The thesis focuses on machine learning methods for Earth Observation (EO) data, more specifically, remote sensing data acquired by satellites and drones. EO plays a vital role in monitoring the Earth’s surface and modelling climate change to take necessary precautionary measures. Initially, these efforts were dominated by methods relying on handcrafted features and expert knowledge. The recent advances of machine learning methods, however, have also led to successful applications in EO. This thesis explores supervised and unsupervised approaches of Deep Learning (DL) to monitor natural resources of water bodies and forests. 

The first study of this thesis introduces an Unsupervised Curriculum Learning (UCL) method based on widely-used DL models to classify water resources from RGB remote sensing imagery. In traditional settings, human experts labeled images to train the deep models which is costly and time-consuming. UCL, instead, can learn the features progressively in an unsupervised fashion from the data, reducing the exhausting efforts of labeling. Three datasets of varying resolution are used to evaluate UCL and show its effectiveness: SAT-6, EuroSAT, and PakSAT. UCL outperforms the supervised methods in domain adaptation, which demonstrates the effectiveness of the proposed algorithm. 

The subsequent study is an extension of UCL for the multispectral imagery of Australian wildfires. This study has used multispectral Sentinel-2 imagery to create the dataset for the forest fires ravaging Australia in late 2019 and early 2020. 12 out of the 13 spectral bands of Sentinel-2 are concatenated in a way to make them suitable as a three-channel input to the unsupervised architecture. The unsupervised model then classified the patches as either burnt or not burnt. This work attains 87% F1-Score mapping the burnt regions of Australia, demonstrating the effectiveness of the proposed method. 

The main contributions of this work are (i) the creation of two datasets using Sentinel-2 Imagery, PakSAT dataset and Australian Forest Fire dataset; (ii) the introduction of UCL that learns the features progressively without the need of labelled data; and (iii) experimentation on relevant datasets for water body and forest fire classification. 

This work focuses on patch-level classification which could in future be expanded to pixel-based classification. Moreover, the methods proposed in this study can be extended to the multi-class classification of aerial imagery. Further possible future directions include the combination of geo-referenced meteorological and remotely sensed image data to explore proposed methods. Lastly, the proposed method can also be adapted to other domains involving multi-spectral and multi-modal input, such as, historical documents analysis, forgery detection in documents, and Natural Language Processing (NLP) classification tasks.

Earth Observation (EO) data is crucial for understanding, managing, and conserving our planet's ecosystem and its natural resources. This data enables humanity to monitor environmental changes, such as natural disasters, urban growth, and climate shifts, assisting informed decisions and proactive measures. Early EO heavily relied on statistical methods and expert domain knowledge, but the advent of machine learning has revolutionized EO data processing, enhancing efficiency and accuracy. Conventional ML models require expensive and labor-intensive data labeling. In contrast, unsupervised ML techniques can learn features from data without the need for manual labeling, making the process more efficient and cost-effective.

This thesis presents a UCL approach utilizing advanced DL models to classify EO data, referred to as UCL4EO. This approach eliminates the need for manual data labeling in training the DL model. The UCL framework comprises i) a DL model tailored for feature extraction from image data, ii) a clustering method to group deep features, and iii) a selection operation to capture representative samples from these clusters. The CNN extracts meaningful features from images, subjected to a clustering algorithm to create pseudo-labels. After identifying the initial clusters, representative samples from each cluster are chosen using the UCL selection operation to fine-tune the feature extractor. The stated process is repeated iteratively until convergence. The proposed UCL approach progressively learns and incorporates salient data features in an unsupervised manner by utilizing pseudo-labels.

UCL started as a proof of concept to show the viability of the method for binary classification on RS and aerial imagery. Specifically, the UCL framework is employed to identify water bodies using three RGB datasets, encompassing both low and high-resolution RS and aerial imagery. While UCL has been extensively examined with RGB imagery, it has been adapted to benefit from the enhanced capabilities of multi-spectral satellite imagery. This adaptation enables UCL to generalize to multi-spectral imagery from Sentinel-2 to detect forest fires in Australia. UCL undergoes subsequent improvements and is further investigated to identify utility poles in high-resolution UAV images. These gray-scale images of utility poles pose computer vision challenges, including issues like occlusion and cropping, where a significant portion of the image contains the background and only a slight appearance of the utility pole. Extensive experimentation on the mentioned tasks effectively showcases UCL's adaptive learning capabilities, producing promising results. The achieved accuracy surpassed those of supervised methods in cross-domain adaptation on similar tasks, underscoring the effectiveness of the proposed algorithm.

The scope of UCL has been extended to encompass multi-class classification tasks in the domain of RS data, referred to as Multi-class UCL. Multi-class UCL progressively acquires knowledge about various categories on multi-scale resolution. To investigate Multi-class UCL, we have used four publicly available datasets of Sentinel-2 and aerial imagery: EuroSAT, SAT-6, UCMerced, and RSSCN7. Comprehensive experiments conducted on the above-mentioned datasets revealed better cross-domain adaptation capabilities compared to supervised methods, thereby demonstrating the effectiveness of Multi-class UCL.

In these investigations, two datasets are generated using Sentinel-2 satellite imagery: one for water bodies - PakSAT and the other for Australian forest fires. However, cloud cover poses a significant challenge by obstructing the satellite's ability to capture clear images of the Earth's surface. To address this issue, available cloud masking techniques are employed to filter out images affected by cloud cover, ensuring the datasets contain only clear and usable data. Later, this thesis examines cloud detection and Cloud Optical Thickness (COT) estimation from Sentinel-2 imagery. We employed machine-learning techniques, achieving better performance than SCL designed by ESA for cloud cover tasks.

In addition to the application in RS data, UCL has been investigated in other domains of EO, such as undersea imagery. Furthermore, UCL has also been used for tasks like natural scene classification, medical imaging, and document analysis, demonstrating its versatility and broad applicability. Further exploration of UCL could involve improving the process of generating pseudo-labels through deep learning techniques.