Wheat planted area detection from the MODIS NDVI time series classification using the nearest neighbour method calculated by the Euclidean distance and cosine similarity measures

Abstract

This research aims to detect the wheat crop in the Northwest region of Rio Grande do Sul (Brazil) using MODIS NDVI time series. Detection of wheat crops presents two difficulties: (a) high variation of wheat phenological curves due to climatic fluctuations during the crop cycle and (b) the plantations are in an environment with different types of rural and urban targets. In solving these problems, we propose a classification based on the nearest neighbour (a specific case of the k-NN method) from the similarity and distance metrics combined with the determination of the best threshold value to individualize the wheat mask. The nearest neighbour classification using minimum distance (Kappa of 0.75) obtained a result equivalent to that of cosine similarity (Kappa of 0.74) as attested by the McNemar test. The planted area result was comparable to official statistics from the Brazilian Institute of Geography and Statistics obtained through direct interviews.