02-02-2026

Geospatial Modelling
For Earth System Science

Arijit Roy
Head Disaster Management Studies department
Indian Institute of remote sensing

Models

1
 02-02-2026

Models and Modelling

What is a model ( Definition) ?

Models are considered as abstractions and

Approximations of reality which is achieved through
the process of Simplification of complex real world
relations to the point that they are understandable
and analytically manageable.

Modeling involves making a representation of something.

Creating a tiny, functioning volcano is an example of modeling.

... Modeling is anything that represents something else, usually

on a smaller scale. Military commanders use physical and
computer modeling to plan offensives.

TYPES OF MODELS

2
 02-02-2026

COMPONENTS OF GEO-SPATIAL MODELS

Spatial data models

3
 02-02-2026

Modelling and Simulation

Modelling is used to explore scenarios and make Simulation is used to predict the behaviour of
prediction a system in certain sets of conditions

Simulation recreates a scenario of a

system using a physical framework

Map Projections: Basic Geospatial Model

Map projection is a
way to flatten a
globe's surface into a
plane in order to make
a map. This requires a
systematic
transformation of the
latitudes and
longitudes of locations
from the surface of
the globe into
locations on a plane.

4
 02-02-2026

Map Projections: Types

Cylindrical Projections: These projections involve wrapping a cylinder around the

Earth and projecting its features onto the cylindrical surface. Examples are the
Mercator, Transverse Mercator, and Miller Cylindrical projections.
Conic Projections: For these projections, a cone is placed over the Earth, and its
features are projected onto the conical surface. Common examples are the
Lambert Conformal Conic and Albers Equal-Area Conic projections.
Azimuthal Projections: Also referred to as planar or zenithal projections, these use
a flat plane that touches the Earth at a single point, projecting the Earth’s features
onto the plane. Azimuthal Equidistant, Stereographic, and Orthographic
projections are examples.
Pseudocylindrical Projections: These projections resemble cylindrical projections
but employ curved lines instead of straight lines for meridians and parallels. The
Sinusoidal, Mollweide, and Goode Homolosine projections are popular examples.

Classification of spatial models by technique

• Dynamic Optimization models (mathematical For changes based on

programming) neighbourhood characteristics

• Cellular Automaton models For changes based on

neighbourhood characteristics

For changes based on trend

• Statistical/Regression models

• Agent based/Multi-agent system models For changes due to impact of

external factor

• Integrated/Hybrid models Models using a combination of 2

or more techniques

5
 02-02-2026

Dynamic Optimization Models

• It derives an ideal or optimal solution for a given system, based on a quantitative
objective.
• Incorporate temporal lags and/or forward looking behavior.
• Models are most often normative and used for policy.
• Spatial dynamic optimization models are often difficult to solve.

Dynamic Optimization Models

A Case Study of Ecosystem Development in Blast Furnace Slag Dumps

Conceptual model of a developing slag dump

ecosystem. Legend: Box represents the state of a
component; valve, rate of change of state; circles,
parts of rate description qualifying for a separate
entity; connector (line), amount and direction of
flow between component; clouds, source or sink
for the flows. Numbers correspond to equations
described

6
 02-02-2026

Dynamic Optimization Models

Cellular Automata Models

 CA models are dynamic simulation models, where • Strengths:

cell transitions are based on the state of the current • Models are very strong at representing local spatial
cell and the states of neighboring cells. patterns
 “Neighbors” can be very broadly defined, and may • Models tend to do well at replicating real-world
include multi-scale influences. spatial patterns, especially fractal structures
 Cellular structures are generally grids, but can be
any cellular structure, in principle. • Weaknesses:
• Models may place too much emphasis on local
interactions
• Models are not strong at representing behavior
• Often, models require projections of rates and
quantities of change to run
• Difficult to calibrate; does not necessarily explicitly
incorporates drivers.

