Methodology¶

This page describes the analytical approach used by the Spatiotemporal LULC Analysis plugin.

Overview¶

The plugin performs post-classification change detection on categorical land use/land cover rasters. Rather than comparing raw spectral data, it compares classified maps to identify where and how land cover classes have changed.

graph LR
    A[Classified Raster Year 1] --> C[Change Analysis]
    B[Classified Raster Year 2] --> C
    C --> D[Transition Matrix]
    C --> E[Area Statistics]
    C --> F[Spatial Outputs]

Analysis Framework¶

Spatial Units¶

The fundamental unit of analysis is the pixel. Each pixel in each input raster contains a categorical class value representing the dominant land cover type at that location.

Temporal Structure¶

The plugin analyzes change across intervals - pairs of consecutive years in your time series:

Input Years	Intervals Analyzed
2010, 2015, 2020	2010→2015, 2015→2020
2000, 2005, 2010, 2015	2000→2005, 2005→2010, 2010→2015

Additionally, a first-to-last comparison ignores intermediate years to show cumulative change.

Valid Pixel Set¶

For each interval, analysis considers only pixels that are valid in both years:

\[\Omega = \Omega_{t_0} \cap \Omega_{t_1}\]

Where:

\(\Omega_{t_0}\) = pixels valid at time 0 (not NoData, within AOI)
\(\Omega_{t_1}\) = pixels valid at time 1 (not NoData, within AOI)
\(\Omega\) = pixels analyzed for this interval

This ensures fair comparison by excluding pixels with missing data in either year.

Core Metrics¶

Area by Class¶

For each year, the plugin counts how many pixels belong to each class:

Pixel count: Number of pixels with that class value
Area: Pixel count × pixel area (in selected units)
Percentage: Class area / total valid area × 100

Transition Accounting¶

For each interval, every pixel falls into one of two categories:

Persistence: Same class in both years
Transition: Different class between years

The transition matrix records all pixel movements:

	To Class A	To Class B	To Class C
From Class A	Persistence	A→B	A→C
From Class B	B→A	Persistence	B→C
From Class C	C→A	C→B	Persistence

Gain and Loss¶

From the transition matrix, we derive:

Gain for class k: Sum of column k (excluding diagonal)
Loss for class k: Sum of row k (excluding diagonal)
Net change: Gain - Loss
Gross change: Gain + Loss

Change Intensity¶

Intensity metrics normalize change by total area and time:

Interval intensity: Fraction of pixels that changed
Annual intensity: Interval intensity ÷ number of years

Spatial Analysis¶

Change Frequency¶

The change frequency raster counts how many times each pixel changed across the entire time series:

Value 0: Pixel never changed (persistent)
Value 1: Pixel changed once
Value n-1: Pixel changed in every interval

This highlights areas of repeated change vs. stable areas.

Change Hotspots¶

Hotspot maps use Kernel Density Estimation (KDE) to create smooth surfaces showing change concentration:

Changed pixels are converted to point locations
A Gaussian kernel is applied to each point
Kernels are summed to create a continuous density surface
Higher values indicate more concentrated change

Parameters:

Kernel: Gaussian
Radius: 1000 map units
Sample limit: 50,000 points (for performance)

Processing Approach¶

Block-Based Processing¶

To handle rasters of any size without exhausting memory, the plugin uses block-based processing:

Rasters are divided into 256×256 pixel blocks
Each block is processed independently
Results are aggregated across all blocks
Memory usage remains bounded regardless of raster size

+-------+-------+-------+-------+
| Block | Block | Block | Block |
| (0,0) | (1,0) | (2,0) | (3,0) |
+-------+-------+-------+-------+
| Block | Block | Block | Block |
| (0,1) | (1,1) | (2,1) | (3,1) |
+-------+-------+-------+-------+

Data Flow¶

graph TD
    A[Input Rasters] --> B[Block Iterator]
    B --> C[Read Block from All Rasters]
    C --> D[Apply NoData Mask]
    D --> E[Apply AOI Mask]
    E --> F[Compute Block Statistics]
    F --> G[Aggregate Results]
    G --> H[Write Outputs]

Methodological Considerations¶

Classification Consistency¶

The plugin assumes that:

Class definitions are consistent across all input years
The same class ID represents the same land cover type
Classification accuracy is acceptable for change detection

Garbage In, Garbage Out

Change detection quality depends entirely on input classification quality. Misclassification in either year appears as false change.

Scale Effects¶

Results depend on the spatial resolution of input data:

Coarse resolution: May miss small-scale changes
Fine resolution: Captures detailed patterns but increases processing time
Mixed pixels: Pixel values represent dominant cover type

Temporal Considerations¶

Interval length matters: Longer intervals capture more cumulative change
Seasonality: Ensure comparable acquisition dates across years
Intermediate changes: May be missed if not captured by input years

Comparison with Other Methods¶

Approach	This Plugin	Alternative Methods
Input data	Classified rasters	Raw imagery, spectral data
Change detection	Post-classification comparison	Image differencing, trajectory analysis
Metrics	Area, transitions, intensity	Spectral change vectors, time series analysis
Spatial outputs	Frequency, hotspots	Change maps, trend surfaces

References¶

The methods implemented in this plugin draw from established LULC change analysis literature:

Post-classification comparison is a standard approach for categorical change detection
Transition matrices follow conventional from-to accounting frameworks
Change intensity metrics are normalized measures suitable for cross-study comparison
Kernel density estimation provides spatial smoothing for pattern visualization

For detailed formulas, see the Formulas page.