives, multispectral sensors, national classification programmes and ordinary spatial-data services converged. A satellite scene stopped being an event. It became another dated observation in a growing geographic record.

Another kind of measurement was becoming equally important. LiDAR replaced reflected colour and broad spectral patterns with dense measurements of surface height, producing point clouds and terrain models detailed enough to alter flood mapping, engineering, forestry and landscape analysis. Earth observation had become routine from above while the shape of the ground itself was becoming newly measurable.