Semi Supervised LULC segmentation with state of the art encoder IJEPA

Abstract

Land-Use/Land-Cover (LULC) mapping from satellite imagery is a fundamental component of environmental monitoring, agricultural planning, disaster management, and urban development assessment. With the rapid growth of remote sensing technology, large volumes of high-resolution satellite images have become readily available for most countries. However, transforming these raw images into meaningful semantic maps requires accurate semantic segmentation models capable of consistently identifying classes such as farmland, water bodies, forested regions, built-up areas, and meadow zones. Also, it has been seen that specifically for satellite imagery data from one climate zone tend to suffer performance degradation when applied to regions with different climate/environment characteristics, indicating that segmentation models generalize better when trained and tested on the same regional data [21] [18]. Experiments on the LoveDA[34] data set show that models trained in one geographic style obtain significantly better segmentation performance when evaluated in regions with similar urban characteristics than when transferred to rural scenes with different land-cover distributions and visual appearances. Hence, Region specific data are required for good segmentation results.