2TSS: Two-tier semantic segmentation framework with enhancement for hotspot detection of solar photovoltaic thermal images
Recently, intelligence-based hotspot detection has been widely used in solar photovoltaic (PV) image applications. However, the semantic segmentation approach has limitations in terms of accuracy, particularly for hotspot thermal images. This study introduces a novel method based on Two-tier Semanti...
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025
|
| Subjects: | |
| Online Access: | http://umpir.ump.edu.my/id/eprint/45020/ http://umpir.ump.edu.my/id/eprint/45020/1/2TSS-%20Two-tier%20semantic%20segmentation%20framework%20with%20enhancement.pdf |
| Summary: | Recently, intelligence-based hotspot detection has been widely used in solar photovoltaic (PV) image applications. However, the semantic segmentation approach has limitations in terms of accuracy, particularly for hotspot thermal images. This study introduces a novel method based on Two-tier Semantic Segmentation (2TSS) framework explicitly aimed at enhancing hotspot detection in thermal images of PV modules. The proposed method is designed with two subsequent stages of segmentation models, including image pre-processing at the initial of the framework. The first tier of segmentation distinguishes between solar PV modules and the background, whilst the second tier enhances the hotspot localization region. This research enhances comprehension of multi-tier segmentation architectures in deep learning, focusing on optimizing performance for solar energy systems through comparative analysis of semantic models. Three different segmentation models, namely U-Net, ResNet 18 and ResNet 50 were tested. The ResNet 50 model demonstrated superior segmentation performance across both tiers with 98% and 85% accuracy respectively. In summary, the proposed method demonstrates that applying a combined enhancement algorithm prior to training for hotspot segmentation promotes superior performance with an accuracy improvement of 2.26% over the non-enhancement approach. |
|---|