Cognitive modeling based on geotagged pictures of urban landscapes using mobile electroencephalogram signals and machine learning models

Abstract

Evaluating the impact of urban landscapes on human cognition is a hot issue in urban studies which has progressed by producing mobile electroencephalogram (EEG) devices. However, it is still challenging to investigate the effects of urban landscapes in remote places. Nowadays, geotagged pictures share much information about urban landscapes worldwide. This work aimed to model the effect of geotagged pictures of urban landscapes on two mental states of attention and meditation using mobile EEG signals with multi-layer perceptron (MLP), random forest (RF), and support vector regression algorithms. Thirty-five picture features from 350 pictures of 39 Iran cities, and EEG signals of 32 healthy adult participants trained models. Cross-validation revealed that all models performed well with slight differences and had good generalizability. Meanwhile, the most accurate results were related to the prediction of the meditation state by RF with R² coefficient of 0.895, root mean square error of 0.149, and mean absolute error of 0.114. Correspondingly, 0.792, 0.178, and 0.14 were similar values for the prediction of attention state by MLP (the least accurate predictions). The Gini index recognized color histogram and HSV (hue, saturation, value) color space as the most important features in predictions. Generally, color features were more important than entity features, confirming the high impact of colors in landscapes. Although this research has some limitations, in line with previous works, we observed that each picture affected participants’ minds differently, existing particular elements in urban landscapes gained attention and meditation levels, and pictures of green space increased attention level more than meditation. Overall, the proposed approach may help to understand how urban landscapes affect citizens’ cognition even in unnoticed and remote places. However, using more conceptual picture features in modeling can improve the findings.