Accurate PhotoVoltaic (PV) panel segmentation from Remote Sensing Imagery (RSI) is essential for large-scale solar inventory generation, capacity estimation, and renewable energy monitoring. However, PV installations vary significantly in size, appearance, spatial resolution, and acquisition conditions, while often occupying only a small portion of the image. In this paper, we present a comparative study of six representative Deep Learning (DL)-based architectures for PV panel segmentation: Fully Convolutional Network (FCN), SegFormer-B3, SegNet, SwinUNet, TransUNet, and U-Net. The models are evaluated under a unified experimental protocol on three publicly available datasets—namely, (i) Global multispectral dataset for Solar Farm IDentification in satellite imagery (GloSoFarID), (ii) a multi-resolution PV dataset, and (iii) Base de Données PhotoVoltaïque (BDPV)—covering multispectral satellite imagery, as well as RGB satellite, aerial, and Unmanned Aerial Vehicle (UAV) imagery at heterogeneous spatial resolutions. All models are trained from scratch using the same data augmentation, hybrid loss function, deep supervision strategy, training configuration, and evaluation metrics. In addition to segmentation accuracy, computational complexity is analyzed in terms of trainable parameters and Floating Point Operations (FLOPs). Experimental results show U-Net achieving the best overall segmentation performance across the three considered datasets, obtaining the highest Intersection over Union (IoU), mean Intersection over Union (mIoU), precision, F1-score, and accuracy, but unfortunately requiring the highest FLOPs. Instead, FCN provides a competitive accuracy–complexity compromise, especially on multi-resolution and BDPV datasets, while SwinUNet and SegFormer-B3 offer lower computational cost but reduced segmentation accuracy. These results highlight the trade-offs between segmentation performance, false positive suppression, model complexity, and deployment efficiency for PV panel mapping under heterogeneous remote sensing conditions.

A Comparative Study of Deep Learning Architectures for Photovoltaic Panel Segmentation in Remote Sensing Imagery / Battaglioli, D., Mazinani, A., Fontechiari, P., Ferioli, N., Belli, L., Davoli, L., Ferrari, G.. - 14283:(2026), pp. 1-14. (SPIE Environmental Remote Sensing Edinburgh, United Kingdom ) [10.1117/12.3116744].

A Comparative Study of Deep Learning Architectures for Photovoltaic Panel Segmentation in Remote Sensing Imagery

Battaglioli, Daniele;Mazinani, Armin;Belli, Laura;Davoli, Luca;Ferrari, Gianluigi
2026-01-01

Abstract

Accurate PhotoVoltaic (PV) panel segmentation from Remote Sensing Imagery (RSI) is essential for large-scale solar inventory generation, capacity estimation, and renewable energy monitoring. However, PV installations vary significantly in size, appearance, spatial resolution, and acquisition conditions, while often occupying only a small portion of the image. In this paper, we present a comparative study of six representative Deep Learning (DL)-based architectures for PV panel segmentation: Fully Convolutional Network (FCN), SegFormer-B3, SegNet, SwinUNet, TransUNet, and U-Net. The models are evaluated under a unified experimental protocol on three publicly available datasets—namely, (i) Global multispectral dataset for Solar Farm IDentification in satellite imagery (GloSoFarID), (ii) a multi-resolution PV dataset, and (iii) Base de Données PhotoVoltaïque (BDPV)—covering multispectral satellite imagery, as well as RGB satellite, aerial, and Unmanned Aerial Vehicle (UAV) imagery at heterogeneous spatial resolutions. All models are trained from scratch using the same data augmentation, hybrid loss function, deep supervision strategy, training configuration, and evaluation metrics. In addition to segmentation accuracy, computational complexity is analyzed in terms of trainable parameters and Floating Point Operations (FLOPs). Experimental results show U-Net achieving the best overall segmentation performance across the three considered datasets, obtaining the highest Intersection over Union (IoU), mean Intersection over Union (mIoU), precision, F1-score, and accuracy, but unfortunately requiring the highest FLOPs. Instead, FCN provides a competitive accuracy–complexity compromise, especially on multi-resolution and BDPV datasets, while SwinUNet and SegFormer-B3 offer lower computational cost but reduced segmentation accuracy. These results highlight the trade-offs between segmentation performance, false positive suppression, model complexity, and deployment efficiency for PV panel mapping under heterogeneous remote sensing conditions.
2026
A Comparative Study of Deep Learning Architectures for Photovoltaic Panel Segmentation in Remote Sensing Imagery / Battaglioli, D., Mazinani, A., Fontechiari, P., Ferioli, N., Belli, L., Davoli, L., Ferrari, G.. - 14283:(2026), pp. 1-14. (SPIE Environmental Remote Sensing Edinburgh, United Kingdom ) [10.1117/12.3116744].
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11381/3079474
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact