Improving Solar Panel Defect Detection in Thermal Imagery Through Temperature Data Integration

Thermal drone inspections have become a practical solution for monitoring large photovoltaic plants, but defect detection in infrared imagery remains challenging because apparent anomalies can be influenced by shadows, vegetation, reflections, background structures, and the limitations of color-based thermal visualization. In this work, we propose a defect detection approach for photovoltaic modules that integrates real temperature metadata directly into a Faster R-CNN detector as a fourth input channel. Instead of relying only on the conventional thermal rendering produced by the camera, the model is trained with both the visual thermal image and a temperature-derived representation constructed from the corresponding per-pixel radiometric values. The method was evaluated on a dataset collected from real inspections performed across heterogeneous environments, comprising 16,052 images and 39,595 annotated defects distributed across four classes: spot, diode, multiple spots, and open circuit. Under the same data volume and training process, the temperature-augmented model outperformed the standard model, improving macro-averaged precision from 0.832 to 0.874, recall from 0.871 to 0.903, and F1-score from 0.851 to 0.887. Qualitative comparisons further show improved detection of subtle defects and a reduction in false positives caused by thermally misleading background patterns. These results support the value of incorporating physically meaningful thermal information into deep-learning pipelines for photovoltaic defect inspection.