Image Classification to identify Transverse Cirrus Band Clouds using Convolutional Neural Networks
Gregory Saini, Thilanka Munasinghe, Kevin Pan · 2021 IEEE International Conference on Big Data (Big Data) · 2021
The NASA IMPACT team is investigating a certain type of cloud called Transverse Cirrus Bands (aka. TCB), that can cause dangerous levels of aviation turbulence for nearby planes. We wanted to collaborate with the NASA IMPACT team by exploring how machine learning models that can detect Transverse Cirrus Band clouds in satellite images. To accomplish this task we decided to use various CNN models and data preprocessing methods to determine which would produce the most accurate detection. For our CNN models, we chose Simple Sequential, AlexNet, ResNet50, LetNet5, Google MobileNetV2, and VGG-16. For our data preprocessing methods we shrank our training images, converted images into grayscale, and turned images into edges by using edge detection. Our results showed that training CNNs on the originally colored satellite images produced the best results for most models. We believe there is still potential for images converted into edges to improve model accuracy, but we would need further work to calculate an appropriate edge detection threshold for satellite images of different sizes.