In recent years the wide availability of high-resolution satellite images has made the remote monitoring of water resources all over the world possible. While the detection of open water from satellite images is relatively easy, a significant percentage of the water extent of wetlands is covered by vegetation. Convolutional Neural Networks have shown great success in the task of detecting wetlands in satellite images. However, these models require large amounts of manually annotated satellite images, which are slow and expensive to produce. In this paper we use self-supervised training methods to train a Convolutional Neural Network to detect water from satellite images without the use of annotated data. We use a combination of deep clustering and negative sampling based on the paper ”Unsupervised Single-Scene Semantic Segmentation for Earth Observation”, and we expand the paper by changing the clustering loss, the model architecture and implementing an ensemble model. Our final ensemble of self-supervised models outperforms a single supervised model, showing the power of self-supervision. / Under de senaste åren har den breda tillgången på högupplösta satellitbilder möjliggjort fjärrövervakning av vattenresurser över hela världen. Även om det är relativt enkelt att upptäcka öppet vatten från satellitbilder, täcks en betydande andel av våtmarkernas vattenutbredning av vegetation. Lyckligtvis kan radarsignaler tränga igenom vegetation, vilket gör det möjligt för oss att upptäcka vatten gömt under vegetation från satellitradarbilder. Under de senaste åren har Convolutional Neural Networks visat stor framgång i denna uppgift. Tyvärr kräver dessa modeller stora mängder manuellt annoterade satellitbilder, vilket är långsamt och dyrt att producera. Självövervakad inlärning är ett område inom maskininlärning som syftar till att träna modeller utan användning av annoterade data. I den här artikeln använder vi självövervakad träningsmetoder för att träna en Convolutional Neural Network-baserad modell för att detektera vatten från satellitbilder utan användning av annoterade data. Vi använder en kombination av djup klustring och kontrastivt lärande baserat på artikeln ”Unsupervised Single-Scene Semantic Segmentation for Earth Observation”. Dessutom utökar vi uppsatsen genom att modifiera klustringsförlusten och modellarkitekturen som används. Efter att ha observerat hög varians i våra modellers prestanda implementerade vi också en ensemblevariant av vår modell för att få mer konsekventa resultat. Vår slutliga ensemble av självövervakade modeller överträffar en enda övervakad modell, vilket visar kraften i självövervakning.

Multi-temporal image analysis has been widely used in many applications such as urban monitoring, disaster management, and agriculture. With the development of the remote sensing technology, the new generation remote sensing satellite images with High/ Very High spatial resolution (HR/VHR) are now available. Compared to the traditional low/medium spatial resolution images, the detailed information of ground objects can be clearly analyzed in the HR/VHR images. Classical methods of multi-temporal image analysis deal with the images at pixel level and have worked well on low/medium resolution images. However, they provide sub-optimal results on new generation images due to their limited capability of modeling complex spatial and spectral information in the new generation products. Although significant number of object-based methods have been proposed in the last decade, they depend on suitable segmentation scale for diverse kinds of objects present in each temporal image. Thus their capability to express contextual information is limited. Typical spatial properties of last generation images emphasize the need of having more flexible models for object representation. Another drawback of the traditional methods is the difficulty in transferring knowledge learned from one specific problem to another. In the last few years, an interesting development is observed in the machine learning/computer vision field. Deep learning, especially Convolution Neural Networks (CNNs) have shown excellent capability to capture object level information and in transfer learning. By 2015, deep learning achieved state-of-the-art performance in most computer vision tasks. Inspite of its success in computer vision fields, the application of deep learning in multi-temporal image analysis saw slow progress due to the requirement of large labeled datasets to train deep learning models. However, by the start of this PhD activity, few works in the computer vision literature showed that deep learning possesses capability of transfer learning and training without labeled data. Thus, inspired by the success of deep learning, this thesis focuses on developing deep learning based methods for unsupervised/semi-supervised multi-temporal image analysis. This thesis is aimed towards developing methods that combine the benefits of deep learning with the traditional methods of multi-temporal image analysis. Towards this direction, the thesis first explores the research challenges that incorporates deep learning into the popular unsupervised change detection (CD) method - Change Vector Analysis (CVA) and further investigates the possibility of using deep learning for multi-temporal information extraction. The thesis specifically: i) extends the paradigm of unsupervised CVA to novel Deep CVA (DCVA) by using a pre-trained network as deep feature extractor; ii) extends DCVA by exploiting Generative Adversarial Network (GAN) to remove necessity of having a pre-trained deep network; iii) revisits the problem of semi-supervised CD by exploiting Graph Convolutional Network (GCN) for label propagation from the labeled pixels to the unlabeled ones; and iv) extends the problem statement of semantic segmentation to multi-temporal domain via unsupervised deep clustering. The effectiveness of the proposed novel approaches and related techniques is demonstrated on several experiments involving passive VHR (including Pleiades), passive HR (Sentinel-2), and active VHR (COSMO-SkyMed) datasets. A substantial improvement is observed over the state-of-the-art shallow methods.