Classification of urban elements according to CO2 emission using Machine learning

Urban elements and their particular identities define and shape cities. The field of

Urban Design has been discussing this term of urban element over several years

and provided several explanations. The term Urban elements consist of two components

urban, and element both are often examined as crucial factors in urban

design and planning processes in cities. The component ´´element´´ refers to the

tool or an instrument forming an identity of the placement. Elements placed in the

urban scene include a feature and can be represented by a plaza, street, bench,

park, building, and neighborhood. However, when such urban elements are joined

and interconnected, they develop a system of urban elements in the cities. According

to Richthofen ´´Urban elements are capable of accelerating the process of

urbanization in a city´´(Richthofen, 2017).

Currently, we are starting to see some concepts for using machine learning processes

in the field of urban design to face urgent global problems such as the

influence of CO2 emissions in the cities. Even though cities occupy less than 2 % of

the earth’s surface, they produce more than 60% of all greenhouse gas emissions

and consume 78% of the produced energy. [1] This project’s overall intention is to

spread awareness about the influence of urban elements on CO2 emissions in the

cities to effectively reduce the carbon emissions in the planning processes.

According to oxford languages, the definition of machine learning is “the use and

development of computer systems that can learn and adapt without following

explicit instructions, by using algorithms and statistical models to analyze and draw

inferences from patterns in data.” [2] The process of machine learning was presented

several years ago, and it has been developed over the years with the contribution

of many researchers. Machine learning’s history starts already in the 1940s,

but it wasn’t until recently that it started to develop rapidly. The technical revolution,

the development of new algorithms, and the availability of tech are some of

the reasons for this rapid development. Machine learning can be considered as a

combination of different inventions and algorithms. (GLADCHUK., V. (2020))

This research examines the machine learning process in the context of training

and recognizing the urban elements and their influence on carbon emission.

The urban scenes enable us to understand the city’s image where urban elements

complement and form the scenes. The objective is to analyze the urban

elements in the urban scene and teach the machine to recognize such elements

according to their carbon emissions influence. This project aims to help

designers and planners acknowledge urban elements’ influence on the CO2

emissions within a city more smartly and effectively.

Method

The purpose of this project is to train a machine learning model to recognize

and classify urban elements according to their carbon emission in an urban

scene by using deep learning-based classification and GAN (generative adversarial

network). Classification is a type of supervised machine learning [3],

where the algorithm’s training is based on labeled data.It learns to recognise

elements according to their specific categories.[4] The classification algorithm

uses a convolutional neural network architecture.

Our project used a collection of several data sets in order to create a dataset

of diverse pixel-based urban scene images with their corresponding classifications.

Some of the images in the dataset were photographed and classified

manually by us. The substantial part of the dataset was mainly collected from

online data libraries. The libraries we used are Camvid library[5] and Cityscape

dataset(Cordets e al.2016), both libraries are pixel-based libraries that focus

on the semantic segmentation of urban scenery and associate pixels with their

corresponding labels. Camvid has 32 labels and Cityscape has 30 labels.

To use the collected data set for our purposes, we relabeled the collected

datasets according to the influence on CO2 emission categories (Carbon source

+, carbon source, carbon neutral, carbon sink, and people). This process was

utilized by identifying the RGBS of the labels in the original labeled dataset and

changing them using a Python code to the RGBs corresponding with our 5 categories

for labeling. The dataset was divided into 90% training and 10 % testing

material. The final dataset was 964 images of 512 * 512 pixels with their corresponding

labeled images.

We used two open-source algorithms (Codes) from GitHub repository [6] for

the training process. We used the same dataset for the training of both algorithms

to recognize and classify the urban elements according to our 5 classifications

in images of urban scenery.

The first was Semantic Segmentation Suite in Tensorflow [7] and the second

was Pix2Pix Gan[8]. Semantic segmentation is understood as the process of

labeling items in an image according to their class. It works by labeling each

pixel in an image with a corresponding class. The prediction then happens for

each pixel in an image. This type of prediction is called dense prediction.[9] The

Semantic Segmentation code from GitHub allows you to choose a model from

16 different models to train the data set with. We trained two of those models (DensNet56 and Mobile UNet) with our dataset.

Each of those models was trained with about 600 Epochs and a checkpoint

was set for every 10 epochs. The basis of the Pix2Pix GAN is the generative

adversarial network, which is an approach for an image-to-image translation. It

works by generating an image based on the input image it gets.[10] The pix2pix

GAN is a conditional GAN because the generating of an image is conditional

to an input image. [11] The Github Code allows you to train the model in both

directions A2B ( Original image to labeled image) and B2A ( Labeled image to

original image). We trained our model in both directions because we saw an

opportunity in training a model to create urban scenarios from color-coded

images. This model was trained with 400 epochs.

After the training process was complete, we applied the trained model to

random urban images. The process and outcomes can be seen in Figure (). The

process can be divided into four steps for both methods. First is the preparation

of a database for the training process. Then, the adjustment of the labels

to correspond to our goals. Then the training of the models and testing them

with several delineated images. Both tools provide different outcomes enabling

us to compare them with our desired results. The classified results from the

trained pix2pix reflected our goal.

SOURCES:

(1) RICHTHOFEN., A. (2018). Element, System and Milieu. Urban elements ( for this citation we need to check which part of the book we

used because there are sub authors ) Stephen Cairns

(2) GLADCHUK., V. (2020). The History of Machine Learning: How did it all start? https://labelyourdata.com/articles/history-of-machinelearning-

how-did-it-all-start/

( I don’t know how to add this is ecaade citation +

(3) Oxford Languages and Google - English | Oxford Languages (oup.com)

(4) Cities and Pollution | United Nations “Cities and Pollution.” Accessed January 7, 2021. https://www.un.org/en/climatechange/climate-

solutions/cities-pollution.

“UNITED NATIONS Climate Change - Summit 2019.” United Nations, United Nations, www.un.org/en/climatechange/cities-pollution.shtml.

(5) LEONEL.,J. (2019). Classification in machine learning. https://medium.com/@jorgesleonel/classification-methods-in-machine-learning-

58ce63173db8

(6) What Is Machine Learning: Definition, Types, Applications and Examples | Potentia Analytics Inc. (potentiaco.com)

(7) LEPELAARS, C. “CamVid (Cambridge-Driving Labeled Video Database).” Kaggle, May 5, 2020. https://www.kaggle.com/carlolepelaars/

camvid. Creative Commons — Attribution-NonCommercial-ShareAlike 4.0 International — CC BY-NC-SA 4.0

(8) CORDTS, M., OMRAN, M., RAMOS, S., REHFELD, T., ENZWEILER, M., BENENSON, R., FRANKE, U., ROTH, S., & SCHIELE, B. (2016). “The

Cityscapes Dataset for Semantic Urban Scene Understanding.” in Proc. of the IEEE Conference on Computer Vision and Pattern Recognition

(CVPR)

(9) (https://github.com/ )

(10) (GitHub - GeorgeSeif/Semantic-Segmentation-Suite: Semantic Segmentation Suite in TensorFlow. Implement, train, and test new Semantic

Segmentation models easily!)

(11) GitHub - phillipi/pix2pix: Image-to-image translation with conditional adversarial nets

(12) JORDAN., J. (2018). Semantic segmentation. https://www.jeremyjordan.me/semantic-segmentation/

(13) BROWNLEE., J. (2019). A gentle introduction to pix2pix generative adversarial network. https://machinelearningmastery.com/a-gentle-

introduction-to-pix2pix-generative-adversarial-network/

(14) How to Develop a Pix2Pix GAN for Image-to-Image Translation (machinelearningmastery.com)