Machine learning-based snow coverage estimation on PV panels under extreme weather conditions

Document Type

Conference Proceeding

Publication Date

6-11-2026

Department

Department of Computer Science; Department of Applied Computing

Abstract

Accurate estimation of snow coverage on photovoltaic (PV) panels is vital for predicting energy losses and scheduling timely snow-shedding operations. Snow accumulation acts as an opaque barrier that blocks sunlight from reaching the PV cells, leading to substantial power losses that can range from 38.9% to 93.2% depending on snow depth, with overall energy losses reaching up to 34% due to snow removal delays. These figures underscore the importance of developing reliable, automated systems that can continuously monitor snow buildup to minimize downtime and maintain energy efficiency, especially in northern regions experiencing prolonged winter conditions. However, fixed cameras used for continuous PV monitoring often capture degraded images at night or during heavy snowfall, resulting in poor illumination and low contrast, which reduces the reliability of machine learning (ML) models. Building upon our previous UAV-based real-time monitoring framework, which achieved a 1.1% mean error in snow coverage estimation using an optimized combination of YOLOv11n-seg and Otsu’s dynamic thresholding, this study extends our investigation to fixed-camera imagery captured over long winter periods. While the prior system effectively mitigated variable illumination in aerial imagery, the fixed-camera configuration introduces new challenges, including persistent low light, specular reflections, and uneven exposure caused by snowflakes and lens fogging. To address these limitations, we propose an image enhancement pipeline that integrates Contrast-Limited Adaptive Histogram Equalization (CLAHE) with unsharp masking, designed to improve local contrast, highlight snow–panel boundaries, and enhance texture details critical for downstream segmentation algorithms. Technically, the enhancement process increases inter-class separability in the intensity histogram, a key factor that improves the performance of Otsu’s dynamic thresholding, as the most effective method for robust snow segmentation under variable illumination. Preliminary evaluation of the enhanced images shows measurable improvements in perceptual and structural metrics, including a 13% reduction in PIQE, 10.7% increase in entropy, and 21.5% improvement in Otsu separability, confirming a substantial gain in information content and discriminative contrast. These improvements translated into substantially better snow coverage estimation, reducing the average Snow Coverage Percentage (SCP) error under nighttime conditions from 14.99% before enhancement to 2.75% after enhancement and improving the robustness of the pipeline under adverse conditions.

Publication Title

Proc. SPIE 14030, Machine Learning from Challenging Data 2026

Link to Full Text

Share

COinS