Date of Graduation

12-2023

Document Type

Thesis

Degree Name

Master of Science in Computer Science (MS)

Degree Level

Graduate

Department

Computer Science & Computer Engineering

Advisor/Mentor

Le, Ngan

Committee Member

McCann, Roy A.

Second Committee Member

Mantooth, H. Alan

Keywords

Computer Vision; Deep Learning; Photovoltaic Energy; Segmentation

Abstract

This thesis introduces a novel approach to Photovoltaic (PV) installation segmentation by proposing a new architecture to understand and identify PV modules from overhead imagery. Pivotal to this concept is the creation of a new Transformer-based network, S3Former, which focuses on small object characterization and modelling intra- and inter- object differentiation inside an image. Accurate mapping of PV installations is pivotal for understanding their adoption and guiding energy policy decisions. Drawing insights from current Deep Learning methodologies for image segmentation and building upon State-of-the-Art (SOTA) techniques in solar cell mapping, this work puts forth S3Former with the following enhancements: 1. Contrary to popular implementations for PV segmentation, S3Former eliminates the need for a classifier network and focuses on learning strong representations at the segmentation network. 2. S3Former introduces a pioneering Transformer-based architecture featuring a Mask-Attention Transformer Decoder. The novel attention mechanism adeptly captures relationships within high-resolution features, accurately identifying minute solar cells and excelling at contextual understanding around PV installations to prevent mischaracterization of background elements with similar features. 3. Adds a Self-Supervised component to enhance feature extraction at the backbone level and create stronger representations for the down-stream segmentation task. Validation of the proposed method is undertaken through extensive experiments on three annotated datasets—one from California and two from France. This diverse set of backgrounds and PV characteristics ensures the robustness of our method in addressing solar PV segmentation challenges. Benchmarking S3Former against current SOTA methods and popular networks for semantic segmentation reveals superior performance across widespread metrics. In conclusion, this work presents a new pathway for accurately mapping current solar installations, contributing to a deeper understanding of solar energy extension. We hope that the methods and processes described in this work contribute to reducing the impact of PV installations on the grid and, ultimately, create a pipeline for automatically detecting solar cells in the future.

Download

Share

COinS