Rajiv Gandhi University Of Knowledge Technologies

APIIIT SRIKAKULAM
“Final Review of Mini Project”
for E-3 2016
Land Classification Using Convolutional Neural Networks
Submitted as part of Mini Project.

<N.V.S.DINESH> <S160649>
<M.R.SUDHAKAR REDDY> <S160347>
<T. KIRAN SIVA SAI> <S160078>
<T. SANDHYA ANITHA> <S160194>

Under the Supervision of:

Mr. SIVA RAMA SASTRY GUMMA
Assistant Professor
Department of Computer Science and Engineering
Rajiv Gandhi University of Knowledge Technologies
Srikakulam – 532402
Email:- srsastry09@rguktsklm.ac.in 1
1
Problem Statement:
Can we create a model that can classify the landscape/land use of a given image???

Abstract:
• Satellite imagery can help us find solutions to the growing number of environmental problems that we face today. It allows us to not
only get a bird’s eye view of what’s around us, but also uncovers parts of the world that are rarely seen. Tapping into the potential of
categorizing land cover and land use around the world means that humans can more efficiently make use of natural resources, hopefully
lowering cases of waste and deprivation. But despite its potential to be incredibly useful, satellite data is massive and confusing, and
making sense of it requires complex analysis.
• This project could help a variety of stakeholders, including conservationists, urban planners, and environmental scientists, survey and
identify patterns in land use to see which natural areas are under threat or which areas are best for urban development. But in this case,
I will tailor this project to a client that is similar to the agriculture imagery company that I currently work for. Imagery companies can use
land use classification models to categorize what’s in each image and optimize their efforts towards the parts of land that are important
to them. For example, an agriculture company will only want to be concerned with land that’s labelled as pasture, annual crop, or
permanent crop. If a satellite is constantly taking images, this project would help save hours of time manually sorting through imagery.
• In addition to sorting imagery, land use classification is important for identifying the parts of an image to which certain analyses are
applied. For example, if the company has different crop stress algorithms for fields that are next to urban areas versus fields that are
next to rivers, this project would help them to automate the process of applying these specialized algorithms and more effectively solve
environmental problems.

2
Objective:
Create a land classification model using a convolutional neural network
Tools:
Pytorch, NumPy, Matplotlib
Algorithm:
a. Data preparation
b. Exploratory Data Analysis
c. Deep Learning Modeling

3
Data Collection:
Methods

● Download data locally

● Translate files from TIF to JPEG format using gdal_translate
● Only extract red, green, and blue bands
● Separate data into training, validation and test sets

4
Satellite Data:

● EuroSAT Sentinel-2 images from the German Research Center for Artificial Intelligence
● Ten different land use classifications
● 13 bands
○ Only RGB used in this project

5
6
7
Exploratory Data Analysis:
Methods
• Use histograms and bar plots to visualize data
○ Differences between categories
○ Differences between color bands

8
9
Basic Convolutional Neural Network:
Methods & Results
● One convolutional input layer, one flatten layer in the middle, and one dense
output layer
● Good to test data preparation and parameters
● Not good for model performance
● 10% accuracy == random guess

10
11
Transfer Learning:
Methods & Results

● Adapt a pre-trained model to this project

● VGG16 model, trained on
● 15 million images in over 22k categories
● Freeze first 12 layers, train last 10
● 94% accuracy > random guessing

12
13
14
 Convolutional layer
early in the model

Pooling layer early in

the model
Convolutional layer
late in the model
Model Tuning:
Method & Results
● Data augmentation
○ Computational challenge to the CNN

○ Increases dataset size

● Batch normalization
○ Improves speed and performance of the neural network
● 56% accuracy
○ Could increase with more epochs and higher learning rate
17
Conclusions & Next Steps:
Conclusion
1.Save on employee costs
Money spent manually categorizing satellite imagery can be saved by
having the model handle it.

2. Use employee talents for projects that add value to the company
a. Research and development
b. Growth strategy
c. Human-centered design

20
Future Explorations:
1. Improve model tuning techniques
– Weight Initialization
– Dropout regularization

2. Explore other base models for transfer learning.

Mobile Net
3.Use satellite data from other sources to corroborate or improve the model

The End

21