Lightweight Transformer-Driven Segmentation of Hotspots and Snail Trails in Solar PV Thermal Imagery Journal: Solar Energy
Deepak Joshi, Mayukha Pal · 2025
Read Lightweight Transformer-Driven Segmentation of Hotspots and Snail Trails in Solar PV Thermal Imagery Journal: Solar Energy by Deepak Joshi, Mayukha Pal (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-20680.txt.
Introduction
Deepak Joshi and Mayukha Pal teach a light transformer to paint solar panel faults in drone heat images. Hotspots and snail trails sap a module's output and can worsen unseen. Their 2025 study starts from 277 aerial thermal images shot by a drone-mounted infrared camera. Contrast boosting, denoising, and normalization prep each frame. A trimmed SegFormer model then segments faults pixel by pixel, trained on hand-labeled regions.
Four rival networks face the same data and prep. The light model wins on accuracy and efficiency, especially for small, ragged defects. Scores span Dice, F1, kappa, overlap, and pixel accuracy. Its small footprint suits edge devices on the drone itself. Farm-scale automatic inspection is the target use.
Catalog record
| CREATOR | Deepak Joshi, Mayukha Pal |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-20680
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Lightweight Transformer-Driven Segmentation of Hotspots and Snail Trails in Solar PV Thermal Imagery Journal: Solar Energy Deepak Joshi, Mayukha Pal Abstract Accurate detection of defects such as hotspots and snail trails in photovoltaic (PV) modules is essential for maintaining energy efficiency and system reliability. This work presents a supervised deep learning framework for segmenting thermal infrared images of PV panels, using a dataset of 277 aerial thermographic images captured by a Zenmuse XT infrared camera mounted on a DJI Matrice 100 drone. The preprocessing pipeline includes image resizing, CLAHE-based contrast enhancement, denoising, and normalisation. A lightweight semantic segmentation model based on SegFormer is developed, featuring a customised Transformer encoder and streamlined decoder, and fine-tuned on annotated images with manually labelled defect regions. To evaluate performance, we benchmark our model against U-Net, DeepLabV3, PSPNet, and Mask2Former using consistent preprocessing and augmentation. Evaluation metrics include per-class Dice score, F1-score, Cohen’s kappa, mean IoU, and pixel accuracy. The SegFormer-based model outperforms baselines in accuracy and efficiency, particularly for segmenting small and irregular defects. Its lightweight design enables real-time deployment on edge devices and seamless integration with drone-based systems for automated inspection of large-scale solar farms. † † graphicalabstract: † † highlights: A lightweight and customized SegFormer-based segmentation framework is developed for the analysis of thermal images in solar photovoltaic (PV) modules. The proposed model is rigorously evaluated using a suite of metrics, including mean Intersection over Union (mIoU), pixel-wise accuracy, F1-score, Dice coefficient, and Cohen’s kappa score. Detailed class-wise segmentation performance is reported, with a specific focus on identifying hotspot and snail trail defects. The framework’s effectiveness is further validated through comparative experiments against established segmentation architectures such as U-Net, DeepLabV3, and PSPNet. 1 Introduction The global push toward sustainable energy has accelerated in response to climate change, energy security concerns, and the environmental degradation caused by the prolonged use of fossil fuels. Traditional energy sources such as coal, oil, and natural gas are not only finite but are also major contributors to greenhouse gas emissions, which are driving global warming and extreme climate events [ 1 , 2 ] . In line with international agreements such as the Paris Climate Accord, countries worldwide are investing heavily in renewable energy sources to reduce carbon emissions and ensure a sustainable energy future. Among these, solar energy has emerged as one of the most promising and widely adopted alternatives due to its abundance, scalability, and zero-emission nature during operation [ 3 ] . Photovoltaic (PV) technology, which directly converts solar radiation into electricity, has seen unprecedented growth in the last decade. According to recent estimates, global PV capacity surpassed 2.2 terawatts by the end of 2024 and is projected to continue expanding rapidly. Solar PV modules are now deployed in residential rooftops, commercial installations, and utility-scale solar farms across the globe [ 4 ] . However, the long-term efficiency and safety of PV systems are compromised by various types of physical and thermal defects such as hotspots, snail trails, microcracks, and delamination. These faults can emerge