SegDesicNet: Lightweight Semantic Segmentation in Remote Sensing with Geo-Coordinate Embeddings for Domain Adaptation

TL;DR

This paper tackles the domain shift challenge in semantic segmentation of high-resolution remote sensing imagery by introducing SegDesicNet, an unsupervised domain adaptation framework that leverages geo-coordinate metadata. It integrates GRID-based geographic encodings projected onto the unit sphere and trains a cosine dissimilarity-based domain loss alongside a lightweight segmentation model to align source and target domains without labels. The approach yields about a 6% mean IoU improvement while reducing model size by roughly 27% on benchmarks such as FLAIR #1 and ISPRS Potsdam, demonstrating strong cross-domain robustness. The work highlights the potential of human-centric geographic reasoning in neural networks and outlines future directions for more efficient geo-encoding and cross-sensor generalization.

Abstract

Semantic segmentation is essential for analyzing highdefinition remote sensing images (HRSIs) because it allows the precise classification of objects and regions at the pixel level. However, remote sensing data present challenges owing to geographical location, weather, and environmental variations, making it difficult for semantic segmentation models to generalize across diverse scenarios. Existing methods are often limited to specific data domains and require expert annotators and specialized equipment for semantic labeling. In this study, we propose a novel unsupervised domain adaptation technique for remote sensing semantic segmentation by utilizing geographical coordinates that are readily accessible in remote sensing setups as metadata in a dataset. To bridge the domain gap, we propose a novel approach that considers the combination of an imageś location encoding trait and the spherical nature of Earthś surface. Our proposed SegDesicNet module regresses the GRID positional encoding of the geo coordinates projected over the unit sphere to obtain the domain loss. Our experimental results demonstrate that the proposed SegDesicNet outperforms state of the art domain adaptation methods in remote sensing image segmentation, achieving an improvement of approximately ~6% in the mean intersection over union (MIoU) with a ~ 27\% drop in parameter count on benchmarked subsets of the publicly available FLAIR #1 dataset. We also benchmarked our method performance on the custom split of the ISPRS Potsdam dataset. Our algorithm seeks to reduce the modeling disparity between artificial neural networks and human comprehension of the physical world, making the technology more human centric and scalable.

Figures (3)

• Figure 1: Our model contains ResNet-18he2016deep asencoder ($\mathcal{E}$), a decoder ($\mathcal{D}$) and the SegDesic module. The encoder and decoder together form a U-Netronneberger2015unet. The structure of the SegDesic module with its input and output sizes is shown on the lower block in the figure. In this diagram, $\dashrightarrow$ indicates its skip connections (with feature concatenation).CONV represents a 2D convolution layer, UPCONV signifies a 2D transposed convolution layer, $*\times *$ refers to the convolution kernel size, $*,/*'$ represents the number of kernels and convolution stride, $\dot*$ represents batch size, and $N$ represent number of class.
• Figure 2: Shown are the training and test set splits of the ISPRS Potsdam dataset (on the left) and the FLAIR #1 dataset (on the right), with custom splits of the source and target domains indicated by colored dots.
• Figure 3: Comparison of segmentation map prediction on the target domain images of SPLIT1 by baseline models to our best model.