Evaluating the performance of deep learning-based segmentation algorithms for land use land cover mapping in a heterogenous vegetative environment

Land Use and Land Cover (LULC) maps are important geospatial information sources for different applications such as city planning, vegetation analysis, natural resource management, natural disaster analysis, and land change determination. In recent decades, the demand for more frequent creation and updating of LULC maps has grown significantly, driven by the rapid and continuous changes occurring on the Earth surface. Moreover, the increased availability of satellite images and processing power led to improvements in LULC mapping. However, traditional classification approaches are prone to several errors emerging from high human interaction and algorithm limitations. In addition, they generally suffer from processing time performance due to software limitations and generally singular hardware configurations, especially when very high resolution (VHR) images are of concern. In this study, we aim to produce LULC maps of the Aksu region of Bursa city Türkiye, using Worldview-3 VHR images and deep learning (DL) methods. We applied two widely used DL architectures, Unet++ and DeepLabv3+, and evaluated results using overall accuracy, average accuracy, error matrix, weighted accuracy, recall, precision, F-1 score, IoU score, and kappa metrics. Among several experimental setups, we achieved the best accuracy with the Unet++ architecture, using the ResNeXt-50 backbone and Adam optimizer, resulting in an approximately 84% IoU score and 91% F-1 score. This study demonstrates that utilizing appropriate datasets and CNN-based segmentation models for LULC mapping ensures efficient, accurate, and high-performance results, significantly contributing to long-term monitoring and sustainable development goals.