Earth observation guideAnalysis

Analysis methods

A processed image is the start of the work. A handful of standard methods turn it into an answer, and this page explains the ones a product description is most likely to name.

The basics

An analyst usually starts by looking. Shown in false colour, plants, water and bare ground stand out from each other at a glance. The methods below then do the same sorting by rule, for every pixel, so it can be repeated over a whole country or a decade of images. An index compares two colours to pick out one thing, such as how green the plants are or where there is open water. Classification sorts every pixel, or every shape in the image, into kinds such as water, forest, crops and buildings. Either the analyst shows the computer examples of each kind first, or the computer groups similar pixels by itself and a person names the groups afterwards. Change detection compares two dates and marks what is different. A time series looks at many dates, learns the usual rhythm of the seasons and flags a change that breaks it. A mosaic stitches many scenes into one picture of a large area, and a composite picks, for each spot, the clearest view from many dates so the clouds disappear. Pan-sharpening, explained on the optical imaging page, adds the detail of a sharp grey image to a coarser colour one.

Why it matters

A product called a vegetation index, a land cover map, a change alert or a cloud-free mosaic is named after the method that made it. The method decides what the numbers mean. Two methods applied to the same images can give different answers, so a buyer should know which one was used and how its result was checked. The AI and machine learning page covers the methods that learn from examples.

Change detection in its simplest form. The same nine fields on two dates, and a map of the three that changed between them.

In depth

Spectral indices

An index combines two or more bands into one number that responds strongly to the property of interest and weakly to everything else. The best known is the normalised difference vegetation index, which uses the contrast between red light, absorbed by chlorophyll, and near-infrared light, scattered strongly by leaves [1].

\mathrm{NDVI} = \frac{\rho_{\mathrm{NIR}} - \rho_{\mathrm{red}}}{\rho_{\mathrm{NIR}} + \rho_{\mathrm{red}}}

(1)

where $\mathrm{NDVI}$ is the normalised difference vegetation index, without unit; and $\rho_{\mathrm{NIR}}$ and $\rho_{\mathrm{red}}$ are the reflectances in the near-infrared and red bands, each a fraction between 0 and 1 without unit.

Water indices follow the same pattern. Xu replaced the near-infrared band in the index of McFeeters, set out as equation 3 on the multispectral imaging page, with a shortwave-infrared band. The modified index enhances open water while suppressing the noise from built-up land, vegetation and soil, so it suits water among towns [2].

\mathrm{MNDWI} = \frac{\rho_{\mathrm{green}} - \rho_{\mathrm{SWIR}}}{\rho_{\mathrm{green}} + \rho_{\mathrm{SWIR}}}

(2)

where $\mathrm{MNDWI}$ is the modified normalised difference water index, without unit; and $\rho_{\mathrm{green}}$ and $\rho_{\mathrm{SWIR}}$ are the reflectances in the green and shortwave-infrared bands, each a fraction between 0 and 1 without unit.

An index is only as comparable as the reflectances behind it. Values from different sensors, or from images at different processing levels, differ because of band widths and atmospheric correction, not only because of the ground.

Classification

Classification assigns each unit of an image to one of a set of classes. In supervised classification the analyst supplies training samples of known class and the classifier learns to separate them, while in unsupervised classification an algorithm groups pixels by spectral similarity and the analyst labels the clusters afterwards. Non-parametric classifiers such as neural networks and decision trees have become important, especially when several data sources are combined, and the choice of method and of input features matters a great deal for accuracy [3]. The unit of classification is the other choice. Per-pixel methods suit imagery whose pixels are about as large as the objects of interest or larger. As pixels became smaller than the objects, object-based image analysis took hold. It first segments the image into groups of similar neighbouring pixels and then classifies those objects using their context as well as their spectra [4].

Change detection

The simplest change detection subtracts one co-registered image from another and treats large differences as change. A common rule sets the threshold a chosen number of standard deviations from the mean difference, since pixels that did not change cluster around the mean and those that did sit in the tails [5].

D = x_2 - x_1,\quad |D - \mu_D| > k\,\sigma_D

(3)

where $x_1$ and $x_2$ are the values of a pixel on the first and second dates, in the same band or index and the same unit; $D$ is their difference, in that unit; $\mu_D$ and $\sigma_D$ are the mean and standard deviation of $D$ over the image, in that unit; and $k$ is the chosen number of standard deviations, without unit.

Other methods compare ratios or principal components of the two dates, or classify each date and compare the two maps. A review of these methods found that different procedures produce different maps of change even in the same place [5], so the method is part of the result.

Time series

Two dates cannot tell a harvest from a clearance. Methods that use every available image fit a model of the normal seasonal cycle and trend to each pixel and flag a break when new observations depart from it, as Figure 1 sketches. The continuous change detection and classification method fits harmonic and trend terms to all available Landsat observations of a pixel and treats a sudden departure from the predictions as abrupt change [6]. Its simplest form, with one harmonic, is

\hat{\rho}(t) = a_0 + a_1\,\cos\frac{2\pi t}{T} + b_1\,\sin\frac{2\pi t}{T} + c_1\,t

(4)

where $\hat{\rho}$ is the predicted reflectance or index of the pixel at time $t$ , without unit; $t$ is the date of the observation, in days; $T$ is the number of days in a year; $a_0$ is the mean level and $a_1$ and $b_1$ set the size and timing of the seasonal cycle, all without unit; and $c_1$ is the gradual trend, per day.

A related method, breaks for additive season and trend, splits a series into trend, seasonal and remainder parts and detects abrupt changes in both the trend and the seasonal pattern [7].

Figure 1. A time series model in sketch. Dots are observations of a vegetation index at one place over four years. A seasonal curve fits them until the break, after which a new, flatter curve fits the rest, as when forest is cleared. The values are illustrative, not measured.

Mosaics and composites

A mosaic joins neighbouring scenes into one image of a large area, matching their geometry and, as far as possible, their brightness along the seams. A composite chooses, for each pixel, one value from many dates. The maximum value composite keeps the observation with the highest vegetation index over a period of days. The technique minimises cloud contamination and reduces the effects of viewing angle, sun angle, shadow, aerosol and water vapour [8]. A composite pixel may therefore come from a different day from its neighbour, which matters for anything that changes quickly. The processing levels page places composites at level 3.

Pan-sharpening

The optical imaging page explains why a satellite carries a sharp panchromatic band beside coarser multispectral ones. Pan-sharpening methods fall into two main families. Component substitution replaces a component of the multispectral image with the panchromatic band, and multiresolution analysis injects the spatial detail extracted from the panchromatic band into each multispectral band. Their results are judged by protocols at full and at reduced resolution. Five data sets from different satellites show that performance varies with the instrument [9].

References

Tucker, C. J. (1979) Red and photographic infrared linear combinations for monitoring vegetation. Remote Sensing of Environment, 8(2), 127-150. https://doi.org/10.1016/0034-4257(79)90013-0
Xu, H. (2006) Modification of normalised difference water index (NDWI) to enhance open water features in remotely sensed imagery. International Journal of Remote Sensing, 27(14), 3025-3033. https://doi.org/10.1080/01431160600589179
Lu, D. and Weng, Q. (2007) A survey of image classification methods and techniques for improving classification performance. International Journal of Remote Sensing, 28(5), 823-870. https://doi.org/10.1080/01431160600746456
Blaschke, T. (2010) Object based image analysis for remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing, 65(1), 2-16. https://doi.org/10.1016/j.isprsjprs.2009.06.004
Singh, A. (1989) Digital change detection techniques using remotely-sensed data. International Journal of Remote Sensing, 10(6), 989-1003. https://doi.org/10.1080/01431168908903939
Zhu, Z. and Woodcock, C. E. (2014) Continuous change detection and classification of land cover using all available Landsat data. Remote Sensing of Environment, 144, 152-171. https://doi.org/10.1016/j.rse.2014.01.011
Verbesselt, J., Hyndman, R., Newnham, G. and Culvenor, D. (2010) Detecting trend and seasonal changes in satellite image time series. Remote Sensing of Environment, 114(1), 106-115. https://doi.org/10.1016/j.rse.2009.08.014
Holben, B. N. (1986) Characteristics of maximum-value composite images from temporal AVHRR data. International Journal of Remote Sensing, 7(11), 1417-1434. https://doi.org/10.1080/01431168608948945
Vivone, G., Alparone, L., Chanussot, J. et al. (2015) A Critical Comparison Among Pansharpening Algorithms. IEEE Transactions on Geoscience and Remote Sensing, 53(5), 2565-2586. https://doi.org/10.1109/TGRS.2014.2361734