Week 8— Warmth of Image

Title: Weather Condition Prediction from Image

Team Members: Berk GÜLAY, Samet KALKAN, Mert SÜRÜCÜOĞLU

E-mails Respectively: berkgulay.cs@gmail.com , abdulsametkalkan@gmail.com , mertsurucuogluu@gmail.com

Welcome to our latest blog post!

Finally, we have completed our project and obtained our latest results (at least for now). :) Before I mention our results and conclusions, I want to give a short info about our latest works from this week and previous weeks. Moreover I am also sure about that you wonder answers of those questions, What will be next?, In which platforms you can follow us? or Where can you find our latest works, datasets, papers?

Let me briefly explain all. Over the previous weeks, alongside of completing and comparing our results from different algorithms, architectures and methods, we also prepared a short introductory video presentation for our project. We gathered our conclusions together , tuned our results and arranged all to present you. We organised our codes, dataset, CNN models and extracted feature models to share as well. We wrote a final paper to explain all details which we got over.

Good news!!! We also opened our codes to you and you can follow us on Github as well. Next, we will be organizing our codes, manuals and introductory stuff on Github. Furthermore we will gather all of our blog posts together on our own Medium publication and you can find this Medium page in Github-Readme of our project later. Another good news from Twitter, “Deep Learning Turkey” page mentioned about our blog posts and did retweet our videos with our requests on their page.

DeepLearningTR mentioned our Medium blogs and projects on their Twitter page

Github (Please follow us there!) :)

GitHub - berkgulay/weather-prediction-from-image

Introductory Video Presentation :

Here our Project Paper ( with ins and outs ;) )

WarmthOfImage.pdf

Let me also briefly show our results and talk about our conclusions after all of these works;

This is our latest CNN architecture which gives best result.

Our best results. These two algorithms with shown architectures overperform.

Some of our test results with respective images,

Different test image’s classification results using CNN architecture shown above

And lastly our conclusions at the end of Warmth Of Image project…

• In this work, we used supervised learning methods which are Convolutional Neural Network(CNN), Support Vector Machine, Decision Tree and Random Forest for image classification. CNN gives the best result with %64.63 and Random forest gives close and another best accuracy with %63.90. Also RF is really time efficient and easy to perform with right feature models. Other methods(DT or SVM) are not good enough for our task and slower.
• The hardest part is trying CNN architectures, because there are so many variations and trying each architecture takes too long(high time cost). If we had strong GPUs, we could try more different architecture and find better architectures maybe.
• Random Forest works faster(really time efficient and can be used in real time applications) and gives good generalized results at the end.
• For image description right features and methodologies lead really good results and in future works other descriptors can be researched and used.
• Cloudy class is tricky and hard to distinguish from other classes because all other weather conditions also include clouds. Taking it out or change with another air type may produce way better weather classification performance.
• Even if rain detection is hard to perform task, we could succeed it for most of the cases.