Avaliação de sinistros agrícolas via sensoriamento remoto orbital e aprendizado de máquina

• Autor(a) principal: Rambo, Eduardo Matias
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do UNIOESTE
• Texto Completo: http://tede.unioeste.br/handle/tede/5150
• topic: Seguro rural; Perfil espectro-temporal; EVI; TIMESAT; Support vector machine; Random forest; Decision tree; Eficiência; Agricultural insurance; Crop insurance; Time-series; EVI; TIMESAT; Support vector machine; Random forest; Decision tree; Efficiency; CIENCIAS BIOLOGICAS::BIOLOGIA GERAL

Agricultural insurance is an important alternative to convert the agricultural sector into a financially stable model, even when faced with adverse events and natural disasters. Agricultural insurance is a financial instrument to reduce risk related to natural disasters by establishing a future contract in which one party is obligated to compensate the damage loss to the other party by paying a premium. Hence, for the producer, it works as a way to substitute an uncertain future financial loss by a reduced, predictable investment. Due to the higher transaction values involved, it is necessary to develop methods to inspect farms and verify the claimed losses. Remote sensing has the potential to support the insurance industry by providing an alternative to crop monitoring in large scales and at a low cost, facilitating the processes of fiscalization and decision making regarding agricultural insurance. Thus, this research aims to develop a methodology to confirm if a claimed loss occurred by applying seasonal trend analysis in Landsat-8/EVI time-series combined with weather. For this effect, information about both affected and non-affected areas are employed, using real inspected farmers (sown with maize, soybean, and wheat), in order to verify the existing pattern between these parameters, indicating their occurrence and distinguishing the natural disaster, by comparing the judicial investigation data provided by the insurance company helping this research. For the analyses, three classifiers were applied: Decision Tree (DT), Random Forest (RF) and Support Vector Machine (SVM). RF classifier achieved 83, 96, and 81% for maize, soybean, and wheat, respectively, when determining whether a natural disaster did or did not occur. SVM classifier achieved 99 and 90% in maize and soybean, respectively, to detect the type of disaster, and RF achieved 86% for wheat at the same task. This methodology has proved to be efficient to confirm and detect a natural disaster, being a viable and important alternative solution for insurance companies to minimize their risks and increase their efficiency, helping in the process of insurance verification and fiscalization of actions related to such agricultural segments.