Evaluating Deep Learning Models for Land Cover Classification via Test-Time Augmentation

The practice of classifying and mapping the various land cover or land use categories such as buildings, water bodies, forests, etc. that exist in a given area is known as land cover classification. It is a key task in geographic information systems (GIS) and remote sensing, with applications ranging from natural resource management and disaster response to urban planning and environmental monitoring. There has been a lot of research focused on detection of land cover using Machine Learning and Deep Learning. Unfortunately, there is not much research on assessing the robustness of models trained for these particular tasks. In this paper, we present a comparative assessment of three cutting-edge deep learning models: UNet++, Feature Pyramid Network (FPN), and Pyramid Scene Parsing Network (PSPNet) for land cover classification on one of the datasets from DeepGlobe 2018 and look over the model's performance on test data which are deformed and augmented in different ways. We assess the efficacy of these models using criteria including Accuracy, Precision, F1-Score, Recall, and IoU on a particular dataset. Our tests show how well each model captures complex spatial patterns and reliably classifies a range of land cover groups as we will be implementing different image augmentation methods on the test dataset. Additionally, we examine each model's advantages and disadvantages in managing various land cover types and talk about how applicable each model is in various scenarios including environmental monitoring and management.