Statistical/Regression Models
Markovian Chain
• These models find a set of best-fit model
coefficients that express a statistical relationship A Markov chain is a stochastic model describing
between a dependent variable (often land use or a sequence of possible events in which the
cover) and a series of independent variables probability of each event depends only on the state
(representing drivers of LUCC) attained in the previous event.

• Models produce a transition probability, Regression based Models

conditional on states of independent variables Regression analysis aims to study the relationship
• Models are only dynamic when some set of rules between one variable, usually called the
is used to generate transitioned landscapes using dependent variable, and several other variables,
those estimated probabilities often called the independent variables. These
models are among the most popular data-driven
models .

Logistic Regression
In statistics, logistic regression (sometimes called
the logistic model or logit model) is used for
prediction of the probability of occurrence of an
event by fitting data to a logit function logistic
curve.

7
 02-02-2026

Commonly Used Spatial Models

• Spatial Models: (Patterns, Fractrals, Relationship)

• Landscape models (fragmentation, Biodiversity richness)
• Species distribution models (Entropy, Genetic Algorithm)
• Complex Network models (least cost, Graph Theory)
• Dynamic Global Vegetation Models (DGVM)
• Species loss and shift modeling (Hybrid/ Integrated)
• Fire-risk and Spread modelling (Integrated Models)

Spatial Models: Parameter Retrieval

Oak nitrogen mass map Sal nitrogen mass map

8
 02-02-2026

Spatial Models: Fractals

A scale invariant power relationship between perimeter and area ratio

P = kAD/2 (1)

Where:
 P is the perimeter of the study area
 A is the area of the study area
 D is the fractal dimension
 k is the constant of proportionality

Fractal dimension, D, may then be determined from the equation 1 as

D = 2*(ln(P)-ln(k))/ln(A) (2)

Spatial Biodiversity Characterization

Landscape comprises the visible features of an area
of land, including the physical elements of landforms
and human elements

 Different ecosystems make a landscape

 Landscape ecology deals with the spatial distribution
of the different ecosystems, their dynamism over time
and space and their interaction

Shaping of a landscape is dynamic landscape

influenced by mosaics; resilience and
thresholds; biocomplexity;
adaptive cycles; Spatial distribution of ecological variability:
Attributes (proxies) of Biological Richness (Aspect: Elevation; temperature; slope)
Climate Terrain Complexity and Geomorphology Source: Google earth
 Moisture availability
 Altitude Studies have reported that if Altitude, slope, aspect do not vary much in a
 Slope landscape then landscape variations like soil moisture account for more that 70%
the vascular plant diversity
 Aspect
 landscape People are part of biodiversity, but people exert forces on biodiversity, which destroy the natural
landscape……………..Hence characterization and monitoring

9
 02-02-2026

Approach for Spatial Biodiversity Characterization

Parallel computing for spatial data processing

Database Preparation

 Two Season IRS LISS – III used Roy et al., 2012

 Hybrid Classification approaches followed

 16518 field plot data of 7606 species database integrated
 120 Vegetation types/habitats mapped

10
 02-02-2026

Fragmentation and Disturbance Index Maps

Biological Richness Maps

Sensitivity Analysis

Model Preformance Analysis

11
 02-02-2026

Biological Richness Map and its use for conservation/prioritization

CA-Markov LULC Change Model

Suitable for modeling spatial and temporal land use change.

 The Markhov process controls temporal dynamics among the cover types through the use of transition probabilities.
 Spatial dynamics are controlled by local rules through a CA mechanism considering either neighborhood
configuration or transition probability.
 GIS and remote sensing data can be used to define initial conditions, to parameterize the CA-M model, to calculate
transition probabilities and determine the neighborhood rule.

