Self-Supervised Pretraining for Power Generation via Dynamic Guided Masking

The escalating reliance on fossil fuel power plants remains a critical driver of global greenhouse gas emissions, necessitating precise and scalable monitoring systems for climate change mitigation. Traditional power generation estimation methods rely on bottom-up self reporting methods, which are time-consuming and subjective. These limitations have motivated growing interest in remote sensing based approaches, where satellite observations provide an objective, scalable, and globally consistent alternative. We present a self-supervised learning (SSL) approach tailored for multi-spectral satellite imagery. Unlike traditional methods, our methodology utilizes a dynamic and guided masking pretext task that forces the model to internalize latent features from high-priority spectral bands. Experimental results demonstrate that the proposed framework beats existing benchmarks in regression and segmentation by 7.7\% and 14.1\% respectively, achieving superior accuracy in power generation estimation without requiring auxiliary segmentation or classification labels.