GEOSPATIAL ANALYTICS AND AI APPLICATIONS

USING SATELLITE DATA PRODUCTS

Shweta Mishra
Scientist -’SF’
j a i s w a l s @ s a c . i s r o . g o v. i n
Space Applications Centre (ISRO)
STTP: SSODIAI:2023
October 9, 2023
 Indian Earth Observation System : In-Orbit Satellites
RESOURCESAT- 2 & 2A, EOS-04
• Three tier imaging : 56 m / 23 m / 5.8 m
• Revisit Capability : 03 / 11 / 03 days
• C-Band SAR (3-50m resolution) / 17 to
24 days repeativity
CARTOSAT-2E & CARTOSAT-3
• 60 cm PAN & 1.5 m Multi-spectral
• 28 cm PAN & 1 m Multi-spectral

SARAL

• Ocean Altimeter

INSAT-3D & 3DR

• 6 Channel Imager – 48 images per day

• 19 Channel Sounder – Atm. Profiles
22
A geographic information system is a complete system for capturing, storing,
querying, analyzing and displaying geographically referenced data.

Geographic Knowledge
Creating
Measuring
Organizing Holistic
Analyzing
Modeling Comprehensive
Systematic
Analytic
Applying
Visual
Planning
Managing
Acting

3
 COMPONENTS OF GIS
A Geographic Information System is not only
about computers, software and data. Its
components include:
• Hardware: the computer on which GIS
operates
• Software: which provides the functions and
tools needed to store, analyze, and display
geographic information.
• Data: geographic data & related tabular data
• People: GIS users who range from specialists
to simple users
• Methods: a designed plan and business rules

4
 Spatial / GIS Data

Data that describe both the location and characteristics of geographic features on the
Earth’s surface.
Spatial Data
Data that defines the locations of spatial features, which may be
Type: Discrete and Continuous
Example: land use, Roads, City location.

Non-Spatial Data
All attributes related to spatial or even not related to spatial are called non-spatial
data.
Example: Population, rainfall data, name of city etc.

55
 Spatial Data Model
# # # #

#
#
#
#

#
#

# #
#

#
#

# Points
#

A data model is the method used to represent real world objects

#
#
#
#
#

in a computer
#

Vector Data
Vector data model uses point and their x-, y- coordinates to construct spatial features of
Lines
points ,lines and polygons (areas).
Example: ESRI Shapefile Areas

Raster Data
Raster data model uses a grids and grid cells to represent spatial variations of the feature.
Example: Satellite Image (geotiff format)

They are different in data structures: Raster data uses a simple structure with rows and
columns while vector data may include simple or composite features.

Image

66
 AN INTEGRATED VIEW
Layers are integrated using explicit location on the earth’s surface, thus geographic location is
the organizing principal.

77
 Spatial Data Model
Data Collection
Data Storage
Field data Data Manipulation
Satellite data Creation of
database Reclassification
Arial Photos Scale Change
Analog Maps Registration
Line Smoothening
Reports

Data Analysis
Modeling
Statistical Analysis
Data Display Overlaying
Buffer Analysis
Images
Graphs
Maps
Tables
ASCII Files

88
 REMOTE SENSING
Science of collecting information
about an object, area or
phenomenon from a distance i.e.
without any physical contact.

Add a Footer 9
Interactions with medium (atmospheric effect)

10
10
Electromagnetic spectrum

11
11
SPECTRAL SIGNATURE
Spectral signature of a ground surface
feature is a set of values for the
reflectance of the feature, each values
corresponding to the reflectance
averaged over a different, well defined
wavelength interval.

Each object has its own signature like

Vegetation, Soil etc.

