SPATIAL AUTOCORRELATION

Spatial autocorrelation measures the degree to which similar values (positive) or contrasting values
(negative) are clustered together in geographic space. Based on Tobler’s First Law of Geography, it
indicates that nearby locations are more related than distant ones. It is crucial for identifying spatial
patterns (clusters or dispersion).

Key Concepts and Types

i. Positive Spatial Autocorrelation: Similar values cluster together (e.g., high-income neighbourhoods
near each other).

ii. Negative Spatial Autocorrelation: Dissimilar values are next to each other, resulting in a
checkerboard pattern.

iii. Zero/Random Spatial Autocorrelation: No spatial pattern; data is randomly distributed.

iv. Global Moran’s I: Global Moran's I is a spatial statistics tool that measures spatial autocorrelation
based on both feature locations and values, determining if a pattern is clustered, dispersed, or random.
It computes a single index, generally ranging from -1 to +1, to indicate whether similar values
(positive) or dissimilar values (negative) are clustered together, or if data is randomly distributed.
Values significantly above indicate positive spatial autocorrelation (similar values cluster). Values
below indicate negative spatial autocorrelation (dissimilar values neighbour each other). The tool uses
spatial weights matrices (like inverse distance, fixed distance, or contiguity) to define neighbour
relationships.

v. Local Indicators of Spatial Association (LISA): Identifies specific "hot spots" or "cold spots" of
high/low values.

Importance of Spatial Autocorrelation

i. Model Accuracy: Violates the assumption of independence in traditional statistics, requiring spatial
regression methods to avoid biased results.

ii. Pattern Recognition: Essential for detecting non-random, underlying structures in environmental,
social, or economic data.

iii. Data Sampling: Helps determine if samples are too close, leading to data redundancy (e.g., in
remote sensing).

Commonly Used Tools

ArcGIS Pro: The "Spatial Autocorrelation (Global Moran's I)" tool.

R/Python: Libraries such as spdep and PySAL are used for analyzing spatial patterns.

QGIS: Hotspot Analysis Plugin– This plugin performs Local Indicators of Spatial Association (LISA)
statistics to find hotspot (Getis-Ord Gi*) and cluster (Moran's I) areas, allowing for direct mapping of
spatial autocorrelation.
1
Page
 SURFACE MODELLING
In Geographic Information Systems (GIS), surface modeling is the process of representing and
analyzing the distribution of a continuous variable over a defined geographic area.

Unlike discrete features (like a building or a road), surfaces represent phenomena that exist
everywhere and have values at every point. While we most commonly associate surface modeling
with elevation (Topography), it is also used for variables like temperature, rainfall, chemical
concentrations, or population density.

1. Key Data Structures

There are two primary ways to represent surfaces in a digital environment:

i. Raster (Digital Elevation Model - DEM): The surface is divided into a grid of identical square cells.
Each cell contains a single value representing the average for that area. It is computationally efficient
and great for mathematical modeling.

ii. Triangulated Irregular Network (TIN): The surface is represented by a set of non-overlapping
triangles. This method uses "mass points" at irregular intervals, allowing for higher precision in
complex areas (like sharp mountain ridges) while using fewer points in flat areas.

2. Core Modelling Techniques

Since we cannot measure every single point on Earth, surface modelling relies on Interpolation to fill
in the gaps between known data points:

i. Deterministic Methods: These use mathematical formulas based on distance (e.g., Inverse Distance
Weighting (IDW)).

ii. Statistical/Stochastic Methods: These use complex spatial statistics to account for the
"autocorrelation" or trends in data (e.g., Kriging).

3. Derivative Surface Analysis

Once a surface model is created, new information can be derived from it, which is essential for
geography and engineering:

 Slope: The degree of steepness at any given point.

 Aspect: The compass direction that a slope faces (crucial for solar energy or vegetation
studies).
 Hillshade: A 3D visual effect that mimics the shadows cast by the sun, making the terrain
easier for the human eye to interpret.
 Contour Lines: Lines connecting points of equal value, commonly seen on topographic maps.

Question: What is meant by surface modelling using spatial autocorrelation?

In the context of GIS and spatial analysis, surface modelling using spatial autocorrelation refers to
creating a continuous surface by leveraging the statistical relationship between nearby points.
2
Page
 While simple interpolation (like IDW) assumes things are more similar just because they are closer,
modelling via spatial autocorrelation—most notably through Kriging—actually measures how that
similarity changes over distance to make more accurate predictions.

