Hands-on-Experience on achieving State of the Art results on classifying EuroSAT satellite images using Keras

A brief description of DATASET:

In this post, I will show you how to obtain near state of the art performance on Sentinel-2 satellite images (RGB color space) provided within the scope of the Earth observation program Copernicus. This post is a small part of my project (Mapping Infrastructure and Monitor Urbanization in European Countries using deep learning) which I did at the German Research Centre for AI. The dataset is freely available and can be downloaded from http://madm.dfki.de/downloads(RGB color space images).

Fig1: Land use and land cover classification based on Sentinel-2 satellite images. Patches are extracted with the purpose to identify the shown class. This visualization highlights the classes annual crop, river, highway, industrial buildings, and residential buildings.

Fig 2: This overview shows sample image patches of all 10 classes covered in the proposed EuroSAT dataset. The images measure 64x64 pixels. Each class contains 2,000 to 3,000 image. In total, the dataset has 27,000 geo-referenced images.

What Strategy Worked

Fine tuning/Transfer Learning of a ResNet-50 architecture already trained on imagenet dataset worked really well for this challenge.

What is Fine Tuning/Transfer Learning of Network?

Fine Tuning is the process where we fine-tune existing networks(also known as Transfer Learning) like vgg-16, resnet-50, googleNet etc. that are already trained on a larger dataset like ImageNet(1.2M labeled images) by continuing training it (i.e. running back-propagation) on…