Comparative Analysis of Unsupervised Pre-Training Models for LULC Classification With Limited Labelled Data

Accurate classification of land use and land cover (LULC) categories from satellite imagery remains a pivotal task in remote sensing and geospatial analysis. The efficacy of classification models is often hindered by the scarcity of labeled samples and the degree of noise present in the labeled data, posing prevalent challenges in the real-world applications. This paper presents a study aimed at addressing these challenges by systematically comparing six combined generative pre-training and discriminative models with two standalone discriminative models in the context of LULC classification. Through meticulous empirical analysis, this study evaluates the performance of both generative and discriminative models when confronted with limited labeled data. The comparison not only highlighted the strengths and weaknesses of each model, but also provided insights into their adaptability and utility under constrained and noisy labeling resources, which are common scenarios in real-world LULC classification tasks. Additionally, the comparison offer insights into how generative models can leverage unlabeled data to augment their predictive capabilities, while also high-lighting the efficiency of traditional discriminative models when dealing with limited labeled data.