1. The Concept of Spatial Autocorrelation

Spatial autocorrelation is based on Tobler’s First Law of Geography: "Everything is related to

everything else, but near things are more related than distant things.

 "Positive Autocorrelation: Similar values cluster together (e.g., if one spot in a field has high
soil salinity, the spot $1\text{ m}$ away likely does too).
 Negative Autocorrelation: Dissimilar values are next to each other (a "checkerboard"
pattern).Zero Autocorrelation: The values are random; there is no spatial pattern.

2. How it is used in Surface Modelling

When we use autocorrelation for surface modelling, we don't just "guess" the weight of a point based
on distance. We use a tool called a Semivariogram to quantify the relationship.

 Measuring the Trend: The system looks at all pairs of sample points and calculates the
variance between them as the distance between them increases.
 Fitting a Model: A line is fitted to these points on a graph (the Semivariogram). This line tells
us the "Nugget" (sampling error), the "Sill" (total variance), and the "Range" (the distance
beyond which points no longer affect each other).
 Kriging: This is the primary surface modelling technique that uses these autocorrelation
statistics. It assigns weights to surrounding points based on the actual statistical behaviour of
the data in that specific region, rather than a fixed mathematical formula.

3. Why this matters (The "Statistical Advantage")

Surface modelling via autocorrelation is often called Geostatistical Modelling. It offers two major
advantages over deterministic methods:

 Uncertainty Quantification: Because it is based on statistics, these models can generate a

second surface—a Standard Error Map. This tells where the surface model is reliable and
where it is "guessing" with low confidence.
 Handling Natural Variation: It is excellent for natural phenomena (like groundwater levels in
an aquifer or soil pH) where the change isn't perfectly smooth but follows a measurable
spatial trend.

Question: Discuss how Tobler’s First Law of Geography dictates spatial autocorrelation
and spatial modelling

Tobler’s First Law of Geography—"Everything is related to everything else, but near things
are more related than distant things"—serves as the philosophical and mathematical
foundation for almost all spatial analysis in GIS. It establishes the "logic" of space that allows
us to move from isolated data points to continuous models.

Here is how that law dictates the two pillars of spatial science:
3
Page

1. The Dictation of Spatial Autocorrelation

 Spatial autocorrelation is essentially the formal measurement of Tobler’s First Law. If the law
is the theory, autocorrelation is the proof.

 The Dependency Principle: Tobler’s Law asserts that data points are not independent.
In traditional statistics, we assume samples are random. In geography, we assume
spatial dependency. Autocorrelation quantifies this: a positive value confirms that
"near things" are indeed similar.
 The Concept of the "Range": Tobler’s Law implies that there is a limit to influence.
Spatial autocorrelation tools (like the Semivariogram) identify the specific distance—
the Range—where the relationship described by Tobler finally breaks down and
things become "distant" enough to be unrelated.
 Clustering vs. Dispersion: By measuring autocorrelation (e.g., using Moran’s I), we
can determine if a landscape follows Tobler’s Law (Clustered), defies it
(Dispersed/Negative Autocorrelation), or is simply chaotic (Random).

2. The Dictation of Spatial Modelling

Without Tobler’s Law, we could never create a map from sample data. The law provides the
"permission" to interpolate.

i. The Weighting Logic

Most spatial models (like IDW or Kriging) use a weighting system where weights decrease as
distance increases. This is a direct mathematical translation of the law. If we didn't believe
near things were more related, we would give equal weight to a sample $10\text{ km}$ away
as we would to a sample $10\text{ m}$ away.

ii. Prediction and Interpolation

Surface modelling is the act of predicting values at unsampled locations. We can only do this
because Tobler’s Law suggests that the value at an unknown point is likely an average of its
neighbours.

SPATIAL INTERPOLATION
Spatial interpolation is the process of using points with known values to estimate (predict) the
values at other unknown locations.

In geography and environmental science, it is physically and financially impossible to

measure every square inch of a landscape. Therefore, we collect samples at specific points—
such as weather stations, soil pits, or water quality sensors—and use interpolation to "fill in
the gaps" to create a continuous surface.

1. The Core Logic

4
Page
 Spatial interpolation is based on the assumption that spatially distributed objects are spatially
correlated. Following Tobler’s First Law of Geography, we assume that an unsampled point
will have a value more similar to its immediate neighbors than to points further away.

