Histogram-based spatio-temporal feature classification of vegetation indices time-series for crop mapping

Abstract

Classification of time-series of vegetation indices (VIs) can be a reliable strategy for identifying and monitoring different crop types. Recently, with the advent of new sensors, the time-series data with high spatial and temporal resolutions have become widely available and used for constructing various VIs time-series. These high-resolution time-series, in addition to temporal information about the crops’ phenology, contain valuable information about the spatial patterns of croplands. This information can be used to increase the performance of crop classification. In order to properly extract both spatial and temporal information from the time-series of VIs, we proposed the concept of histogram-based spatio-temporal (HST) features. These features represent each pixel in a time-series by the histogram of its spatio-temporal neighborhood. The HST features, like any other histogram-based features, are characterized by high dimensionality and sparseness. Consequently, the common classification algorithms cannot be employed for their classification. To address this issue, we presented Support Vector Machines (SVM) using an intersection kernel, which is specifically proposed for classification of histogram-based features. Time-series of three different vegetation indices, namely, Normalized Difference Vegetation Index (NDVI), Soil Adjusted Vegetation Index (SAVI), and Red Edge Normalized Difference Vegetation Index (NDVIRE) were considered to evaluate the performance of the HST features. The results of experimental tests showed that the HST features by yielding the overall accuracy of 88.31%, 87.27% and 84.36% for NDVIRE, NDVI, and SAVI respectively are much more informative than other textural features used for comparison. Moreover, we provided a detailed analysis of the performance of the HST features concerning the size of the spatio-temporal neighborhood and the number of histogram’s bins.