Detection and mapping of solar panels based on Deep learning, instance segmentation and very high-resolution multiband unmanned aerial vehicle (UAV) photogrammetric survey.

Abstract

Solar energy is becoming a pivotal resource in the field of electricity generation as it seeks to create clean energy which makes the earth environmentally friendly. Unmanned Aerial Vehicle (UAV) on the other hand has become a low-cost data collection technology used for remote sensing in Geoscience. With 300 days of sunshine in the Kingdom of Spain and 3,321 hours of annual sunshine in Castellon, residents are eager to harness the solar energy at their disposal through the mass installation of solar panels. In light of this, there is an urgent need for studies on the accurate detection of solar panels with minimal edge loss and energy quantification of these solar panels in order to create a spatial data infrastructure that serves as a geospatial tool for urban planning and green policy implementation. A case in point is the spatial identification of solar clusters which can aid in the location of EV charging stations in urban areas. To this end, this thesis addresses these challenges by conducting a UAV photogrammetric survey and developing a deep learning model for the accurate detection, mapping, and quantification of solar panels in Castellon, Spain using ESRI ArcGIS Pro and Drone2Map. To ensure accurate detection and minimal detection edge loss, we make use of instance segmentation which combines the use of object detection and semantic segmentation. This method accurately delineates the boundary of each solar panel in the study area from a very high-resolution UAV multiband photogrammetry survey while minimising false detection through the addition of a normalised digital surface model. The deep learning model was trained on a record high 0.03m spatial resolution RGBnDSM UAV imagery with the state-of-the-art instance segmentation deep learning architecture called Mask RCNN on the ResNet-101 backbone in diverse weather conditions. This research was tested on real world scenarios, and we achieved a mean accuracy of 0.8445, a recall of 0.9162 and an F1 score of 0.8782 where a higher intersection over union was set at 0.75. Following these results, the deep learning model was applied to an urban planning use case to determine spatial distribution of solar clusters. The findings of this research contribute to advancing solar energy integration into urban landscapes with a robust and accurate geospatial framework.