2. The Components of Interpolation

To perform spatial interpolation, it generally need three elements:

 Control Points: The known data points (e.g., X, Y coordinates and a value Z, like
elevation or temperature).
 Grid/Surface: The area where the user want to predict values.
 Mathematical Function: The algorithm (e.g., IDW, Kriging, Spline) that decides how
much influence each control point has on the unknown location.

3. Types of Interpolation Methods

Interpolation methods are broadly categorized into two groups:

A. Deterministic vs Statistical/Stochastic

i. Deterministic Methods

These create surfaces based on the geometric properties of the points or mathematical
formulas.

 Inverse Distance Weighting (IDW): Assigns weights to neighbors based on their

distance; closer points have more influence.
 Thiessen Polygons (Voronoi): Assigns the value of the nearest sample point to every
location within a surrounding polygon. No "blending" occurs between points.
 Spline: Fits a flexible mathematical surface through the points, similar to bending a
sheet of rubber to touch every data point.

ii. Geostatistical Methods

These are based on statistical models that include autocorrelation (the statistical relationship
between points).

 Kriging: This is the most popular geostatistical method. It doesn't just look at
distance; it looks at the spatial "trend" or variance in the data to provide a prediction
and a measure of certainty (error).

B. Global vs Local

i. Global interpolation:

Global interpolation uses the entire dataset to create a mathematical surface that covers the
whole study area.
5

 How it works: A single mathematical function is calculated that fits all the sample
Page

points simultaneously. If the user change the value of just one point—even one on the
 far edge of the map—it will theoretically change the value of every other point on the
predicted surface.
 Key Characteristic: They are designed to capture the general trend or "macro-
structure" of the data rather than local variations.
 Example: Trend Surface Analysis (Polynomial Regression).
 Best for: Identifying large-scale patterns, such as a steady increase in temperature
from south to north across a continent.

ii. Local interpolation:

Local interpolation uses only a small subset of nearby points (a "neighbourhood") to calculate
the value at an unsampled location.

 How it works: The algorithm moves across the map, creating a unique calculation for
each spot based on the nearest neighbours within a specified search radius. Changing
a data point at one end of the map will have zero effect on predictions at the other
end.
 Key Characteristic: They are designed to capture local variations and "micro-
structures." They are much more sensitive to "peaks" and "valleys" in the data.
 Examples: Inverse Distance Weighting (IDW), Kriging, and Nearest Neighbour.
 Best for: Most environmental and geographic tasks, such as mapping soil nutrients,
elevation, or water salinity in a specific lake.

Feature Global Interpolation Local Interpolation

Data Usage Uses every point in the dataset. Uses only nearby "neighbour" points.

Sensitivity Low; ignores local "outliers." High; captures local fluctuations.

Primary Goal To identify a broad, overall trend. To accurately predict specific locations.

Complexity Mathematically simple (one formula). More complex (repeated calculations).

Visual Result Usually a very smooth, simple surface. A detailed, often "bumpy" surface.

4. Practical Applications

 Meteorology: Creating a continuous temperature or rainfall map from a few dozen

weather stations.
 Hydrology: Mapping the depth of a water table based on well data.
 Geology: Estimating the concentration of minerals or pollutants in the soil.
 Urban Planning: Estimating population density or noise levels across a city.
6
Page

In summary, spatial interpolation is the tool that transforms point data (dots on a map) into
surface data (a complete picture of the landscape).
 MOVING AVERAGING IN SPATIAL INTERPOLATION

In the context of spatial interpolation, a moving average (often referred to as a "windowing"

or "sliding neighborhood" average) is a local, deterministic method used to estimate values
for a continuous surface.

While it shares the same name as the time-series tool used in finance, its spatial application is
focused on smoothing out data noise and identifying local trends across a 2D or 3D
landscape.

1. How it Works: The Sliding Window

The moving average method calculates the value of an unsampled location by taking the
mathematical mean of all known data points that fall within a specific search window (or
neighborhood).

 The Window: User defines a radius (circular) or a grid size (rectangular) around the
target point.
 The Calculation: All control points within that window are summed and divided by
the count of points.
 The Movement: The "window" slides across the entire study area, calculating a new
average for every cell in the output raster.

