Solar Photovoltaic Module Detection Based On Multi-sources Hyperspectral Data

Kurzfassung

Solar photovoltaic is a promising and growing resource for green energy generation. Detecting and monitoring the performance of photovoltaic modules is necessary. The use of remote sensing data has the potential to avoid labor- and time-consuming fieldwork. However, current research on PV detection mostly focuses on airborne and spaceborne color images, which require high spatial resolution i.e., about 30 cm per pixel, and can lead to confusion with the module types due to similar structure (e.g., solar thermals) or color (e.g., dark PV arrays on dark roofs). Hyperspectral images allow the exploitation of the material-specific spectral characteristics of PV modules to identify and separate them from spectrally similar surfaces. The city of Oldenburg in Germany was selected as the study area because of the presence of various PV installations, including large PV power plants and small PV modules, etc. In order to train a classifier, a set of laboratory measurements of two PV samples and other similar material samples with a goniometer setting of 60 spectra per sample and a large library containing HyMap spectra of 31 urban surface material classes (5627 spectra samples) were implemented. In addition, the airborne HySpex data over the city of Oldenburg were used, which were acquired in July 2018 and pre-processed to a spatial resolution of 1.2 m. In this study, we defined and developed a combination of spectral indices for PV materials based on laboratory measurements of PV modules and a large urban surface material spectral library, and then applied the spectral indices to airborne HySpex data. The thresholds for each spectral index were first set manually based on the combined spectral library of the multiple sensors and data acquisition platforms. To obtain a more robust and adaptive threshold derivation approach, the large combined spectral library is used to train a machine learning model to determine the adaptive thresholds for each spectral index. Therefore, we detect PV materials from airborne HySpex data with the (1) manually defined thresholds and (2) the thresholds derived by the machine learning model. Finally, the PV detection results of both versions were compared and discussed. The results show that both approaches can provide accurate PV area classifications. Specifically, the combination of spectral indices is able to accurately detect PV modules in different arrangements and within different environments, without requiring explicit training samples for each setting, but purely based on their spectral characteristics. The spectral indices derived from robust spectral features can effectively identify PV areas, and a clear advantage of the presented approach is that it works well even in the absence of large amounts of labeled training data. The machine learning model overcomes the need for local training data to set thresholds for spectral indices and therefore aims to provide an adaptive and extensible PV spectral library. For the detection of new types of PV modules (e.g. thin-film PV), the features for new PV materials must be included. The proposed approach aims at providing an applicable, robust, and transferable method for PV module detection using hyperspectral data by exploring the physical-based robust spectral features of PV modules combined with machine learning techniques to overcome the manual definition of thresholds. The benefits are not limited to detecting the areas and sizes of PV modules, but may also enable the detection of surface dust thickness and defection of PV modules in the near future. Our goal is to create greater awareness of the potential and applicability of airborne and spaceborne imaging spectroscopy data for PV module identification and monitoring. Moreover, this work is intended to enrich the current spectral library for urban surface materials, with particular attention to specific materials such as PV modules, and to make it adaptable and expandable.