12
 Resolutions
Renewable Energy Applications: Forecasting Solar Insolation at a given location

AI Product #6: AI for Half Hourly forecasting of Solar Insolation for point location

1. Spectral: refers to the no. of spectral bands and bandwidth of the

sensors.
Example:
For IRS-LISS-II camera 4 bands (in micro meter),
blue: 0.45 – 0.52
Green: 0.52 – 0.59
Red: 0.62 – 0.68
NIR: 0.77 – 0.86
2. Spatial: refers to the ability of sensor system to separate two closely
spaced objects.
Example: 5.8m

Ground pixel size for PAN is 5.8 m 5.8m

 Resolutions
Renewable Energy Applications: Forecasting Solar Insolation at a given location

AI Product #6: AI for Half Hourly forecasting of Solar Insolation for point location

3. Radiometric: no. of discrete levels into

which a signal may be divided. Refers gray
levels.
Example:
LISS-III data is 7 bit data. 7 bit data has 0 to
128 gray level. (gray levels = 2n, where n is
no. of bit)

4. Temporal: revisit period of the space craft for

a given point of location.
Example:5 days
AWiFS 56m LISS-III 23 m PAN 5.8 m

LISS-IV 5m Cartosat-3 25cm

1515
 • It is location reference system for
GEOGRAPHIC COORDINATE SYSTEM spatial features on the Earth’s surface.
(GCS)
• Geographic Coordinate System is
defined by longitude and latitude.
Both are angular measures.
• Using Equator as 0 degree latitude,
we can measure the latitude value of
point as 0 to 90 degree north- south
of the equator. It is used for
measuring location in N-S direction.
• Using Prime meridian as 0 degree
longitude, we can measure the
longitude value of point as 0 to 180
degree east-west of the Prime
meridian. It is used for measuring
location in E-W direction.

16
 MAP PROJECTIONS

It is process of representing location of spatial feature lies in three dimensional surface on to the
two dimensional plane.
Map projections can be grouped either by preserved properties or projection surface. On the basis of
preserve properties, projections are defined into four categories:
• Conformal Projection (keep shapes same)
• Equidistant projection (keep distances same)
• Equivalent Projection (keep area same)
• Azimuthal Projection (retains accurate directions)
Use: Population map should be in equivalent Projection.

In contrast, an equidistant projection would be better for mapping distance ranging fro missile sites
Example: Lambert Conformal Conic Projection,
Albers Equal Area Projection.

17
17
 SPATIAL ANALYSIS TECHNIQUES

GIS Data manipulation

• Spatial data manipulation
• Non-spatial data manipulation
GIS database retrieval
• What lies here?
• Where lies this category?
• Criteria based query
• Statistics
Overlay analysis: Integrate more than one layer. (Union, Identity, Intersect), Subset and masking (Clip, Erase)
Proximity analysis: Generate buffer around a feature by applying distance criteria
3D analysis: Generate 3D perspective view using elevation data.
Network analysis: Find out shortest path using distance criteria, Find out optimum path using time criteria

18
18
 GIS

Desktop GIS Web GIS Mobile GIS

19
19
• Desktop GIS Software: Mapping software that is installed onto and runs
on a personal computer and allows users to display, query, update, and
analyse spatial data and the information linked to those locations
Example: QGIS, ArcGIS, IGIS, Envision Autodesk etc.

• WebGIS Software: Web GIS is a Geographic Information System

distributed across networked computer environment to integrate,
disseminate and communicate geographic information visually over the
Internet.
Example: GeoServer, UMN Map Server, ArcGIS Server

20
20
 WEBGIS TYPICAL ARCHITECTURE

Client Middle Ware Server

Spatial request

Web Server Web GIS

Web WWW Server
Browser
Maps, HTML, Image..

GIS
software GIS
database

PostgreSQL

21
21
 WEBGIS SOFTWARE
Components Software
Backend
Map Server/ Web GeoServer, Mapserver
GIS Server
Web Server Apache Tomcat

Data Server/ Spatial PostgreSQL, OracleSpatial

Database
Programming JavaScript, Java, Python etc.
Language

Frontend
HTML, CSS, Javascript or Javascript Framework

22
22
 23
23
 AI for Renewable
Energy Applications
 HALF-HOURLY SOLAR INSOLATION
FORECASTING USING INSAT3D PRODUCT
AT BOPAL LOCATION

• Pyranometer, Temperature and Humidity

sensors installed at SAC Bopal Campus.
Data from these are fetched continuously
Solar Irradiance study setup: Instruments
and visualisation is available on VEDAS.