12
 02-02-2026

CA-Markov LULC Change Model: Model Caliberation

CA-Markov LULC Change Model: Future LULC Prediction Map

13
 02-02-2026

Development of Agent Based Change Model

Model requirement 1. Need for a model with multiclass predictability

2. Need for a driver based model
for Indian 3. Driver data input should not be a (0, 1) input as in Clue-S
Conditions 4. Drivers should dictate the land conversion decision rules as in India
unlike in Europe

Development of Agent Based Change Model

MODEL STRUCTURE Data Used

SPATIAL PATTERN OF LOGIT

Temporal LULC Maps
LAND USE CLASSES REGRESSION (1985, 1995, 2005)
& DRIVING FORCES

Driver Data:
 Population Density
WEIGHTAGES OF (village level)
DRIVING FORCES
 Climate (Rainfall;
Temperature)
 Topography (elevation,
slope)
CHANGE IN PROBABILITY SURFACE
FOR ALL LAND USE
ALLOCATION OF ACTUAL  Infrastructure (roads,
DRIVING FORCES CHANGES LAND USE
TYPES (T-2) (T-1)
railways)

-LOGIT
CONDITIONAL
DATABASE CORRELATION REGRESSION IMAGE RS/GIS DATABASE
ORGANIZATION -TRANSITIONAL COMPOSITION
PROBABILITY

14
 02-02-2026

Development of Agent Based Change Model

BL\W EG\SE
L CL FL WB DF BU G MF

BL\WL 0.8554 0.051 0.0224 0.0173 0.0008 0.0272 0.0001 0.0257

CL 0.0086 0.893 0.005 0.0387 0.0033 0.0125 0.0001 0.0388

FL 0.0483 0.1016 0.6372 0.0257 0.03 0.035 0.0048 0.1173

WB 0.0028 0.0036 0.0006 0.9711 0.0011 0.0012 0.0001 0.0195

DF 0.002 0.0843 0.0011 0.0054 0.7046 0.0017 0.0762 0.1247

BU 0.0082 0.0955 0.0462 0.0005 0.0001 0.8417 0.0029 0.0049

EG\SE
G 0.0000 0.0012 0.0000 0.0001 0.0766 0.0001 0.8911 0.0309

MF 0.0041 0.2633 0.0088 0.0064 0.0608 0.0146 0.0293 0.6126

Development of Agent Based Change Model

Kappa= (d-q) / (N-q)

d= 341, q= 50, N= 400,
Kappa= (341-50) / (400-50)= 291/350, = 0.8314.
Where: -
q= No. of cases expected in diagonal cells by chance.
q= n (row) * n (column) / N
d= Diagonal total of cells,
N= Total of column / rows which should be equal

Classification Accuracy

= d / N = 341/400= 0.8525.

15
 02-02-2026

Land Use Land Change dynamics in Western Himalaya

Decadal LULC
1975 - 2015

LULC classes
WB
CL
BL
SL
GL
LULC Class 1975 (Km2) 2015 (Km2) EF

Scrubland (SL) 6772 6083 DF

BU
Cropland (CL) 18870 21206 PL
SI
Built up (BU) 715 951
Barren land (BL) 7333 10381
Snow/Ice (SI) 21461 16975
Grassland (GL) 8315 9608
Deciduous forest (DF) 7380 7164
Evergreen forest (EF) 31038 29569 Predicted map of
Plantation (PL) 2454 2533 2055 using Land
Waterbodies (WB) 2448 2317 Change Modeller

Species Distribution Modelling

Data Used 1. Annual mean temperature
2. Mean diurnal range
3. Isothermality
4. Temperature seasonality
5. Maximum temp. of warmest month
6. minimum temp. of coldest month
7. Annual temperature range
8. Mean temperature of wettest quarter
9. Mean temperature of driest quarter
[Link] temperature of warmest
quarter
[Link] temperature of coldest quarter
[Link] precipitation
[Link] of wettest month
[Link] of driest month
[Link] seasonality
[Link] of wettest quarter
[Link] of driest quarter
[Link] of warmest quarter
[Link] of the coldest quarter

16
 02-02-2026

Species Distribution Modelling

Results/Outputs Model performance and validation

Std. dev. (blue) as a result of

subsample is a measure of The sensitivity of
how the model’s predictability model is being tested
changes over the different with independent field
suitability ranges.
data

