.png){kind=icon}
.png){kind=icon}
Return
PANEL: A photovoltaic-specific vision-language model for zero-shot and few-shot PV interpretation tasks in remote sensing
.png){kind=icon}
PRE
AI
R
Z
P
J
X
Q
Y
J
DOI：10.1016/j.isprsjprs.2026.05.012.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Large-scale photovoltaic (PV) mapping from remote sensing (RS) is crucial for solar energy assessment and infrastructure development. While vision-language models (VLMs) have advanced language-guided image understanding, existing RS CLIP models underperform on PV across changing regions, sensors, and imaging conditions due to the scarcity of PV imagery and aligned prompts in open RS multi-modal datasets. To address this, we curate the first worldwide PV vision-language dataset, comprising over one million image-text pairs spanning 0.1-20 m resolutions, and design aligned flexible binary prompts that explicitly distinguish PV-present from PV-absent imagery. We then build PANEL (PV-specific vision-lANguage modEL) by retraining CLIP using a tailored pipeline featuring extensive augmentation and prompt diversification. For zero-shot inference, we introduce PANEL Surgery to reinforce vision-language alignment and extract PV-relevant priors for PV interpretation tasks. For few-shot adaptation, we propose a plug-in Knowledge Assistance Module (KAM) that injects these priors into diverse baselines via gated convolution and cross-attention fusion. To enable rigorous evaluation, we establish seven PV-specific benchmarks spanning five interpretation tasks and observe that PANEL consistently surpasses representative CLIP models in zero-shot classification (>40% accuracy), semantic localization (>45% mSC), segmentation (>25% IoU), detection (>24% recall), and image generation (>7% ΔCLIP reduction). In few-shot settings, KAM generalizes across multiple architectures with minimal labels, and ablation studies confirm the necessity of vision-language priors and the complementary design of each component. We further introduce a label-efficient sampling strategy leveraging VLM-derived similarity scores, consistently outperforming random selection under tight annotation budgets. Collectively, these results establish PANEL as a unified, PV-specific VLM for robust zero-/few-shot interpretation while reducing annotation costs in real-world RS for solar energy. PANEL and benchmarks are publicly available at https://huggingface.co/UEXdo .
Keywords:
photovoltaic mapping
vision-language model
remote sensing
zero-shot learning
few-shot adaptation
Journal
IF:
12.2
Papers:
4.3K
Citations:
3.2W
View more organizations