Tuning-Free Incremental Learning via Multilevel Prototypical Parts for Few-Shot Land Cover Segmentation

Abstract

Dynamic Earth observation applications often require land-cover segmentation models to recognize newly emerging classes from only a few pixel-level annotations while retaining previously learned knowledge. However, gradient-based incremental fine-tuning (FT) is prone to severe overfitting and catastrophic forgetting under such data-scarce conditions, and conventional prototype-based methods usually rely on a single global prototype that is insufficient for complex remote sensing scenes. To address these challenges, we propose a tuning-free multilevel prototypical parts network (MLPPN) for few-shot incremental land-cover segmentation. Inspired by recognition-by-components (RBCs) theory, MLPPN first constructs a discriminative library of multilevel prototypical parts from base classes, and then registers novel classes by gradient-free clustering and prototypical imprinting with all network parameters frozen. This design mitigates parameter drift during incremental learning and captures the heterogeneous part-level composition of land-cover objects. Experiments on four remote sensing and UAV datasets under two-shot and five-shot settings show that MLPPN achieves competitive segmentation accuracy and stable inference efficiency without incremental backpropagation.