Segmentation of Satellite Images of Solar Panels Using Fast Deep Learning Model

Authors: Wani, M.A. and Mujtaba, T.

Journal: International Journal of Renewable Energy Research

Volume: 11

Issue: 1

Pages: 31-45

eISSN: 1309-0127

Abstract:

Segmenting satellite images provides an easy and cost-effective solution to detect solar arrays installed on building tops and on ground over a region. Solar panel detection is the first step towards image based estimation of energy generation from the distributed solar arrays connected to a conventional electric grid. Segmentation models for small devices require light weight procedures in terms of computational effort. State-of-the-art deep learning segmentation models have the disadvantage that these require long training times, large number of floating-point operations (FLOPS) and tens of millions of parameters which make these models less suitable for devices with limited computational power. This paper proposes a deep learning segmentation architecture that is suitable for small devices. The proposed architecture combines features of Mobilenet classification architecture and Unet architecture in such a way such that it is efficient in terms of computational effort and produces segmentation results with good accuracy. The results of the proposed model are compared with the results obtained by various state-of-the-art segmentation models. The results demonstrate that the proposed model is computationally efficient as it requires less number of model parameters, less training time, consumes less number of FLOPS and produces good segmentation results with competitive accuracy.

Source: Scopus