The use of artificial neural networks to classify the social vulnerability of municipalities in Rio Grande do Norte State, Brazil

Abstract: The objective was to apply artificial neural networks to classify municipalities (counties) in Rio Grande do Norte State, Brazil, according to their social vulnerability. This was an ecological study using 17 variables that reflected epidemiological, demographic, socioeconomic, and educational indicators for the year 2010. The sources were the Human Development Atlas for Brazil and the Brazilian Institute of Geography and Statistics. For classification of the municipalities, the study applied the artificial neural networks of the PNN and Multilayer feedforward types, resulting in a classification in five categories of vulnerability: very high, high, medium, low, and very low. The networks’ training phase used the minimum and maximum values, 25th and 75th percentiles, and medians for the 17 selected variables. The Multilayer feedforward network with six nodes showed the best results. The municipalities from the Metropolitan Area (Natal, Parnamirim) and the eastern and western Seridó micro-regions (Caicó, Currais Novos, São José do Seridó, Jardim do Seridó, Parelhas, Carnaúba dos Dantas) showed the lowest levels of vulnerability. The municipalities with high and very high vulnerability were located in the East of the state, in the micro-regions of the Northeast Coast (João Câmara, Touros, Caiçara do Rio dos Ventos) and Southern Coast (Nísia Floresta, São José do Mipibu, Arês, Canguaretama). The neural network classified the municipalities with high precision, distinguishing those with extreme vulnerability from those with better social indicators.摘要：本研究旨在应用人工神经网络对巴西北大河州（Rio Grande do Norte）的市镇（县）开展社会脆弱性分类研究。本研究为生态学研究，采用2010年的17项指标变量，涵盖流行病学、人口学、社会经济及教育维度。数据来源为巴西人类发展地图集（Human Development Atlas for Brazil）与巴西地理与统计研究所（Brazilian Institute of Geography and Statistics）。针对市镇分类任务，本研究采用了概率神经网络（Probabilistic Neural Network, PNN）与多层前馈型神经网络两种模型，将社会脆弱性划分为极高、高、中等、低、极低五个等级。神经网络的训练阶段以17项选定变量的最小值、最大值、25分位数、75分位数及中位数作为参考基准。其中，包含6个节点的多层前馈神经网络表现最优。大都市圈（纳塔尔（Natal）、帕纳里米尔（Parnamirim））以及东部和西部塞里杜微区域（卡伊库（Caicó）、库赖斯诺沃斯（Currais Novos）、圣若泽杜塞里杜（São José do Seridó）、雅尔迪姆杜塞里杜（Jardim do Seridó）、帕雷利亚斯（Parelhas）、卡瑙巴杜斯当塔斯（Carnaúba dos Dantas））所辖的市镇，社会脆弱性水平最低。社会脆弱性等级为高和极高的市镇，分布于该州东部的东北海岸微区域（若昂卡马拉（João Câmara）、图鲁斯（Touros）、卡伊萨拉杜里约多斯文图斯（Caiçara do Rio dos Ventos））与南部海岸微区域（尼西亚弗洛雷斯塔（Nísia Floresta）、圣若泽杜米皮布（São José do Mipibu）、阿雷斯（Arês）、坎瓜雷塔马（Canguaretama））。该神经网络模型分类精度较高，可有效区分极端脆弱性市镇与社会发展指标更优的区域。