Caracterización del daño causado por granizo en la región agrícola argentina utilizando datos de sensores remotos satelitales ópticos y radares

Abstract

Weather hazards are becoming more frequent and intense as a result of climate change. Hailstorms usually cause the total loss of the harvest, exceeding the financial capacity of agricultural producers. For this reason, in Argentina, this risk is transferred to specialized agricultural insurance companies. After a hailstorm, the affected field is inspected by an insurance claims adjuster to assess yield loss. Assessment accuracy depends largely on detection of Homogeneous Damage Zones (HDZ). Currently, these areas are quantified in situ and their Identification becomes complex in large plots with tall crops and heterogeneous damage. This Thesis presents an algorithm for automatic detection of homogeneous hail damage. Unsupervised machine learning techniques are applied to vegetation indices calculated from satellite data provided by the Sentinel -1 and -2 missions of the EU’s Copernicus Programme. The first step of the algorithm is the pre-processing of the images provided by both missions, eliminating the noise of the microwave signal and the interference of the atmosphere in the spectral signal. Five microwave and five spectral indices were evaluated before and after a hailstorm in zones with different degrees of damage. The time series and rates of change of these indices were used as input variables in the K-means method for clustering pixels from each plot into different ZHDs. To validate the quality of the pixel groupings in ZHD, we compared the means of the percentages of damage evaluated in situ with the data of each of the zones defined by the algorithm. Homogeneity between the ZHD was tested using ANOVA (one-way) test. induced changes. Validation of the suggested algorithm showed that in 87.01% of cases there was significant evidence of differences in average damage between zones determined by the algorithm within the plot. Thus, the algorithm presented in this Thesis allowed efficient detection of homogeneous hail damage zones, which is expected to improve accuracy and transparency in the characterization of hailstorm events. This new method was published in the journal Agronomy (Sosa et al., 2021b). The journal's impact factor is 3,417, and it ranks 18/91 (Q1) in the "Agronomy" category of the JCR (Journal Citation Reports). On May 9, 2022, the article has been cited in two publications (Ha et al., 2022; Watson-Hernández et al., 2022). In addition, it has been presented in an oral communication at the Third International Congress on Geomatics Engineering, held in Valencia in July 2021, and was published in the proceedings of the congress (Sosa et al., 2021a). Also, on May 25, a poster will be presented at the XVIII Edition of the Spanish Conference on Biometrics. Simultaneously, the new method developed within the framework of this thesis was incorporated into an application for mobile devices. The method is currently in use by more than 50 the agricultural risk loss adjusters of the Argentine insurance cooperative "La Segunda".