MAPLE: Modality-Adaptive Prior-Guided Learning With Language-Conditioned Spatial Adaptation for Multimodal Remote Sensing Segmentation

Abstract

Multimodal remote sensing segmentation benefits from complementary information across heterogeneous data sources, yet effectively leveraging such information while preserving the generalization ability of pretrained models remains challenging. In this article, we propose modality-adaptive prior-guided learning (MAPLE), a unified framework that adapts a frozen foundation backbone for multimodal learning via lightweight modality-aware mechanisms and language-guided priors. By introducing adaptive feature modulation and semantic guidance, the proposed approach effectively mitigates modality discrepancies and improves spatial prediction consistency. In addition, an adaptive reweighting strategy is employed to address class imbalance and enhance the learning of underrepresented categories. Extensive experiments on multiple benchmarks demonstrate that MAPLE achieves state-of-the-art accuracy with the smallest trainable footprint among the compared methods, while a comprehensive efficiency analysis of total parameters, floating-point operations (FLOPs), inference latency, and memory consumption further confirms a favorable accuracy–efficiency trade-off. The source code is available at https://github.com/shiroSamaNanoda/MAPLE.git