Semi-Supervised Techniques for Solar Panel Segmentation in Aerial Images

There has been an ever-growing awareness in recent years about climate change, a phenomenon which is caused primarily by the massive use of fossil fuels by humans. One of the main solutions to this issue is to switch to low-carbon and renewable forms of energy, which include solar energy. To promote the use of photovoltaic panels, it is useful to have an extensive and updated database of installed panels and plants, which could help with performance evaluations, but the availability of such maps is scarce. A solution to this problem is to employ artificial intelligence systems that automatically detect panels in aerial images, specifically by detecting, segmenting and classifying every panel in a given image. This is the objective of Instance Segmentation, one of the most frequently-tackled tasks in the field of Computer Vision. However, the effective training of a Deep Learning model is at times impeded by the scarcity of data: it is expensive and time-consuming to gather enough annotated data to form a reasonably large dataset. This is the case for the dataset used for this thesis, which does not have enough annotations to obtain satisfying results. It therefore becomes necessary to employ methods to artificially increase the size of the dataset, such as semi-supervised learning and Unsupervised Domain Adaptation. The former allows an algorithm to generate annotations by itself on a set of unlabeled data after training on a small portion of labeled data, while the latter consists in taking an algorithm that has satisfying performances on a given domain and apply it to a different but related domain where data is scarce. The goal of this thesis is therefore to evaluate the efficacy of semi-supervised methods for a model trained with the purpose of detecting and segmenting solar panels in aerial images. To this end, two such methods were explored and adapted for the task at hand. The first is Noisy Boundaries, a framework for semi-supervised instance segmentation which provides additional components to resist and exploit the noise inherent in the boundaries of artificially-generated annotations. The second is the union of DACS, a framework that performs Unsupervised Domain Adaptation, with Instance Mixing techniques, with the end goal of increasing the size of the dataset by copying and pasting groups of instances from labeled images to unlabeled images, along with the automatic generation of labels on the latter. Said techniques are evaluated and compared, highlighting their differences and demonstrating their effectiveness in contexts with scarce annotations.