Satellite imagery and machine learning for aridity disaster classification using vegetation indices

Abstract

Central Java Province is one of provinces in Indonesia that has a high aridity risk index. Aridity disaster risk monitoring and detection can be done more accurately in larger areas and with lower costs if the vegetation index is extracted from the remote sensing imagery. This study aims to provide accurate aridity risk index information using spectral vegetation index data obtained from LANDSAT 8 OLI satellite. The classification of drought risk areas was carried out using k-nn with the Spatial Autocorrelation method. The spectral vegetation indices used in the study are NDVI, SAVI, VHI, TCI and VCI. The results show a positive correlation and trend between the spectral vegetation index influenced by seasonal dynamics and the characteristics of the High R.A. and Middle R.A. drought risk areas. The highest correlation coefficient is SAVI with a High R.A. amounted to 0.967 and Middle R.A. amounted to 0.951. The results of the Kappa accuracy test comparison show that SVM and k-nn have the same accuracy of 88.30. The result of spatial prediction using the IDW method shows that spectral vegetation index data that initially as an outlier, using the k-nn method, the spectral vegetation index data can be identified as data in the aridity classification. The spatial connectivity test among sub-districts that experience drought was done using Moran’s I Analysis.