Published January 1, 2026 | Version v1

Deep Learning Based Semantic Segmentation of Satellite Imagery for Land Cover Classification

Description

Land cover classification from remote sensing imagery underpins agricultural monitoring, urban planning, disaster response, and environmental assessment, and deep learning has become the dominant paradigm for extracting land cover information from satellite scenes at pixel-level granularity. This survey reviews recent advances in deep learning-based land cover classification and semantic segmentation, organizing the literature into convolutional neural network (CNN)-based encoder–decoder architectures, attention and self-attention mechanisms, transformer and hybrid CNN–Transformer architectures, and optimization-augmented and foundation-model-adapted approaches. We summarize the datasets, evaluation metrics, and comparative performance reported across the reviewed studies, and use a representative hybrid EfficientNet–Vision Transformer framework as a case study to illustrate current design trends. Persistent challenges are identified, including class imbalance for narrow structures such as roads, the computational cost of hybrid transformer-based models, limited cross-dataset generalization, and the scarcity of densely annotated data. The survey concludes by outlining promising future directions, including lightweight hybrid architectures, self-supervised and weakly supervised learning, adaptation of large vision foundation models, multi-sensor data fusion, and explainable land cover mapping.

Files

IJSET_V14_issue4_238.pdf

Files (579.2 kB)

Name Size
md5:e4a3ec3e5156dbbb17cbffad28b371cb
579.2 kB Preview Download

Additional details