Table 1.

Comparison between the proposed method and other research identifying solar panels using only thermal images with a complex background.

Method	Description	Weaknesses
Dhimish et al. [34]	Identifies hot spots in an RGB image with a color ramp.	Does not identify the affected solar panel. The thermal images have only a panel in high resolution.
Libra et al. [35]	Identifies hot spots in an RGB image with a color ramp.	Does not identify the affected solar panel. Does not segment the image
Liao et al. [31]	Identifies hot spots in an RGB image with a color ramp. Classic method based on filters for a binary classification of the image between faulty and non-faulty areas.	Does not identify the affected solar panel. Assumes that in all the photo there are only panels Lacks of methods to classify segments.
Alsafasfeh et al. [26]	Segmentation based on hot pixels detection. Classic method based on Canny edge detection	Does not identify the affected solar panel. Lacks of methods to classify segments.
Addabbo et al. [32]	Panel detection with classic methods based on template matching using normalized cross-correlation Tested on large dataset	The thermal images present a few panels in high resolution. It does not report or present solutions for the feautures 1, 4, and 5 of the complex background.
Alfaro-Mejía et al. [28]	Classic method based on two techniques. Performs an image transformation to orthogonize the detected panel	The thermal images present a few panels in high resolution. It does not present solutions for any of the complex backgrounds
Uma et al. [33]	Classic method with segmentation of the image using the k-means clustering algorithm.	k-means is an unsupervised classification. It does not present solutions for the feautures 3, 4, and 5 of the complex background.
Zhu et al. [30]	learning with an algorithm based on a fully convolutional neural network and a dense conditional random field	It does not propose a solution to identify panels that remained undetected by the deep learning method.
Greco et al. [29]	Deep learning with a convolutional neural network framework called ’You only Look Once’ (YOLO) Tested on large dataset	It does not propose a solution to identify panels that remained undetected by the deep learning method.