Analysis of Preprocessing Techniques, Keras Tuner, and Transfer Learning on Cloud Street image data

A Convolutional Neural Network is a powerful tool that has been extensively used for image classification. One specific area of application is remotely sensed images of meteorological phenomena such as cyclones and high latitude dust events. Such images are complicated in nature and hence may require special techniques for feature identification. Cloud streets are another such phenomenon that occurs in nature and is mainly captured in images taken by artificial satellites. In this work, deep learning models are implemented on NASA-IMPACT teams' cloud street dataset. Three preprocessing techniques were tested to address the drawbacks of the dataset. Gaussian blur, census transformation to extract textural features, data augmentation, and removal of noise were implemented. Then techniques such as Keras tuner are also utilized for hyperparameter tuning to help achieve maximum accuracy. The results show the efficiency with which Keras tuner attempts to direct towards optimal hyperparameters restricting the number of iterations to a low value and obtaining a test dataset accuracy as high as 80.96%. Lastly, binary classification of Cloud Street satellite images is performed by leveraging the benefits of Transfer Learning and pre-trained models. The various architectures that were tested in this work were namely, EfficientNetB7, AlexNet, VGG19 and, InceptionNetV3. Transfer learning provides a quick approach to build deep learning models with good accuracy scores achieving high accuracies of 80.89% for the VGG19 architecture.