Rapid development of artificial intelligence (AI) systems amplify many concerns in society. These AI algorithms inherit different biases from humans due to mysterious operational flow and because of that it is becoming adverse in usage. As a result, researchers have started to address the issue by investigating deeper in the direction towards Responsible and Explainable AI. Among variety of applications of AI, facial expression recognition might not be the most important one, yet is considered as a valuable part of human-AI interaction. Evolution of facial expression recognition from the feature based methods to deep learning drastically improve quality of such algorithms. This research work aims to study a gender bias in deep learning methods for facial expression recognition by investigating six distinct neural networks, training them, and further analysed on the presence of bias, according to the three definition of fairness. The main outcomes show which models are gender biased, which are not and how gender of subject affects its emotion recognition. More biased neural networks show bigger accuracy gap in emotion recognition between male and female test sets. Furthermore, this trend keeps for true positive and false positive rates. In addition, due to the nature of the research, we can observe which types of emotions are better classified for men and which for women. Since the topic of biases in facial expression recognition is not well studied, a spectrum of continuation of this research is truly extensive, and may comprise detail analysis of state-of-the-art methods, as well as targeting other biases.

Social media continues to grow, permanently capturing our digital footprint in the form of texts, photographs, and videos, thereby reflecting our daily lives. Therefore, recent studies are increasingly recognising passively crowdsourced geotagged photographs retrieved from location-based social media as suitable data for quantitative mapping and assessment of cultural ecosystem service (CES) flow. In this study, we attempt to improve CES mapping from geotagged photographs by combining natural language processing, i.e., topic modelling and automated machine learning classification. Our study focuses on three main groups of CESs that are abundant in outdoor social media data: landscape watching, active outdoor recreation, and wildlife watching. Moreover, by means of a comparative viewshed analysis, we compare the geographic information system- and remote sensing-based landscape organisation metrics related to landscape coherence and colour harmony. We observed the spatial distribution of CESs in Estonia and confirmed that colour harmony indices are more strongly associated with landscape watching and outdoor recreation, while landscape coherence is more associated with wildlife watching. Both CES use and values of landscape organisation indices are land cover-specific. The suggested methodology can significantly improve the state-of-the-art with regard to CES mapping from geotagged photographs, and it is therefore particularly relevant for monitoring landscape sustainability.

No abstract

<p>Integrated use of citizen science (crowdsourcing in general) and remote sensing is essential to comprehend the complexity of the notion of landscape, based on subjective experience and objective structure of environment. Organisation-related landscape attributes, such as landscape diversity and orderliness, as well as the extent of colour harmony, greenness, and transport accessibility, were recently recognised as indicators for visual and recreational values of environment. However, it is currently an open research question, whether mentioned anthropocentric nature-related values are dependable on these landscape attributes, quantifiable with GIS and remote sensing, and accurate mapping of aesthetic and recreational landscape services is important to answer this question. Image hosting services and social networks provide a huge source of evidence on the aesthetic and recreational landscape experience, allowing for mapping the intangible anthropocentric values with publicly shared georeferenced photographs. Therefore, we aimed to apply automated image recognition with Clarifai service to assign each photograph with tags, reflecting its content, and further topic modelling (a variety of textual analysis) to group the tags into the categories.</p><p>In this study, we used combined Flickr and VK.com dataset for 2016-2018 years, collected via official APIs within the territory or Estonia; outdoor photographs were grouped into three classes: aesthetic landscape experience, outdoor recreation activities and wildlife watching. Non-relevant photographs and photographs with repeating content from the same author were excluded from analysis; a dataset of >10000 photographs was finally analysed. Cloud-free summertime Landsat-8 mosaic for 2018 was used to estimate the landscape diversity, orderliness, colour harmony extent, greenness and other metrics, whereas digital elevation model and land use/land cover model were used to map landscape coherence, terrain ruggedness, and indicate transport accessibility. Contrary to previous findings, users of Flickr and VK.com tend to take photographs of lower landscape diversity and lower greenness. We confirm that, according to the photographs being studied, water presence, terrain ruggedness, and transport accessibility are the best indicators of recreational experience. Colour harmony of land cover and landscape coherence are moderately higher for actual outdoor photographs.</p><p>Performance of the mentioned indicators varies among the groups of photographs, wildlife watching is the least predictable class of recreational landscape services. The applicability of remote sensing-based mapping of landscape attributes and textual analysis of tags, extracted for outdoor photographs, is examined and discussed. Our results contribute to the deeper understanding of landscape pattern and processes, responsible for visual and recreational values, as well as the methodology is based on the integrated quantitative approach, supporting evidence-based landscape science and decision-making.</p>

<p>Over the past decade, we witnessed a rapid growth in the use of social media data when assessing  cultural ecosystem services (CESs), like modelling the supply-demand relationships. Researchers increasingly use user-generated content (predominantly geotagged pictures and texts from Flickr, Twitter, VK.com) as a spatially explicit proxy of CES demand. However, for modelling CES supply most of such studies relied on simplistic geospatial data, such as land cover and digital elevation models. As a result, our understanding of the favourable environmental conditions underlying good landscape experience remains weak and overly generic.</p><p>Our study aims to detect the spatial disparities between population density and CES supply in Estonia in order to prioritise them for further in-depth CES assessment and green and blue infrastructure improvements. We relied on Flickr and VK.com photographs to detect the usage of three CESs: passive landscape watching, active outdoor recreation, and wildlife watching (biota observations at organism and community levels) with automated image content recognition via Clarifai API and subsequent topic modelling. Then, we used Landsat-8 cloudless mosaic, digital elevation and digital surface models, as well as land cover model to derive 526 environmental variables (textural, spectral indices and other indicators of landscape physiognomy) via the Google Earth Engine platform. We conducted an ensemble environmental niche modelling to analyse the relative strength and directions of relationships between these predictors and the observed occurrence of CES demand. Based on multicollinearity and relative importance analysis, we selected 21 relevant and non-collinear indicators of CES supply. With these indicators as inputs, we then trained five models, popular in environmental niche modelling: Boosted Regression Trees, Generalized Linear Model, Multivariate Adaptive Regression Spline, Maxent, and Random Forest. Random Forest performed better than the other models for all three CES types, with the average 10-fold cross-validation area under curve > 0.9 for landscape watching, >0.87 for outdoor recreation, and >0.85 for wildlife watching. Our modelling allowed us to estimate the share of the Estonian population residing in the spatial clusters of systematically high and low environmental suitability for three considered CESs. The share of the population residing in the clusters of low environmental suitability for landscape watching, outdoor recreation, and wildlife watching is 5.5%, 3.1%, and 7.3%, respectively. These results indicate that dozens of thousands of people in Estonia (population is >1.3 million) likely have fewer opportunities for everyday usage of considered CESs. However, these results are biased as there was not enough evidence in social media for CES use in some of these areas.</p><p>Although our results should be treated with caution, because social media data are likely to contain a considerable sampling bias, we have demonstrated the added value of remote sensing data for CES supply estimation. Given nearly global and continuously updated satellite imagery archives, remote sensing opens new perspectives for monitoring the loss and gains in landscape suitability for CES across temporal and spatial scales. As such, we can better account for the intangible underlying geospatial features that can influence  economic and environmental decision-making.</p>