Matching environmental data produced from remote sensing images to demographic data in Sub-Saharan Africa

In a globalized context increasingly impacted by climate change, demographic studies would gain from taking environmental data into account and be carried out at the transnational level. However, this is not always possible in Sub-Saharan Africa, as matching harmonized demographic and environmental data are seldom available. The large amount of data regularly acquired since 2015 (in 2019 only, Sentinel satellites from the European Space Agency produced 7.54 PiB of open-access data) are an opportunity to produce relevant standardized indicators at the global scale. Several indicators have been developed to help understanding geographical realities in a consistent (i.e., not location dependent) manner. Among them, local climate zones (LCZ) have been proposed by WUDAPT (World Urban Database and Access Portal Tools) to systematically label urban areas [2]. Their goal is to provide a map of the world following this legend, in open access, that can later be used by researchers for a wide range of studies. This data has been used to understand energy usage [1], climate [3] or geoscience modeling [10] or land consumption [5]. An important amount of work has been dedicated in the recent years to the automatic generation of such data, from sensors such as Landsat 8 or Sentinel 2. In a research competition organized by the IEEE IADF, several methods have been proposed to map LCZ from Landsat, Sentinel 2 and OpenStreetMap data [11]. Another recent study focused on the usage of Convolutional Neural Networks (CNNs) to tackle the task of automatically mapping LCZ using deep learning [7] and a large-scale benchmark dataset was proposed in [12], with a baseline of an attention-based CNN. However, these works mostly focused on developed urban areas. For instance, the challenge of [11] targeted Berlin, Hong Kong, Paris, Rome, São Paulo, Amsterdam, Chicago, Madrid, and Xi’An. This is problematic, as developed cities are generally well mapped through governmental censuses, and that spatial generalization of machine learning based methods is a challenge [6]. It is therefore necessary to develop adapted methods for developing areas [9]. In this work, we explore different methods to predict LCZ from Sentinel-2 data. The originality of the approach is to train a convolutional network-based model (ResNet34 [4]) on clusters of data representing similar morphological features as our target city (Ouagadougou, Burkina Faso). To select relevant cities as training data, we use the classification proposed in [8] and intersect relevant cities with those represented in the large-scale LCZ dataset [12]. As such, our dataset is composed of areas covering Karachi an Islamabad, Pakistan, Cairo, Egypt and Hong-Kong, China. Preliminary results show that ResNet34 [4] achieves good performance when training it on images representing similar morphological features (Overall accuracy: 94%). We perform LCZ classification on Ouagadougou with this model. It exhibits two main findings: • Resnet34 [4] can be generalized to an unseen area which has similar morphological features to the training areas. • Some of LCZ classifications are dependent to seasonal variations. In particular, we observed that some classes which do not contain vegetation (i.e., expected to be invariant to seasons) have not been similarly predicted when looking at several seasons. We attribute this phenomenon to the lack of seasonal changes within the training dataset, which does not consider weather fluctuations. Figure 1 shows LCZ classifications of Ouagadougou according to the seasons. Results are globally consistent, and seasonal misclassifications are more frequent when looking at the outskirts of the city. Red classes are buildings, so are expected to be invariant to the seasons. As well as for seasons, this result highlights the necessity to generate data for rural areas, as training on urban areas does not appear to generalize well when inferring on rural areas. To study the correlation between population data and LCZ, we cross-referenced Ouagadougou’s population density data with our classification results in figure 2 and investigate the population density per LCZ class. As expected, class with compact mid-rise buildings are correlated with a high population density. Compact low-rise buildings are associated to a lower population density, and natural areas are predicted as not populated areas, which validate the results of our model. References: [1] Paul John Alexander, Gerald Mills, and Rowan Fealy. Using lcz data to run an urban energy balance model. Urban Climate, 13:14–37, 2015. [2] Benjamin Bechtel, Paul J Alexander, Jürgen Böhner, Jason Ching, Olaf Conrad, Johannes Feddema, Gerald Mills, Linda See, and Iain Stewart. Mapping local climate zones for a worldwide database of the form and function of cities. ISPRS International Journal of Geo-Information, 4(1):199–219, 2015.2 [3] Jan Geletič, Michal Lehnert, Petr Dobrovoln`y, and Maja Žuvela-Aloise. Spatial modelling of summerclimate indices based on local climate zones: expected changes in the future climate of brno, czech republic. Climatic Change, 152(3-4):487–502, 2019. [4] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. [5] Jingliang Hu, Yuanyuan Wang, Hannes Taubenböck, and Xiao Xiang Zhu. Land consumption in cities: A comparative study across the globe. Cities, 113:103163, 2021. [6] Emmanuel Maggiori, Yuliya Tarabalka, Guillaume Charpiat, and Pierre Alliez. Can semantic labeling methods generalize to any city? the inria aerial image labeling benchmark. In2017 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pages 3226–3229. IEEE, 2017. [7] Chunping Qiu, Michael Schmitt, Lichao Mou, Pedram Ghamisi, and Xiao Xiang Zhu. Feature importance analysis for local climate zone classification using a residual convolutional neural network with multi-source datasets. Remote Sensing, 10(10):1572, 2018. [8] Hannes Taubenböck, Henri Debray, Chunping Qiu, Michael Schmitt, Yuanyuan Wang, and Xiao Xiang Zhu.Seven city types representing morphologic configurations of cities across the globe. Cities, 105:102814, 2020. [9] John E Vargas-Muñoz, Sylvain Lobry, Alexandre X Falcão, and Devis Tuia. Correcting rural building annotations in openstreetmap using convolutional neural networks. ISPRS journal of photogrammetry and remote sensing, 147:283–293, 2019. [10] Hendrik Wouters, Matthias Demuzere, Ulrich Blahak, Krzysztof Fortuniak, Bino Maiheu, Johan Camps,Daniël Tielemans, and Nicole Van Lipzig. The efficient urban canopy dependency parametrization (sury)v1. 0 for atmospheric modelling: description and application with the cosmo-clm model for a Belgian summer. Geoscientific Model Development, 9(9):3027–3054, 2016. [11] Naoto Yokoya, Pedram Ghamisi, Junshi Xia, Sergey Sukhanov, Roel Heremans, Ivan Tankoyeu, BenjaminBechtel, Bertrand Le Saux, Gabriele Moser, and Devis Tuia. Open data for global multimodal land use classification: Outcome of the 2017 IEEE GRSS data fusion contest. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(5):1363–1377, 2018. [12] Xiao Xiang Zhu, Jingliang Hu, Chunping Qiu, Yilei Shi, Jian Kang, Lichao Mou, Hossein Bagheri, Matthias Häberle, Yuan sheng Hua, Rong Huang, et al. So2sat lcz42: A benchmark dataset for global local climate zones classification. arXiv preprint arXiv:1912.12171, 2019