UAV-Assisted Solar Panel Condition Monitoring Under Domain Shift: Comparative Evaluation of Classical ML, CNN, and Vision Transformer Models

This study compares five classical machine-learning classifiers, five convolutional neural network (CNN) architectures, and a Vision Transformer (ViT-Base) for binary Clean/Dirty classification of solar-panel images. A source-group split assigned 842 original images to 589 development and 253 untouched internal-test images. External tests used 704 real unmanned aerial vehicle (UAV) crops (609 Clean, 95 Dirty) from one site and flight and a separate 527-image Kaggle binary subset. Synthetic dust, sand, and clonecrop subsets were analyzed separately. With geometric augmentation, seed-42 internal-test Matthews correlation coefficient (MCC) was 0.843 for the support vector classifier (SVC), 0.992 for DenseNet121, and 0.976 for ViT-Base. Across three seeds, UAV MCC was 0.170 ± 0.032 for DenseNet121 and −0.007 ± 0.012 for ViT-Base at threshold 0.5. ViT-Base retained ranking ability, with receiver operating characteristic area under the curve (ROC-AUC) of 0.664 ± 0.029. Values following ± denote sample standard deviations across seeds; UAV bootstrap intervals were calculated separately for MCC and ROC-AUC. A controlled DenseNet121 experiment examined ImageNet pre-training and geometric augmentation. Central processing unit (CPU) benchmarks and an offline video demonstration assessed crop classification without establishing onboard real-time operation. The results show limited transfer to UAV imagery, with uncertainty conditional on the observed acquisition and grouping.

Authors

Institutions

Publication Details

Journal
Electronics
Published
2026-11-19
DOI
https://doi.org/10.3390/electronics15225401
Primary Topic
Industrial Vision Systems and Defect Detection
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

UAV-Assisted Solar Panel Condition Monitoring Under Domain Shift: Comparative Evaluation of Classical ML, CNN, and Vision Transformer Models

Ivana Berković, Milica Mazalica, Dragan Ivetić, Dalibor Dobrilović et al.
Electronics
Industrial Vision Systems and Defect Detection
article

UAV-Assisted Solar Panel Condition Monitoring Under Domain Shift: Comparative Evaluation of Classical ML, CNN, and Vision Transformer Models

Ivana Berković, Milica Mazalica, Dragan Ivetić, Dalibor Dobrilović, Aleksandra Stojkov, Sinisa Mihajlovic
article en

Abstract

This study compares five classical machine-learning classifiers, five convolutional neural network (CNN) architectures, and a Vision Transformer (ViT-Base) for binary Clean/Dirty classification of solar-panel images. A source-group split assigned 842 original images to 589 development and 253 untouched internal-test images. External tests used 704 real unmanned aerial vehicle (UAV) crops (609 Clean, 95 Dirty) from one site and flight and a separate 527-image Kaggle binary subset. Synthetic dust, sand, and clonecrop subsets were analyzed separately. With geometric augmentation, seed-42 internal-test Matthews correlation coefficient (MCC) was 0.843 for the support vector classifier (SVC), 0.992 for DenseNet121, and 0.976 for ViT-Base. Across three seeds, UAV MCC was 0.170 ± 0.032 for DenseNet121 and −0.007 ± 0.012 for ViT-Base at threshold 0.5. ViT-Base retained ranking ability, with receiver operating characteristic area under the curve (ROC-AUC) of 0.664 ± 0.029. Values following ± denote sample standard deviations across seeds; UAV bootstrap intervals were calculated separately for MCC and ROC-AUC. A controlled DenseNet121 experiment examined ImageNet pre-training and geometric augmentation. Central processing unit (CPU) benchmarks and an offline video demonstration assessed crop classification without establishing onboard real-time operation. The results show limited transfer to UAV imagery, with uncertainty conditional on the observed acquisition and grouping.

ElectronicsVol. 15(22)
University of Novi Sad (RS)
Industrial Vision Systems and Defect Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon