Generalized Few-Shot Meets Remote Sensing: Discovering Novel Classes in Land Cover Mapping via Hybrid Semantic Segmentation Framework

TL;DR

The paper tackles the challenge of updating high-resolution land-cover maps to incorporate newly appearing classes under limited labeled data. It presents SegLand, a GFSS-based framework with three components: data preprocessing to balance and augment data, a hybrid segmentation module combining multiple base learners and a Projection onto Orthogonal Prototypes (POP) network, and an ultimate fusion stage to produce robust maps that reflect both base and novel classes. Key contributions include novel data augmentation (NovelCutMix), a four-model base-learner ensemble with average fusion, and a two-stage POP training regime with orthogonality constraints and backbone/decoder selections (notably Swin Transformer + UperNetPlus) that enable effective novel-class updating without degrading base-class performance. The approach achieves top performance in the OpenEarthMap Land Cover Mapping Few-Shot Challenge, demonstrating practical potential for fast, scalable updates of HR land-cover maps with limited annotated data, and offering a pathway to more responsive Earth observation products and downstream applications.

Abstract

Land-cover mapping is one of the vital applications in Earth observation, aiming at classifying each pixel's land-cover type of remote-sensing images. As natural and human activities change the landscape, the land-cover map needs to be rapidly updated. However, discovering newly appeared land-cover types in existing classification systems is still a non-trivial task hindered by various scales of complex land objects and insufficient labeled data over a wide-span geographic area. In this paper, we propose a generalized few-shot segmentation-based framework, named SegLand, to update novel classes in high-resolution land-cover mapping. Specifically, the proposed framework is designed in three parts: (a) Data pre-processing: the base training set and the few-shot support sets of novel classes are analyzed and augmented; (b) Hybrid segmentation structure; Multiple base learners and a modified Projection onto Orthogonal Prototypes (POP) network are combined to enhance the base-class recognition and to dig novel classes from insufficient labels data; (c) Ultimate fusion: the semantic segmentation results of the base learners and POP network are reasonably fused. The proposed framework has won first place in the leaderboard of the OpenEarthMap Land Cover Mapping Few-Shot Challenge. Experiments demonstrate the superiority of the framework for automatically updating novel land-cover classes with limited labeled data.

Figures (11)

• Figure 1: Illustration of discovering novel classes in the general land-cover mapping process. (a) Earth observation provides HR remote-sensing images. (b) Base classes contain sufficient training set. (c) Novel classes suffer from insufficient labeled data (i.e., few-shot support set).
• Figure 2: Overall flowchart of the GFSS-based land-cover mapping framework (SegLand), containing three main parts: (a) Data pre-processing, (b) Hybrid segmentation, and (c) Ultimate fusion.
• Figure 3: Illustration of CutMix augmentation strategy. The images of the base training set are regarded as 'backgrounds' for novel classes.
• Figure 4: Illustration of CutMix augmentation example. The novel class of vehicle is mixed in a random position of the image.
• Figure 5: The overall framework of the adopted POP framework for GFSS, which consists of two phases: (a) Base class learning, (b) Novel class updating.