Abstract:

The separation of small-scale residential solar panels (RSPs) through the utilisation of satellite imagery has emerged as a significant challenge in the field of data science within the renewable energy industry. The purpose of this paper is to present a cross-learning-driven approach, as well as its extension, adaptive Cross Nets, which is intended for the automatic segmentation of RSPs in satellite imagery. For the purpose of improving the accuracy of RSP segmentation, the proposed methods make use of a collection of generic U-Nets that collaborate with one another. Initially, each generic U-Net that is a part of the Cross Nets ensemble is independently initialised by employing either transfer learning or traditional initialisation techniques. All of these techniques are used. Following that, we put into action a novel training strategy known as cross learning, which serves as a constraint in order to optimise the Cross Nets in a more efficient manner. During this process, every U-Net is responsible for updating its parameters on an individual basis at each epoch. Subsequently, the parameters of the U-Net that has performed the best at particular epochs are adapted. The dependence of generic U-Nets on precise initialisation is reduced through the use of cross learning, which ultimately results in improved optimisation