CyCoRS: a self-supervised representation learning method for radar sounder data analysis

As a pivotal element in the energy transition, distributed photovoltaic (PV) systems play a crucial role in resource assessment, grid scheduling, and long-term safeguarding of the national environment. Consequently, their precise identification is of utmost importance. Remote sensing technology has emerged as an indispensable approach for identifying distributed PV systems, primarily due to its advantages in wide coverage, cost-effectiveness, and periodic monitoring capabilities. However, existing methodologies for identifying distributed PV systems remain susceptible to interference from complex backgrounds, compromising identification accuracy. Moreover, the compact spatial scale of distributed PV panels introduces additional challenges to achieving precise recognition in remote sensing imagery. To address these challenges, this paper proposes High - Resolution Spectral Feature Fusion Network (HR-SFNet), a novel model built upon HRNet, by incorporating a spectral feature extraction module to enable effective fusion of spatial and spectral features. This paper validates the HR-SFNet model using Bing Satellite imagery, with selected villages in Jiaozuo City designated as the study area. The experimental results demonstrate that HR-SFNet achieves an Intersection over Union (IoU) of 84.43% and an F1 score of 91.46%, representing improvements of 1.31% and 0.81% respectively compared to HRNet. The model effectively addresses the challenge of PV panel detection being susceptible to complex background interference, enhancing the accuracy of identifying small-target PV panels in remote sensing imagery. These results validate the practical utility of HR-SFNet for distributed PV identification and underscore its significance in promoting the high-quality development of renewable energy applications.

The increasing number of operating radar sounder instruments requires automatic techniques to analyze the radargrams they acquire. Radargrams contain information on the position, composition and extent of subsurface features, that is of utmost importance to monitor the geological structure of planet shallow subsurface. The analysis of radargrams is often carried on manually or with classical statistical methods, which are usually tuned during inference and require the usage of hand-crafted features. Deep learning techniques have recently been introduced for radargram analysis, driven by advances in computational power and the increasing availability of large datasets from multiple instruments. Despite this, the deep-learning based techniques presented so far in the literature are tied to specific task, such as layer tracking, semantic segmentation or subsurface lake detection. In this work, we propose CyCoRS, a self-supervised representation learning method for radar sounder data. CyCoRS encodes radargrams into meaningful features, which can in turn be employed on a variety of data analysis tasks. The method is based on Markov chains transitions casted on each sample of radargram rangelines. These transitions are required to follow horizontal patterns and return to their initial point. We learn the parameters of a pixel-wise feature encoder by optimizing a cycle-consistency objective which does not require labeled data during training. In tests we show that, upon training, our features are reusable on both semantic segmentation and layer tracking. We prove the effectiveness of the learned representation by label propagation (i.e. zero-shot learning) performed with ad-hoc developed algorithms.