Species Loss modeling

Bayesian Method
jackknife estimates
Time taken for a fragment to
Species Area Relationships
lose 50% of the species

Ferraz et al. 2003, PNAS, 100(24), 14069–14073

% of Niche area covered

Collins et al, 2002, Evolutionary Ecology % of invasive species in the niche
Research, 4: 457–464

Schwartz et al, 2000, Oikos, 122, 297–305

CLIMATE ENVELOP
MODELS Extent of
Connectivity
Lewis, 2006 Phil. Trans. R. Soc. B
RATE AND EXTENT OF 361, 163–171
SPECIES LOSS

17
 02-02-2026

Distribution of Rhododendron arboreum in Himachal Pradesh for present and

IPCC 2050 & 2080 scenario

A2a 2050 A2a 2080

Present

B2a 2050 B2a 2080

Number of dominant species in temperate broadleaf forest of Himachal Pradesh

A2a 2050 A2a 2080

Present

B2a 2050 B2a 2080

18
 02-02-2026

Species Association Dynamics under Climate Change Scenarios

Methods
Species
distribution maps
from ensemble
modelling

Conversion into
Binary maps

Algorithms

Species
association maps
for baseline and
Present (baseline) scenario
climate change
scenarios

Modelling Wildlife Corridors: Elephants

Rameshan et al. 2014, Intl J of Geoinformatics

19
 02-02-2026

Modelling Wildlife Corridors: Tigers

Prey/Copredator-
Predator Game
Habitat theoretic Model
Suitability

Habitat
Connectivity

Predicting the probable links and nodes important for species migration in
climate change scenario
Future LULC maps

2012

IPCC 2050 b2

Network
analysis

Potential dispersal of Future potential species distribution

Arundinaria maling w.r.t.
degree of Probability of
Connectivity in Darjeeling

20
 02-02-2026

Earth system Models

Earth System Modelling

Interaction of subsystems of earth gives rise to habitable environment

Atmospheric circulation and radiation

Climate Model Sea Ice

Land physics
Ocean circulation and hydrology

Atmospheric circulation and radiation

Allows Interactive CO2

Earth System Model

Sea Ice Plant ecology, land use,

Ocean ecology and and Biogeochemistry
chemistry

Land physics
Ocean circulation and hydrology

21
 02-02-2026

Earth System Modelling:

• Interaction of subsystems of earth gives rise to habitable environment.

• Over the time earth system model developed from simple to complex i.e. from General Circulation
Model (GCMs) to Dynamic Global Vegetation Model (DGVMs) to represent changes in terrestrial
habitats.
• Main application of DGVMs is to capture and simulate the transient changes in the vegetation
cover(Foley et al. 2000 ).
• Dynamic Global Vegetation Model are process-based numerical models which simulate the large-scale
dynamics of the terrestrial ecosystem.
• Several DGVMs have been developed by different research groups like HYBRID & SDGVM (U.K.) & IBIS
(SAGE, USA) and LPJ-GUESS (LU, Germany).
• DGVMs consider Plant Functional Types or traits for simulation of the existing vegetation.

RCM DGV
GCM M

Climate Models

 Climate models divide the world into

3D blocks called grid cells.
 Each cell has information about
whether it is covered by water or land
and the land use – for example, farm,
mountain or city.
 Each cell also uses physical and
chemical laws to simulate the
movement of air, water and energy.
 The mathematical equations are
turned into computer code.
Additional coding allows the different
grid cells to interact.
 To complicate things even further,
climate happens over a period of
years, so the models run on hourly,
daily or weekly timescales.

22
 02-02-2026

Dynamic Global Vegetation Model (DGVM)

 Computer program that simulates shifts in potential
vegetation and its associated biogeochemical and
hydrological cycles as a response to shifts in climate

 Simulate monthly or daily dynamics of ecosystem

processes.

 Used most often to simulate the effects of future climate

change on natural vegetation and its carbon and water
cycles.