• AI based Model is developed for

forecasting next 48 hours insolation at 30
minute interval for this location.
 Renewable
Half-hourly Solar Energy Applications:
Insolation Forecasting Forecasting
using Solar Insolation atproduct
INSAT3D a given location
at Bopal location

AI Product #6: AI for Half Hourly forecasting of Solar Insolation for point location
• Input: INSAT-3D (half hourly solar insolation
product of previous 7 days ).
• Output: Next 48 hours forecast (half hourly,
three hourly and sixe hourly frequency)
• Model-type: Sequential Model
• Model-details: LSTM (Long-Short Term Memory
Model)
• Performance: RMSE: 98.44 W/m2, MAE: 43.51
W/m2 (over 48 hours forecast for half hourly
forecast)
• Future Scope: Model to be applied and tested
over NIWE SRRA stations.
• Uses: Solar power plants and power supply
companies running on solar power requires
insolation prediction in advance to calculate
future power production to manage energy
requirements.
 Half-hourly Solar Insolation Forecasting using INSAT3D product for Gujarat

AI Product #7: AI for forecasting of Solar Insolation (Gujarat) using INSAT-3D GHI

• Input: INSAT-3D (half hourly

product of previous 3 days).
• Output: Next 6 hours forecast (half
hourly frequency)
• Model-type: Sequential Model
• Model-details: ConvLSTM
• Future Scope: Model will be
applied over India.
• Uses: Solar power plants and
power supply companies running
on solar power requires insolation
prediction in advance to calculate
future power production to
manage energy requirements.

28
 AI for extracting Solar plants using LISS-IV Data
Renewable Energy Applications: Mapping Solar Panels across India
AI Product #9: AI for extracting Solar plants using Satellite Data

Use Case Objective:

Automatic Extracting Solar Power Plants

from very high resolution satellite imagery

Methodology:

Developed deep learning VGG-Unet with pre-

trained weights of VGG16 network.

29
AI for extracting Solar Rooftop Panels using High resolution Satellite Imagery
Renewable Energy Applications: Mapping Solar Rooftop Panels in cities
AI Product #8: AI for extracting rooftop solar panels

Results:

Use Case Objective:

Automatic Extracting building rooftop panels

from very high resolution satellite imagery

Methodology:

Developed deep learning VGG-Unet with pre-

trained weights of VGG16 network.
AI for snow cover
monitoring
 AI for Snow cover monitoring for Himalayan Regions
Snow and Glacier Applications: Snow Cover Monitoring for Himalayan Regions

Use Case Objective:

Automatic Extraction of Snow areas/ snow cover
monitoring using IRS AWiFS satellite data

Methodology:

Convolutional neural network (CNN)

is used for extracting snow cover daily for Himalayan
regions.

32
HANDS ON
PYTHON

33
Introduction

 Strong community support

 Object-oriented programming
 Large standard library with many add-ons
 Python Scientific Add-ons:
 GDAL – geospatial raster support
 numpy: numpy.scipy.org - Numerical Python
 scipy: www.scipy.org - Scientific Python
 matplotlib: www.matplotlib.org - graphics library
Development Environment / Code Editors

 Anaconda-Spyder (IDE)
 IDLE
 Eclipse
 PyCharm (IDE) by JetBrains
 Sublime Text
 Atom
 Thonny (IDE) – built in Python 3, good for beginners
 Visual Studio code –Required to add extension to create
python development environment
 Spyder Variable Explorer/
File / Object Explorer

Editor

Ipython Console
Standard Data Types

 Python has five standard data types :-

 Numbers
 String
 List
 Tuple
 Dictionary
Numpy (Numeric/Numerical Python)

• Numpy is an open-source add-on module

that provide common mathematical and
numerical routines as pre-compiled fast
functions.
• It provides basic routines for
manipulating large arrays and matrices
of numeric data.
• The NumPy (Numeric Python) package
provides basic routines for manipulating
large arrays and matrices of numeric
data.
• Import numpy as np
PANDAS - Python Data Analysis Library

• pandas is an open source, BSD-licensed

library.
• It provides easy-to-use data structures
and data analysis tools for the Python.
• Its key data structure is called the
DataFrame.
• DataFrames allow you to store and
manipulate tabular data in rows of
observations and columns of variables.
• pandas allows you to work with CSV, Excel,
JSON, and SQL database files, pull them
into dataframes
• Syntax:
dataframe = pd.read_csv(‘file_name.csv')
GDAL
• GDAL- Geospatial Data
Extraction Library -
accessible through C,
C++, and Python.
• Reads, writes and
manipulate rasters.
• Converts image, in
memory, into a format
Numpy arrays
• Propagates projection
and transformation
information
• Handles No Data
 GDAL Command Line Utilities

• gdalinfo - report information about a file.

