Dataset of UAV-acquired thermal and visible images for solar photovoltaic hotspot detection using MLX90640BAB and raspberry Pi 5.

Hotspots are among the most common and damaging thermal anomalies in solar photovoltaic (PV) modules, indicating localized cell defects, shading or soiling that reduce energy yield and can shorten module lifespan [1,2]. However, publicly available thermal image datasets for PV hotspot detection remain scarce, particularly datasets acquired with low-cost thermal sensors in South Asian operating conditions. This article presents a dataset of UAV acquired thermal fusion and visible images of operating rooftop PV modules collected in Dhaka, Bangladesh. Fifty raw samples were captured across two sites: 25 at Mirpur on 06 June 2026 and 25 at Mohammadpur on 22 June 2026. Images were acquired using a custom built quadcopter carrying a Raspberry Pi 5 interfaced with an MLX90640BAB far infrared thermal array sensor which produced color thermal fusion images with on screen maximum temperature readouts and a Raspberry Pi Camera NoIR V2 for visible imagery. The 50 raw samples were expanded to a 5000 image dataset through pre-processing, data augmentation and annotated with bounding boxes for a single Hotspot class (4,626 labeled instances in the training split). A YOLO26s detector trained on the dataset for 95 epochs on an NVIDIA A100 GPU achieved a precision of 0.947, recall of 0.889, mAP@0.5 of 0.936 and mAP@0.5:0.95 of 0.878 on the validation split which confirms that the dataset supports supervised hotspot detection research. The dataset is suitable for benchmarking object detection algorithms by developing low cost aerial PV inspection pipelines and training edge deployable models for automated PV monitoring.