GISCI 241 Chap.9 Raster Analysis and Change Detection
Raster Analysis and Change Detection
Raster analysis can operate locally on corresponding cells or focally on a neighbourhood. Mean and median filters reduce local variation, while weighted kernels can enhance boundaries. Change detection adds a temporal comparison, but valid subtraction requires compatible registration, spatial support, bands and acquisition conditions.
This chapter calculates a multiband Euclidean change vector, explains why magnitude loses direction, and separates detecting a difference from attributing its cause. Thresholds and validation turn continuous differences into defensible mapped change.
What this chapter covers
- 01
Local cell operations and aligned raster arithmetic
- 02
Focal neighbourhoods, means, medians and kernels
- 03
Smoothing, edge enhancement and lost detail
- 04
Registration, support and spectral comparability
- 05
Signed differences and Euclidean change magnitude
- 06
Threshold choice, validation and causal attribution
Worked example · free
Compute a two-band change vector
- +1First band difference equals 50 minus 62, giving negative 12.
- +1Second band difference equals 58 minus 34, giving positive 24.
- +1Square the differences to obtain 144 and 576.
- +1Sum the squared differences to obtain 720.
- +1Take the square root to obtain magnitude 26.83 value units.
- +1Retain the signed vector because magnitude alone loses response direction.
Key terms
- Local operation
- A raster transformation using values at one corresponding cell location across one or more layers.
- Focal operation
- A raster transformation using values from a defined neighbourhood around each target cell.
- Kernel
- A small matrix of weights applied to a moving neighbourhood for smoothing or enhancement.
- Image registration
- Geometric alignment that makes corresponding pixels represent the same ground locations.
- Image difference
- A signed subtraction of compatible pixel values between dates or layers.
- Change magnitude
- A nonnegative summary of multiband displacement, often calculated with Euclidean distance.
- Change threshold
- A stated cutoff that converts a continuous change measure into changed and unchanged categories.
Raster Analysis and Change Detection FAQ
What is the difference between local and focal analysis?
A local operation uses the target cell and aligned values at that location, such as band arithmetic or image subtraction. A focal operation uses surrounding cells in a window, such as a mean, median or edge kernel. Window size sets the spatial scale and must be reported.
Why does misregistration create false change?
A stable sharp feature occupies slightly different pixels in the two images. Subtraction then removes the feature from one side and adds it on the other, producing paired negative and positive edges. Improve geometric alignment and quantify residual error before interpreting or changing the detection threshold.
Does a large change magnitude reveal the cause?
No. Magnitude says that a multiband response moved far in feature space. It loses direction and does not distinguish fire, harvest, flood, phenology, atmosphere or processing mismatch. Retain signed differences and combine timing, spatial context and independent evidence before attributing a process.
What should be checked before choosing a change threshold?
First establish registration, spatial support, band compatibility, seasonal and atmospheric comparability, and a no-change reference distribution. Then inspect the continuous signed differences and the costs of false alarms versus missed change. A threshold is a decision rule supported by evidence, not a natural boundary revealed automatically by the image.
Exam move
Work one 3 by 3 focal mean and median manually, then explain which features each filter suppresses. Sketch an edge kernel and state that enhancement changes contrast rather than creating new information. For change detection, memorise the prerequisite audit: registration, spatial support, spectral compatibility, season and atmosphere, and reference evidence.
Calculate signed band differences before Euclidean magnitude and keep direction in the interpretation. Practise spotting registration ghosts at stable edges and classification flicker between dates. When reviewing a change map, ask how the threshold was chosen, what unchanged reference areas show and whether detection has been incorrectly written as cause.
Repeat the focal calculation with one extreme value and compare mean and median outcomes. Translate the window from cells into ground dimensions so the smoothing scale is explicit. For edge enhancement, identify which positive and negative weights respond to a boundary and explain why the output is derived contrast rather than a newly observed feature.
Build a change-preparation checklist before touching the formula: align geometry, match support, confirm band definitions, control atmosphere and season, and select stable reference areas. Work two-band and three-band difference vectors, keeping signed components beside the Euclidean magnitude. Draw a false-registration example with positive and negative stripes around a stable road.
Then compare continuous difference, thresholded change and post-classification transition maps, noting the new assumption introduced at each step. Choose a threshold from reference evidence or a stated error cost, and inspect values just below it. Finish with separate statements for detection and attribution.
The image may locate an unusual change signal, while field observations, event records or another sensor are needed to establish its cause.