Continuous UV level estimation using photochromic PLA surfaces: A transformer-based heterogeneous deep ensemble approach

This study presents a high-precision, continuous machine vision system for estimating ultraviolet (UV) radiation levels using low-cost, 3D-printable polylactic acid-based photochromic material (PLA-PCM) surfaces. Filament color transitions under different UV intensities were recorded at eight ambient color temperatures, and a zero-padding spatial arrangement protocol was applied to the 252×56 images that underwent color changes to extract stable temporal features. The methodology employs a deep transformer encoder (DTE) with multi-head self-attention (MSA), computing global temporal dependencies in parallel within 222 milliseconds and ensuring edge-AI compatibility. Multi-spectral feature vectors from the RGB, HSV, and CIELAB color spaces, augmented with ambient color temperature and DCT components, were constructed for 59,000 frames derived from 236 video sessions at 10 Hz. A strict zero-data-leakage protocol was enforced through video-level holdout and stratified temporal splits. Each fold equally represented the fast-onset and slow-decay phases of the photochromic reaction within a five-fold cross-validation. The resulting transformer-based heterogeneous deep ensemble (T-HDE), integrating Transformer, SVR, CatBoost, LightGBM, ExtraTrees, and MLP regressors with Nelder-Mead optimized weights, achieved R 2 = 0.9938 on the held-out test set with a homoscedastic error distribution, indicating that prediction accuracy remains equally reliable across low and high UV exposure levels. In single-frame snapshot mode, R 2 = 0.9811 was attained, confirming that even without temporal sequence information, the zero-padding-based spatial features alone provide robust UV estimation. These findings demonstrate that PLA-PCM surfaces, coupled with the T-HDE architecture, can serve as precise, continuous, and cost-effective UV dosimeters for IoT-based environmental monitoring and edge-AI sensing platforms.