Unsupervised Modular Adaptive Region Growing and RegionMix Classification for Wind Turbine Segmentation

TL;DR

This work tackles automatic wind turbine blade segmentation under limited data by reframing pixel-wise segmentation as region-level binary classification. It introduces Modular Adaptive Region-Growing (MARG), which integrates Dual-Threshold Region-Growing with Adaptive Thresholding and Region Merging, and couples it with RegionMix data augmentation to synthesize diverse region configurations. The method demonstrates state-of-the-art segmentation accuracy and strong cross-site generalization across windfarms, while remaining interpretable and CPU-friendly. The results suggest that unsupervised, region-based strategies can achieve robust blade delineation with limited labeled data, enabling scalable maintenance analytics in wind farms.

Abstract

Reliable operation of wind turbines requires frequent inspections, as even minor surface damages can degrade aerodynamic performance, reduce energy output, and accelerate blade wear. Central to automating these inspections is the accurate segmentation of turbine blades from visual data. This task is traditionally addressed through dense, pixel-wise deep learning models. However, such methods demand extensive annotated datasets, posing scalability challenges. In this work, we introduce an annotation-efficient segmentation approach that reframes the pixel-level task into a binary region classification problem. Image regions are generated using a fully unsupervised, interpretable Modular Adaptive Region Growing technique, guided by image-specific Adaptive Thresholding and enhanced by a Region Merging process that consolidates fragmented areas into coherent segments. To improve generalization and classification robustness, we introduce RegionMix, an augmentation strategy that synthesizes new training samples by combining distinct regions. Our framework demonstrates state-of-the-art segmentation accuracy and strong cross-site generalization by consistently segmenting turbine blades across distinct windfarms.

Figures (18)

• Figure 1: Schematic overview of our proposed framework, refer to \ref{['fig:marg-class']} for details. The wind turbine blade is segmented by first identifying salient regions and then, classifying these regions into blade or background. Regions obtained by our MARG are depicted with different colors.
• Figure 2: Seed selection process.Left: equidistant seed candidate pixels in red. Middle: window $\aleph_{h,w}^{(2)}$ around a particular seed pixel $\mathbf{c}_{h,w}$. Right: Sobel's output.
• Figure 3: Cartesian Neighbors vs. Modular Neighbors. Image $\mathbf{X}$, DTRG with cartesian neighbors, and DTRG with modular ones (DTMRG), respectively. Regions are shown with different colors.
• Figure 4: Distinct levels of coverage based on selected region-growing thresholds $\tau^{s}$ and $\tau^{l}$.Left: Original image $\mathbf{X}$; Middle: Very high $\tau^{s}$ and $\tau^{l}$, $\Pi = 100\%$ but it generates over-grown regions; Right: Very stringent $\tau^{s}$ and $\tau^{l}$, leading to low coverage $\Pi$; under-segmentation. Regions are depicted with different colors.
• Figure 5: Comparative analysis of segmentation performance metrics across varying $\tau^{s}$ values, differentiated by $\tau^{l}$ levels. The golden star indicates $\tau^{s*}$, the blue one $\tau^{l*}$, while the purple star indicates the couple of thresholds $\tau^{s*}$,$\tau^{l*}$.