Toward More Robust Multiclass Aerial Solar Panel Detection and Classification

Abstract

The challenge of identifying and monitoring multiple types of solar panels has not been studied. Solar panels can be single, double, or double with a water heater on top. Some are packed closely together. Due to installation requirements, additional solar panels may have any random orientation. When combined with the difficulties of detecting different types of panels, this arbitrary orientation negatively affects the effectiveness of deep learning algorithms by resulting to false positive and erroneous panel classifications. Furthermore, no research on the identification of various solar panel types has been done yet. In this study, we concentrate in on two key problems: first, the detection of different types of solar panels; and second, the arbitrary orientation of these panels. Our method does not use horizontal bounding box, rather it leverages horizontal bounding boxes and generates rotated bounding box during the train time. Using our method, we were able to precisely identify three various types of solar panels with various orientations. We show a comparison of their differences for the identification of three different types of solar panels, including water heater photovoltaic (WPV), farm type photovoltaic (FPV), and SPV, in terms of box loss, objectness loss, classification loss, precision, and recall (single photovoltaic).