Cloud formations often obscure optical satellite-based monitoring of the Earth’s surface, thus limiting Earth observation (EO) activities such as land cover mapping, ocean color analysis, and cropland monitoring. The integration of machine learning (ML) methods within the remote sensing domain has significantly improved performance for a wide range of EO tasks, including cloud detection and filtering, but there is still much room for improvement. A key bottleneck is that ML methods typically depend on large amounts of annotated data for training, which are often difficult to come by in EO contexts. This is especially true when it comes to cloud optical thickness (COT) estimation. A reliable estimation of COT enables more fine-grained and application-dependent control compared to using pre-specified cloud categories, as is common practice. To alleviate the COT data scarcity problem, in this work, we propose a novel synthetic dataset for COT estimation, which we subsequently leverage for obtaining reliable and versatile cloud masks on real data. In our dataset, top-of-atmosphere radiances have been simulated for 12 of the spectral bands of the Multispectral Imagery (MSI) sensor onboard Sentinel-2 platforms. These data points have been simulated under consideration of different cloud types, COTs, and ground surface and atmospheric profiles. Extensive experimentation of training several ML models to predict COT from the measured reflectivity of the spectral bands demonstrates the usefulness of our proposed dataset. In particular, by thresholding COT estimates from our ML models, we show on two satellite image datasets (one that is publicly available, and one which we have collected and annotated) that reliable cloud masks can be obtained. The synthetic data, the newly collected real dataset, code and models have been made publicly available.

Seagrass ecosystems are pivotal in marine environments, serving as crucial habitats for diverse marine species and contributing significantly to carbon sequestration. Accurate classification of seagrass species from underwater images is imperative for monitoring and preserving these ecosystems. This paper introduces Unsupervised Curriculum Learning (UCL) to seagrass classification using the DeepSeagrass dataset. UCL progressively learns from simpler to more complex examples, enhancing the model's ability to discern seagrass features in a curriculum-driven manner. Experiments employing state-of-the-art deep learning architectures, convolutional neural networks (CNNs), show that UCL achieved overall 90.12 % precision and 89 % recall, which significantly improves classification accuracy and robustness, outperforming some traditional supervised learning approaches like SimCLR, and unsupervised approaches like Zero-shot CLIP. The methodology of UCL involves four main steps: high-dimensional feature extraction, pseudo-label generation through clustering, reliable sample selection, and fine-tuning the model. The iterative UCL framework refines CNN's learning of underwater images, demonstrating superior accuracy, generalization, and adaptability to unseen seagrass and background samples of undersea images. The findings presented in this paper contribute to the advancement of seagrass classification techniques, providing valuable insights into the conservation and management of marine ecosystems. The code and dataset are made publicly available and can be assessed here: https://github.com/nabid69/Unsupervised-Curriculum-Learning—UCL.

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.

The recognition of inner speech, which could give a ‘voice’ to patients that have no ability to speak or move, is a challenge for brain-computer interfaces (BCIs). A shortcoming of the available datasets is that they do not combine modalities to increase the performance of inner speech recognition. Multimodal datasets of brain data enable the fusion of neuroimaging modalities with complimentary properties, such as the high spatial resolution of functional magnetic resonance imaging (fMRI) and the temporal resolution of electroencephalography (EEG), and therefore are promising for decoding inner speech. This paper presents the first publicly available bimodal dataset containing EEG and fMRI data acquired nonsimultaneously during inner-speech production. Data were obtained from four healthy, right-handed participants during an inner-speech task with words in either a social or numerical category. Each of the 8-word stimuli were assessed with 40 trials, resulting in 320 trials in each modality for each participant. The aim of this work is to provide a publicly available bimodal dataset on inner speech, contributing towards speech prostheses.

We conduct relatively extensive investigations of automatic hate speech (HS) detection using different State-of-The-Art (SoTA) baselines across 11 subtasks spanning six different datasets. Our motivation is to determine which of the recent SoTA models is best for automatic hate speech detection and what advantage methods, such as data augmentation and ensemble, may have on the best model, if any. We carry out six cross-task investigations. We achieve new SoTA results on two subtasks—macro F1 scores of 91.73% and 53.21% for subtasks A and B of the HASOC 2020 dataset, surpassing previous SoTA scores of 51.52% and 26.52%, respectively. We achieve near-SoTA results on two others—macro F1 scores of 81.66% for subtask A of the OLID 2019 and 82.54% for subtask A of the HASOC 2021, in comparison to SoTA results of 82.9% and 83.05%, respectively. We perform error analysis and use two eXplainable Artificial Intelligence (XAI) algorithms (Integrated Gradient (IG) and SHapley Additive exPlanations (SHAP)) to reveal how two of the models (Bi-Directional Long Short-Term Memory Network (Bi-LSTM) and Text-to-Text-Transfer Transformer (T5)) make the predictions they do by using examples. Other contributions of this work are: (1) the introduction of a simple, novel mechanism for correcting Out-of-Class (OoC) predictions in T5, (2) a detailed description of the data augmentation methods, and (3) the revelation of the poor data annotations in the HASOC 2021 dataset by using several examples and XAI (buttressing the need for better quality control). We publicly release our model checkpoints and codes to foster transparency.

