Deep learning model improves PV segmentation in remote sensing images

A research team from South Korea’s Jeju National University has developed a novel deep learning PV segmentation technique for remote sensing images. Called the Solar-Mamba, the novel method is a PV-aware state-space network that unifies efficient long-range dependency modeling with geometry-constrained feature learning.

“Solar-Mamba is, to our knowledge, the first deep learning framework that explicitly incorporates the geometric and structural properties of PV panels their rectangular shape, grid-like row-column arrangement, and precise boundary edges, directly into the model’s architecture,” explained first author Kinza Nazir to pv magazine.

“Rather than treating solar panels as generic objects, we designed a dedicated Photovoltaic State-Space (PVSS) block that simultaneously reasons about global rooftop context, fine panel edges, and rectangular shape priors,” she added. “The integration of a grid-aware scanning mechanism further aligns the model’s sequential processing with how PV arrays are physically laid out. This combination of domain-specific design principles with efficient state-space modeling is what sets Solar-Mamba apart from prior convolutional neural networks (CNNs) and Transformer-based approaches.”

The Solar-Mamba technique begins with taking a preprocessed remote-sensing image and passing it through a U-Net-style encoder to extract features at different spatial scales. At each encoder stage, its PVSS block processes the features through three parallel branches: a Vision Mamba branch scans them horizontally, vertically, and diagonally to recognize the regular grid arrangement of PV arrays; a convolutional branch captures local textures, thin panel edges, and gaps between modules; and a shape-aware branch uses edge information and elongated filters to reinforce the panels’ rectangular geometry.

The outputs of these branches are then combined through learnable projection. In the decoder, the model progressively restores the image resolution using features transferred from the corresponding encoder stages, while an edge-refinement component helps sharpen panel boundaries. Finally, Solar-Mamba produces a pixel-level mask identifying the PV-covered areas. During training, a hybrid loss function penalizes missed panels, inaccurate panel regions, and blurred or irregular boundaries, enabling the model to generate masks that closely follow the actual shape and layout of the PV modules.

The researchers evaluated Solar-Mamba on two publicly available PV segmentation datasets, namely the multi-resolution PV dataset and the high-resolution photovoltaic segmentation (HRPVS) dataset. The first is a large multi-resolution dataset from Jiangsu Province, China, comprising 3,716 annotated image tiles acquired from satellites, with 0.8 m resolution; aerial platforms with a resolution of 0.3 m; and unmanned aerial vehicles (UAVs) with a resolution of 0.1 m. It includes both rooftop and ground-mounted PV systems across diverse environments.

The HRPVS  dataset, on the other hand, contained 717 manually annotated rooftop PV installations in Heilbronn, Germany. Generated from 0.15 m Google Earth imagery and divided into 280 image tiles, the dataset focuses on complex urban rooftops featuring shadows, occlusions, and diverse building geometries.

The Solar-Mamba was trained and evaluated separately on each dataset, using 80% of the images for training and 20% for testing. It was benchmarked against three categories of state-of-the-art segmentation models: CNN-based architectures, selected for their strong local feature extraction capabilities; Transformer-based architectures, chosen for their ability to model long-range global context through self-attention; and state-space-based architectures, included to assess the benefits of efficient long-range dependency modeling and isolate the contribution of Solar-Mamba’s photovoltaic-aware design.

“The most surprising finding was the strength of our boundary accuracy. We introduced a dedicated metric, boundary mean intersection over union (B-mIoU), to specifically evaluate how precisely the model delineates panel edges, not just region overlap,” explained Nazir. “Solar-Mamba achieved a B-mIoU of 84.0%, which was particularly impressive given that transformer-based models, despite their global attention mechanisms, fell below 79% on the same metric.”

On the HRPVS dataset, Solar-Mamba achieved a mean intersection over union (mIoU) of 89.84%, while the best CNN-based model scored 83.43%, the best Transformer-based model had 86.31%, and the best state-space model got 86.78%. On the multi-resolution dataset, Solar-Mamba again ranked first, reaching an mIoU of 90.12%, compared with 84.63% for the best CNN, 87.61% for the best Transformer, and 87.73% for the best state-space model.

“The model also proved to be the most computationally efficient in our comparison: with only 26.4 million parameters, 85.3 giga floating-point operations (GFLOPs), and a 33.9 ms inference time, it outperformed all CNN, Transformer, and Mamba-based baselines in accuracy while using less memory and compute,” Nazir added. “The idea that a model can simultaneously be the most accurate and the most lightweight was not something we expected so decisively from our experiments.”

In conclusion, Nazir said that her group has a few directions for follow-up research. “The most immediate priority is extending Solar-Mamba toward global-scale solar mapping, moving from individual image tiles to continuous, large-area monitoring,” she said. “We also plan to integrate temporal analysis, so the model can track panel degradation and support predictive maintenance over time. On the architecture side, we are working on lightweight variants suited for deployment directly on drone processors for real-time aerial monitoring.”

The novel method was presented in “Solar-mamba: A photovoltaic-aware state-space network for boundary-aware solar panel segmentation,” published in Applied Energy.