SiRCub – Brazilian Agricultural Crop Recognition System

This paper presents a novel approach to classify agricultural crops using NDVI time series. The novelty lies in i) extracting a set of features from the each and every NDVI curve, and ii) using them to train a crop classification model using a Support Vector Machine (SVM). Specifically, we use the TIMESAT program package to: 1) smooth the time series, 2) decompose them into agricultural seasons –a season is the period between sowing and harvesting–, and 3) extract the features for each season. The 11 crop features we extract include the start and end of season, its amplitude, and the curve gradients of the sprouting and senescence periods, among others. Once we have the collection of features, they are fed into an SVM system –we use the LIBSVM library–, together with a collection of annotations about the land use of the corresponding time series. These annotations represent the type of crop for a given location and agricultural season, and they are provided by the specialists of the Embrapa. As a result we obtain a classification model that allows for identifying different crop classes. Our methodology is generic and can be applied to a variety of regions and crop types. We have develop a system called SIRCUB (Sistema de Reconhecimento de Culturas brasileiro), which implements such methodology. Thus, we describe in this paper the architecture of the system and the crop model learning methodology.