2. Simple vs. Weighted Moving Averages

In GIS, we distinguish between how much influence each point in the window has:

 Simple Moving Average: Every point inside the window is treated equally, regardless
of its distance from the centre. This is rarely used for precise mapping as it creates
"blocky" transitions.
 Weighted Moving Average (The Basis for IDW): Points closer to the centre of the
window are given more weight than points near the edge. Inverse Distance Weighting
(IDW) is essentially a sophisticated version of a weighted moving average.

3. Key Characteristics

 Smoothing Effect: Moving averages are primarily used to "smooth" data. They
remove extreme outliers (high peaks or deep pits) to show the general local behaviour
of the variable.
 Smoothing vs. Exactness: Unlike IDW (which is an "exact" interpolator), a simple
moving average is often a smoothing interpolator. It may produce a value at a sample
location that is different from the actual measured value because it is averaging that
point with its neighbours.
 Impact of Window Size:
o Small Window: Retains more detail but can be "noisy" if the input data has
errors.
o Large Window: Produces a very smooth, generalized surface but may hide
important local features (like a small hill or a specific pollution source).
7
Page

4. Applications in Geography and Engineering

 i. Digital Elevation Model (DEM) Generalization: If the user have high-resolution
LiDAR data that is too "jittery" (showing every bush or rock), a spatial moving
average can smooth the surface to show the actual terrain slope.
ii. Bathymetric Mapping: Smoothing out sonar soundings in a river bed to identify the
general channel path without getting distracted by small depressions in the silt.
iii. Shoreline Analysis: When reconstructing the Sundarbans coastline, a moving average
can help "average out" the jagged edges of digitized historical maps to find the most
probable mean coastline position.

INVERSE DISTANCE WEIGHTING (IDW)

Inverse Distance Weighting (IDW) is a deterministic method of spatial interpolation used to

estimate values at unsampled locations based on the values of nearby measured points.

It is one of the most commonly used tools in GIS (like ArcGIS or QGIS) because it is
mathematically straightforward and easy to interpret.

1. The Mathematical Logic

IDW is based entirely on Tobler’s First Law of Geography. It assumes that the influence of a
known point decreases as the distance to the unknown location increases.

The "weight" assigned to a sample point is inversely proportional to its distance from the
point the user is trying to predict.

2. The Power Parameter (p)

The most critical setting in IDW is the Power (p). This determines how quickly the influence
of a distant point drops off:

 Low Power (e.g., p=1): Distant points have a significant influence, resulting in a
"smoother" surface where values are averaged out across larger areas.
 High Power (e.g., p=2 or 3): The influence of distant points drops off very rapidly.
The resulting surface will be dominated by the closest points, creating more "local"
variation.

3. Key Characteristics of IDW

 Exact Interpolator: If the user perform IDW at the exact location of a sample point,
the result will be the exact value of that point. It does not "smooth over" known data.
 No Extrapolation: IDW cannot predict values that are higher than the maximum
sample value or lower than the minimum sample value. The predicted values will
always stay within the range of data.
 The "Bull's Eye" Effect: A common visual artefact of IDW is the appearance of
circular patterns (bull's eyes) around sample points that have significantly higher or
lower values than their neighbours.

4. When to use IDW

8
Page

IDW is best used when:

 1. The variable the user is measuring is locally influenced (like noise levels or chemical
spills).
2. The input has a dense and evenly distributed set of sample points.
3. The user wants a quick, deterministic result without performing the complex
statistical calculations required for methods like Kriging.

Question: Spatial Interpolation using Inverse Distance Weighting (IDW)

A researcher is estimating the elevation at a central point, Location X. Three nearby sampling stations
(A, B, and C) have provided the following observational data. Using the Inverse Distance Weighting
(IDW) method with a power parameter of p = 2, calculate the predicted value for Location X.

Sampling Station Distance from Location X (di) Observed Value (Zi)

Station A 10 m 1
Station B 20 m 2
Station C 40 m 4
9
Page
 KRIGING

Kriging is a sophisticated geostatistical method of spatial interpolation. Unlike deterministic

methods (like IDW) that rely on simple geometry, Kriging is based on statistical models that
include autocorrelation—the statistical relationship among measured points.

It is often called the "Best Linear Unbiased Predictor" (BLUP) because it aims to minimize
the error variance of the predicted values.