DGVM integrated four types of processes

Broadleaf Tree

Plant Geography Koppen, Box, Shrub

MAPSS C3 Grass
Soil

Biogeochemistry Miami, TEM,

Century D Competition (10 days)
G Plant functional type (PFT)
Biophysics SiB, BATS,LSM V
Deciduous, evergreen trees
M Shrub
Grass
Foret, FORSKA
Vegetation Dynamics Crop

Dynamic Global Vegetation Model

Dynamic Global Vegetation Model (DGVM) is a computer program that simulates shifts in potential vegetation and its
associated biogeochemical and hydrological cycles as a response to shifts in climate.

23
 02-02-2026

Predicting Plant Functional Types (PFTs) in response to Climate Change in

Western Himalaya Using Dynamic Global Vegetation Model: LPJ – GUESS
Changes in Plant Functional Types Grid-wise change in Dominant PFT
(PFTs) are critical in evaluating the TeNE (Temp. Needle Leaf Evergreen
climate change impacts and can be
simulated in DGVMs which are
process based models

Inputs

• Annual CO2 Concentration – Mauna

Loa
• Precipitation, Mean temperature,
Cloud Cover, Minimum and
Maximum Temperature, Wetdays,
Sunshine – CRU
• TRY Data for Model Parameterization
• Soil Texture Data – FAO
• Annual Nitrogen Concentration – Conclusion
Global • PFTs shift their potential areas to higher altitudes.
• Some PFTs (TeNE) is not able to keep pace with the climate change and hence will decrease
in future which is also seen in IPCC AR 5 (2014).
• Anthropogenic disturbance should be taken into account in the Model.

Early Warning/ Forest fire risk

• DSS which involves factors influencing fire danger

• Indexing into different classes of fire danger viz. No
danger, Low, Moderate, High and very High for rapid
characterization.
• Implementing the mitigation measures for controlling
fires.
• Two biophysical components:
– DYNAMIC INDEX: based on weather parameters such as air
temperature, relative humidity, wind speed and rainfall;
– STATIC INDEX: based on the constant parameters such as
fuel characteristics, topographic conditions, vegetation
type, edaphic conditions.
– This is then integrated to provide the fire danger rating

Accuracy of 86% to 92%.

With respect to actual fire occurrence

Presently the model is calibrated for Western Himalayan region of India

24
 02-02-2026

Fire Danger Maps Overlaid With Fire Points

May21, 2018
May22, 2018 Accuracy was computed for each day,
assuming that fires fell in No fire, Low
and moderate fire regions were as un-
identified fire pixels by the forest fire
danger index. Accuracy was estimated
on the basis of number of fires fell in
different fire danger classes.
S. No. Date No. of fire incidents Accuracy
May22, 2018
No, Low, High, Very (%)
Moderate High
1 21-05-18 221 1096 83.21
2 22-05-18 113 735 86.67
3 23-05-18 81 478 85.50
4 24-05-18 89 456 83.66
5 25-05-18 68 726 91.43
6 26-05-18 218 976 81.74
7 27-05-18 85 768 90.03
8 28-05-18 18 349 95.09
9 29-05-18 12 116 90.62

Automation of Fire Danger Index

Since this product is dynamic and need to be run daily hence the entire system has been automated and has been
calibrated for 2 states of Western Himalaya (HP, UK)

MODIS NRT
datasets

MOD09GA
(Surface MOD11_L2
ASTER GDEM
reflectance (LST)

Pre processing

PST
Pre processing

Dynamic Fire Static Fire

Danger Index Danger Index
MNDFI

Band 2
Band 5
Band 7 PMI
Forest Fire Danger Index

 Daily product 30 minutes after the MODIS (TERRA) pass is provided High Risk Areas

 Beta version experimental products disseminated [Link]

25
 02-02-2026

Fire Spread Model

Hybris Model using a combination of WRF, Complex

Network and Cellular Automata for predicting the
movement of fire across the landscape

THANKS

26