Evaluation of Images Related to Climate Change with Deep Learning Models

Abstract

This study is developing a Deep Learning model automating the coding of drawings students provide about climate change phenomena in our world, as a learning contribution through formative assessment. We started first with ResNet50 architecture, but ultimately, we settled on MobileNetV2 reduced architecture for the sake of being able to integrate with mobile- and web-based applications. The challenge is the model has very few examples in the training set to work with, so we decided augmenting the data (i.e., rotate, zoom, flip horizontally,) will help the model generalize more reliably. The model achieved training accuracy of 92% and validation accuracy of 90%. Moreover, we were able to reduce the model size about 85% through optimization. Our model outputs not a simple classification, it also produces explanatory feedback for each class, and we have made possible for the feedback to be read by the student about their idea. Our findings are indicating, it is possible to use AI-based systems to teach how to investigate integrated fields like environmental education. Future studies will include multi-label classification, explainable AI (XAI) methodologies and dataset sizes will also increase.