We investigate the performance of a state-of-the-art (SoTA) architecture T5 (available on the SuperGLUE) and compare it with 3 other previous SoTA architectures across 5 different tasks from 2 relatively diverse datasets. The datasets are diverse in terms of the number and types of tasks they have. To improve performance, we augment the training data by using a new autoregressive conversational AI model checkpoint. We achieve near-SoTA results on a couple of the tasks - macro F1 scores of 81.66% for task A of the OLID 2019 dataset and 82.54% for task A of the hate speech and offensive content (HASOC) 2021 dataset, where SoTA are 82.9% and 83.05%, respectively. We perform error analysis and explain why one of the models (Bi-LSTM) makes the predictions it does by using a publicly available algorithm: Integrated Gradient (IG). This is because explainable artificial intelligence (XAI) is essential for earning the trust of users. The main contributions of this work are the implementation method of T5, which is discussed; the data augmentation, which brought performance improvements; and the revelation on the shortcomings of the HASOC 2021 dataset. The revelation shows the difficulties of poor data annotation by using a small set of examples where the T5 model made the correct predictions, even when the ground truth of the test set were incorrect (in our opinion). We also provide our model checkpoints on the HuggingFace hub¹. https://huggingface.co/sana-ngu/HaT5_augmentation https://huggingface.co/sana-ngu/HaT5.

Building open-domain conversational systems (or chatbots) that produce convincing responses is a recognized challenge. Recent state-of-the-art (SoTA) transformer-based models for the generation of natural language dialogue have demonstrated impressive performance in simulating human-like, single-turn conversations in English.This work investigates, by an empirical study, the potential for transfer learning of such models to Swedish language. DialoGPT, an English language pre-trained model, is adapted by training on three different Swedish language conversational datasets obtained from publicly available sources: Reddit, Familjeliv and the GDC. Perplexity score (an automated intrinsic metric) and surveys by human evaluation were used to assess the performances of the fine-tuned models. We also compare the DialoGPT experiments with an attention-mechanism-based seq2seq baseline model, trained on the GDC dataset. The results indicate that the capacity for transfer learning can be exploited with considerable success. Human evaluators asked to score the simulated dialogues judged over 57% of the chatbot responses to be human-like for the model trained on the largest (Swedish) dataset. The work agrees with the hypothesis that deep monolingual models learn some abstractions which generalize across languages. We contribute the codes, datasets and model checkpoints and host the demos on the HuggingFace platform.

This paper introduces a machine learning-based approach for detecting electric poles, an essential part of power grid maintenance. With the increasing popularity of deep learning, several such approaches have been proposed for electric pole detection. However, most of these approaches are supervised, requiring a large amount of labeled data, which is time-consuming and labor-intensive. Unsupervised deep learning approaches have the potential to overcome the need for huge amounts of training data. This paper presents an unsupervised deep learning framework for utility pole detection. The framework combines Convolutional Neural Network (CNN) and clustering algorithms with a selection operation. The CNN architecture for extracting meaningful features from aerial imagery, a clustering algorithm for generating pseudo labels for the resulting features, and a selection operation to filter out reliable samples to fine-tune the CNN architecture further. The fine-tuned version then replaces the initial CNN model, thus improving the framework, and we iteratively repeat this process so that the model learns the prominent patterns in the data progressively. The presented framework is trained and tested on a small dataset of utility poles provided by “Mention Fuvex” (a Spanish company utilizing long-range drones for power line inspection). Our extensive experimentation demonstrates the progressive learning behavior of the proposed method and results in promising classification scores with significance test having p−value<0.00005 on the utility pole dataset.

Massive wildfires not only in Australia, but also worldwide are burning millions of hectares of forests and green land affecting the social, ecological, and economical situation. Widely used indices-based threshold methods like Normalized Burned Ratio (NBR) require a huge amount of data preprocessing and are specific to the data capturing source. State-of-the-art deep learning models, on the other hand, are supervised and require domain experts knowledge for labeling the data in huge quantity. These limitations make the existing models difficult to be adaptable to new variations in the data and capturing sources. In this work, we have proposed an unsupervised deep learning based architecture to map the burnt regions of forests by learning features progressively. The model considers small patches of satellite imagery and classifies them into burnt and not burnt. These small patches are concatenated into binary masks to segment out the burnt region of the forests. The proposed system is composed of two modules: 1) a state-of-the-art deep learning architecture for feature extraction and 2) a clustering algorithm for the generation of pseudo labels to train the deep learning architecture. The proposed method is capable of learning the features progressively in an unsupervised fashion from the data with pseudo labels, reducing the exhausting efforts of data labeling that requires expert knowledge. We have used the realtime data of Sentinel-2 for training the model and mapping the burnt regions. The obtained F1-Score of 0.87 demonstrates the effectiveness of the proposed model.