This repository features training/testing and data visualisation methods for convolutional neural networks. Models and results have been saved for the MNIST and euroSAT databases for local testing (model parameters may have to be adjusted to match).
Achieved 97.7% accuracy on Best two for a non-pretrained model for the euroSAT database for a small CNN. The results for this model can be seen below against the bacth count x-axis.
This is an example sample showing a correctly identified patch. Its second guess highlights how the model views similar images, also considering that the urban 'Residential' patch could be a highway, potentially due to the presence of straight roads cutting through the entire image.
