Weakly Supervised Deep Learning Scheme for Solar Panel Mapping Using Remote Sensing Images

Jue Zhang

Abstract: With the availability of high-resolution aerial images, automatic solar panel mapping (SPM) is feasible with machine learning algorithms. It enables consistent and frequent monitoring of solar panel installations at rooftops in the residential areas, which serves as the foundation step for subsequent social-economic analysis. The introduction of weakly supervised learning makes machine learning low-cost by reducing the demand for manual annotation of training samples and generating labeled samples automatically, which are called pseudo labels (PLs). PLs, however, are less reliable and contain errors. This thesis is to address these issues in the weakly supervised approach for SPM. To mitigate the negative impacts caused by PLs’ inconsistent quality, a Self-Paced weakly supervised learning method with Residual Aggregated Network (SP-RAN) is proposed. With the confidence-aware loss and self-paced label correction strategy, SP-RAN adaptively enhances the contribution of high-quality PLs to the model training and minimizes the negative impacts of the bad-quality ones. By formulating the weakly supervised SPM task as a label noise problem, an Uncertainty Adjusted Label Transition (UALT) method is developed. An uncertainty-adjusted re-weighting strategy and a trace regularizer are proposed to reduce the complexity as well as improve accuracy in the estimation of the instance-dependent transition matrix. To take advantage of complementary PLs, a complementary joint learning framework based on Attention-based Dual Stream Network (ADSNet) is proposed. To synthesize complementary PLs, ADSNet consists of a target mapping branch to produce mapping results with sharp boundaries and a proximity detection branch for complete localization. Given the unavoidable inclusion of proximity regions in joint training, the subspace contrast loss is proposed to enhance feature discrimination between the targets and their proximity. The aerial image datasets collected from Google Static Maps and the Geocentric Datum of Australia 2020 project were utilized to validate the proposed methods. Comprehensive comparison with the state-of-the-art methods and ablation studies demonstrates the stability and superiority of the proposed solutions. This indicates the feasibility of applying weakly supervised SPM for the automatic monitoring of solar panel installations.