Keywords: Artificial intelligence, Vision Transformers, Sustainability, Remote sensing
Abstract: Renewable energy such as solar is key to ensuring access to affordable and sustainable energy generation. Surveying its adoption patterns globally is pivotal to measuring and evaluating renewable energy access and creating a more efficient and equitable grid. Leveraging high-resolution imagery to detect solar PVs has proven to be a more exhaustive way of covering all PVs, including residential PVs that can be very challenging to track via conventional surveying methods. While the literature has developed models to classify and segment PV installations, residential PV is still challenging to identify using medium resolution ($\approx$ 30 cm/pixel or above) remote-sensing products. This work explores different fine-tuning (FT) strategies of pre-trained ViT models for classification tasks in smaller dataset settings. While FT offers an opportunity for fast and computationally efficient model deployment, practitioners have to be cautious about the effects of fine-tuning on OOD classification and how advances in text attention mechanisms do not necessarily map to image architectures. Moreover, the LoRA technique (Low-Rank Adaptation) is identified as an efficient method for fine-tuning, enhancing the model's adaptability to specific tasks while preserving its generalizability. Despite these advancements, achieving robust OOD classification in a foundational model context remains a challenging task.
Submission Number: 26
Loading