due to manufacturing imperfections, mechanical stress, or prolonged environmental exposure, and may lead to power loss, thermal degradation, and even safety hazards if left undetected [ 5 ] . Among the commonly observed faults, hotspots and snail trails are particularly critical due to their impact on performance and their prevalence in real-world deployments. A hotspot forms when a portion of a PV cell experiences higher resistance—often due to shading, cell cracks, or soiling—which leads to localized heating as current is forced through the impaired region. This not only reduces the power output but also accelerates material degradation and can even result in module failure if not addressed [ 6 ] . On the other hand, snail trails appear as discolored, worm-like patterns on the surface of PV modules. These are typically caused by microcracks within the silicon cells, often induced by mechanical stresses during transportation, installation, or operation. Over time, moisture ingress and chemical reactions with silver paste traces at these cracks lead to visible discoloration. While snail trails may initially seem superficial, they often signal underlying structural defects that contribute to long-term efficiency losses and reliability concerns in PV systems [ 7 ] . Spataru et al. [ 8 ] introduced a two-dimensional matched filtering technique to localize micro-cracks in photovoltaic (PV) modules. Their method generates a binary location map by applying post-processing on filtered outputs, allowing for spatial identification of potential crack sites. Similarly, Rashmi et al. [ 9 ] leveraged the Mamdani fuzzy logic algorithm to extract micro-cracked areas and quantify their surface coverage in PV cell imagery, enabling area-based defect discrimination. Although both approaches offer targeted crack localization, they rely on classical image processing heuristics that are sensitive to parameter tuning. As noted by Jiang et al. [ 10 ] , such techniques depend heavily on expert-defined configurations for denoising and feature extraction. This manual dependency restricts their adaptability when operating under diverse or unpredictable imaging conditions, especially in the presence of complex backgrounds and varying textures. Traditional image processing techniques have served as the foundation for early photovoltaic (PV) module defect analysis. Methods such as thresholding, edge detection, region growing, clustering, and graph-based segmentation have been widely adopted to detect surface anomalies. For example, S. Gayathri et al. [ 11 ] applied a hybrid Otsu-Canny edge detection algorithm to isolate micro-cracks in silicon solar cell images, and Montanez et al. [ 12 ] proposed a thermal analysis tool that segments individual PV modules from UAV-acquired infrared images using an automatic pipeline combining Otsu thresholding and contour detection to compute temperature statistics for fault identification. In a similar vein, Prasshanth et al. [ 13 ] proposed another UAV-based study combined histogram-based statistical features (mean, variance, skewness), texture features derived from Gray-Level Co-occurrence Matrix (GLCM), and frequency-domain features using Discrete Wavelet Transform (DWT) to characterize thermal anomalies, followed by Rough Set Theory (RST) for classification and rule generation, offering a multi-feature perspective for PV fault analysis. Cao et al. [ 14 ] enhanced morphological and edge-based techniques for improved precision. Attia et al. [ 15 ] leveraged K-means clustering with hyperspectral imaging for real-time PV fault detection, and Zhang et al. [ 16 ] used a combination of Prewitt and Canny filters to identify micro-cracks. Despite their early promise, these techniques often exhibit poor generalization under complex environmental conditions and diverse imaging scenarios, limiting their robustness in real-world deployments. To improve scalability and coverage in PV diagnostics, aerial inspection using UAVs has emerged as a practical solution. Quater et al. [ 17 ] proposed a UAV-based monitoring framework utilizing both thermal and optical imaging to detect a variety of module faults across solar farms.In a related effort, Chris et al. [ 18 ] implemented an automated detection system combining drone-captured thermal images with binarization and contour extraction to isolate deteriorated PV modules. Their system leveraged bounding box localization and integrated color-based analysis to enhance visual differentiation of damaged regions. Expanding on this direction, Aghaei et al. [ 19 ] developed a thermographic image analysis algorithm involving grayscale conversion, Gaussian filtering, binary segmentation, and Laplacian-based boundary modeling to highlight fault regions. Their subsequent studies [ 20 , 21 ] investigated how flight altitude impacts defect visibility and