A Sensor Agnostic Domain Generalization Framework for Leveraging Geospatial Foundation Models: Enhancing Semantic Segmentation viaSynergistic Pseudo-Labeling and Generative Learning

TL;DR

Remotely sensed semantic segmentation suffers from cross-sensor variability and limited labeled data. The authors propose a sensor-agnostic domain generalization framework that fuses soft-alignment pseudo-labeling with source-to-target MAE generative pre-training on geospatial foundation models to improve cross-domain performance. They provide mathematical insights into how MAE-based generative learning yields a dynamic, confidence-weighted influence of unlabeled target pixels, and validate the approach on hyperspectral and multispectral datasets, showing robust improvements over strong baselines. The work demonstrates practical potential for robust GeoAI across diverse sensors and regions, with open-set and cross-domain scenarios well-addressed by the proposed synergy between MAE and domain alignment.

Abstract

Remote sensing enables a wide range of critical applications such as land cover and land use mapping, crop yield prediction, and environmental monitoring. Advances in satellite technology have expanded remote sensing datasets, yet high-performance segmentation models remain dependent on extensive labeled data, challenged by annotation scarcity and variability across sensors, illumination, and geography. Domain adaptation offers a promising solution to improve model generalization. This paper introduces a domain generalization approach to leveraging emerging geospatial foundation models by combining soft-alignment pseudo-labeling with source-to-target generative pre-training. We further provide new mathematical insights into MAE-based generative learning for domain-invariant feature learning. Experiments with hyperspectral and multispectral remote sensing datasets confirm our method's effectiveness in enhancing adaptability and segmentation.

Figures (8)

• Figure 1: Illustration of the proposed multi-task learning framework, represented by a total loss function comprising three learning objectives, each with its associated loss term: the primary segmentation task ($\mathcal{L}_{Seg}$), and two auxiliary tasks-domain alignment ($\mathcal{L}_{DA}$) and MAE-based self-supervision ($\mathcal{L}_{MAE}$).
• Figure 2: Illustration of the target and source domains used from the FLAIR dataset. The red and blue icons mark the geographical locations and data folders for the target and source domains, respectively.
• Figure 3: Prithvi model with adapters: feature extractor $f_\theta$, segmentation head $h_{\theta_{\text{Seg}}}$, and MAE generative head $g_{\theta_M}$. Layers are color-coded. Parameters: $L$ = transformer depth, $H$ = attention heads, $d$ = embedding dimension, $F$ = output width; $c$, $h$, $w$ = input channels, height, and width.
• Figure 4: Comparative Segmentation Inference using the C2Seg-AB Dataset. Display includes (a) Ground Truth Mask, (b) Our Method, (c) PCS, (d) Zero Shot, (e) GDA, (f) CDS, (g) MIC, (h) CIA_UDA, (i) UDA_ME_BS, and (j) Colorbar.
• Figure 5: Comparative Segmentation Inference using FLAIR Dataset. Display includes (a) Ground Truth Mask, (b) Our Method, (c) PCS, (d) Zero Shot, (e) GDA, (f) CDS, (g) MIC, (h) CIA_UDA, (i) UDA_ME_BS, and (j) Colorbar.