• gdal_translate - Copy a raster file, with control of
output format.
• gdalwarp - Warp an image into a new coordinate
system.
• gdal_merge.py - Build a quick mosaic from a set of
images.
• gdal_rasterize - Rasterize vectors into raster file.
• gdal_transform - Transform coordinates.
• gdal_polygonize.py - Generate polygons from
raster.
• gdal_fillnodata.py - Interpolate in nodata regions.
OGR
• OGR: OpenGIS Simple Features Reference: Tools for
manipulating vector data
 ogrinfo - Lists information about an OGR supported data
source
 ogr2ogr - Converts simple features data between file
formats
• It perform various operations during the process such as
spatial or attribute selections, reducing the set of
attributes, setting the output coordinate system or even re-
projecting the features during translation.
 GDAL and NumPy

• Process:
Get raster band(s)

Convert the raster

band(s) to a NumPy
Open GDALDataset
array using
ReadAsArray()

Write out GDAL Dataset Process the raster

Convert the NumPy
array to GDAL raster band(s) using NumPy
bands using functionality
WriteAsArray()

Taken from: geography.middlebury.edu/data/python/geoprogramming.ppt

 THANK YOU
j a i s w a l s @ s a c . i s r o . g o v. i n
h t t p s : / / v e d a s . s a c . g o v. i n
• Sequential Modeling: These models input or output sequences of
data. Sequential data includes text streams, audio clips, video clips,
time-series data and etc. Recurrent Neural Networks (RNNs), LSTM is
a popular algorithm used in sequence models.
Applications of Sequence Models: (I) Speech recognition (II) Time Series
Forecasting (iii) Sentimental Analysis
• Computer Vision: Computer vision is a field of artificial intelligence
(AI) that enables computers and systems to derive meaningful
information from digital images, videos and other visual inputs.
• Object or Facial Detection and Recognition
• Medical Imaging: It assists doctors in diagnosis of dangerous condition
• Solar Power Plant Detection

47
 Neural Network
• Neural networks are algorithms
explicitly designed as an inspiration
for biological neural networks. Initially,
the goal was to create an artificial
system that functioned similarly to the
human brain. Neurons and layers are
the primary components of neural
networks.
• Each neuron is given a number of
inputs, weighted sum is performed,
activation function is applied and
output is given. The network has a
loss function which is used to
minimize the error in weights.
https://towardsdatascience.com/machine-learning-for-beginners-an-introduction-to-neural-
networks 48
Recurrent Neural Networks
• Time series data is a collection of observations obtained through repeated
measurements over time.
• A recurrent neural network (RNN) is a special type of an artificial neural network adapted to work
for data that involves sequences. RNNs have the capability to store the states or information of
previous inputs to generate the next output of the sequence. Recurrent neural networks, also
known as RNNs, are a class of neural networks that allow previous outputs to be used as inputs.
In theory RNNs can make use of information in arbitrarily long sequences, but
in practice they are limited to looking back only a few steps (Vanishing gradient
problem).

https://colah.github.io/posts/2015-08-Understanding-LSTMs/ 49
 UNET
• UNET is an extension of an encoder-
decoder fully convolutional network. The
convolutional layers perform feature
extraction by employing the filters in these
layers to learn low and high dimensional
features as they iteratively get trained.
• The intuition behind UNET is to encode the
image passing it through a CNN as it gets
downsampled and then decode it back or
upsample it to obtain the segmentation
mask.
• The features to be detected in the mask
depends on the learned weight filters,
upsampling & downsampling blocks (which
can also be made learnable) and the
concatenations & skip connections.

https://towardsdatascience.com/understanding-semantic-segmentation-with-unet-6be4f42d4b47
50