.png){kind=icon}
.png){kind=icon}
Return
RS-LightSSL: An Efficient Lightweight Self-Supervised Learning Framework for Remote Sensing Multitasks
.png){kind=icon}
PRE
AI
DOI：10.1109/TGRS.2026.3660842.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Recent years have witnessed the rise of remote sensing (RS) edge artificial intelligence (AI), showing exceptional performance in various downstream tasks by efficiently deploying vision-pretrained models on resource-constrained edge devices. However, mainstream knowledge distillation (KD) methods are prone to accuracy bottlenecks, as they often overlook task-agnostic knowledge essential for model generalization. In contrast, human cognition evolves as a progressive process from simple to complex concepts, gradually developing and mastering problem-solving strategies. Inspired by this human learning process, we introduce RS-LightSSL, a new and effective lightweight model learning framework for RS edge AI. During the pretraining phase, we introduce the large model-driven progressive masked image modeling (LMP-MIM) strategy, which leverages large models to guide the acquisition of rich task-agnostic knowledge, thereby enhancing model generalization. In the fine-tuning phase, we implement a bifurcated convolution-enhanced adapter (Fine-Adapter), which integrates an additional convolutional branch with a parameter-efficient fine-tuning (PEFT) adapter, thus bolstering the retention of task-specific knowledge. Extensive experimental validation has been conducted across three downstream tasks, accompanied by comprehensive ablation studies. The proposed RS-LightSSL achieves performance comparable to large models with less than 40% parameters, while achieving state-of-the-art (SOTA) results among similar-sized models.
Keywords:
Edge artificial intelligence (AI)
lightweight model
parameter-efficient fine-tuning (PEFT)
remote sensing (RS)
self-supervised learning (SSL)
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W