Mapping Land Coverage Using Drone Imagery and Efficient Image Transformer Models

Abstract

Solving environmental and anthropological problems requires effective public policy, which is only attainable when given accurate land use accounting. However, developing nations lack systems that can track these resources in real-time. Recent technological improvements have enabled Unmanned Aerial Vehicles (UAVs) to be deployed easily and inexpensively to image forests, farmland, and settlements. Manually processing such images to generate useful insights into environmental conditions and land use is time-consuming and expensive. Current automated tools utilize machine learning models like the Convolutional Neural Network, which require large dataset sizes. Nevertheless, collecting the required data is still slow and expensive. This study adapts and improves Vision Transformer models. Specifically, this study modifies the architecture of the Data Efficient Image Transformer (DEIT) to classify land cover images more accurately. This framework will provide developing nations with the accurate data and resources to create public policy. The optimized model reached a phenomenal F1-score of 95% and AUC of 99.85%. The outcome of this research is a precise machine learning-based framework that can produce accurate land coverage maps, a major stepping stone required for building a sustainable future.