1. How Kriging Works: The Two-Step Process

Kriging is more complex than other methods because it requires a two-stage workflow:

Stage A: Variography (Spatial Autocorrelation)

Before predicting any values, Kriging analyzes the "spatial structure" of the data. It uses a
Semivariogram to plot how the difference between point values increases as the distance
between them increases.

 Nugget: The error or variation at a distance of zero (measurement error).

 Sill: The point where the variance levels off.
 Range: The distance at which points are no longer spatially related.

Stage B: Prediction

Once the statistical trend (the semivariogram) is established, Kriging assigns weights to
10

nearby points to calculate the unknown value. Unlike IDW, these weights are based on the
fitted model from Stage A, not just the raw distance.
Page
 Fig: Semivariogram

2. Key Features of Kriging

 Weighting by Trend: If three points are in a line, Kriging recognizes that the middle
point "shields" the far point, and adjusts weights accordingly. IDW cannot do this.
 Error Estimation: Kriging is unique because it provides a prediction surface AND an
error surface (standard deviation). This tells exactly where the map is reliable and
where more samples are needed.
 Smoothing: Depending on the "Nugget" value, Kriging can be an exact or a
smoothing interpolator.

Question: When to use IDW and Kriging

Use IDW when... Use Kriging when...

11

You have a small dataset. You have a large dataset (>50 points).
Page

You need a quick result. You have time for statistical validation.
 The surface is simple/smooth. The surface has complex spatial trends.
You don't need to know the error. You need an error/uncertainty map.
Data is evenly distributed. Data is clustered or has directional trends.

TYPES OF KRIGING

1. Ordinary Kriging (OK)

This is the most widely used version of Kriging.

 The Assumption: It assumes that the mean is constant but unknown across the entire
study area.
 Mechanism: It focuses on the local variation between points rather than a broad trend.
It is essentially an "exact" interpolator that aims to minimize the variance of the
prediction error.
 Best Use: General-purpose environmental mapping where there is no obvious "tilt" or
steady increase/decrease in values across the map.

2. Universal Kriging (UK)

Also known as "Kriging with a Trend."

 The Assumption: It assumes the mean is not constant but follows a predictable
structural trend or mathematical "drift" (e.g., elevation consistently increasing as
someone move north).
 Mechanism: It models the data in two parts: a deterministic "trend" and the remaining
random "residuals."
 Best Use: Variables that have a clear geographic bias, like temperature changing with
latitude or altitude.

3. Block Kriging

While most Kriging predicts values for a single point, Block Kriging predicts values for an
area (a "block").

 Mechanism: It calculates the average value over a specified grid cell or spatial unit.
 Best Use: Resource estimation in mining or soil science. For example, a farmer
doesn't need to know the soil pH at one microscopic point; they need the average pH
for a $10\text{m} \times 10\text{m}$ area to apply fertilizer correctly.

4. Non-Linear Kriging

Standard Kriging assumes a linear relationship, but some variables don't follow this rule.

 Examples: Indicator Kriging and Disjunctive Kriging.

 Mechanism: It transforms the data into a non-linear format before interpolation.
12

Indicator Kriging, for instance, converts values into 0s and 1s (e.g., "Is salinity above
Page

the legal limit?").

  Best Use: When dealing with data that has many extreme "outliers" or when the user
need to map the probability of exceeding a threshold.

5. Stratified Kriging

 Mechanism: The study area is divided into different "strata" or zones based on a
categorical variable (like land use type or soil series). Kriging is then performed
independently within each zone.
 Best Use: In complex landscapes like the Sundarbans, where the spatial behaviour of
salinity might be completely different in a tidal creek vs. a dry agricultural field.

6. Co-Kriging

This method uses secondary data to help predict the primary variable.

 The Assumption: The primary variable (which is expensive to measure) is highly

correlated with a secondary variable (which is easy/cheap to measure).
 Mechanism: If the user has only 10 rainfall stations but 100 elevation points, Co-
Kriging uses the relationship between elevation and rain to improve the rainfall map.
 Best Use: Mapping variables with sparse sample points by leveraging a more densely
sampled, related variable.

7. Probabilistic Kriging

This is often used interchangeably with Indicator Kriging or Bayesian approaches.

 Goal: Instead of predicting the exact value of a point, it predicts the probability that a
point exceeds a certain value.
 Best Use: Risk management. For example on cyclone-induced flood risk in Ghatal
Town, probabilistic kriging could tell: "There is a 75% chance that the water level
here will exceed 2 meters."
13
Page