A spatio-temporal feature extraction algorithm for crop mapping using satellite image time-series data

Crop type identification is a prerequisite for several agricultural analyses. Thus, various methods have been used to accurately identify different crop types. Classification of satellite image time-series (SITS) data is probably the most efficient one, among these methods. Recently, the SITS data with high spatial and temporal resolution have become widely available. This category of SITS data, in addition to information about the temporal phenology of crops, provides valuable information about the spatial patterns of the croplands. This information, if extracted properly, can increase the accuracy of crop classification. In this paper, we proposed a novel feature extraction algorithm in order to extract this information. The proposed feature extraction algorithm is a two-step algorithm. In the first step, an image segmentation method is used to partition the time-series data into several homogenous segments. The pixels of each segment share similar spatial and temporal characteristics. In the second step, the algorithm fits a polynomial function to the average value of pixels of each segment. Finally, the coefficients of the fitted polynomial function are considered as the spatial-temporal (spatio-temporal) features. The effectiveness of the proposed spatiotemporal features was evaluated based on their obtained crop classification accuracies. In this paper, the SITS data were constructed by extracting normalized difference vegetation index (NDVI) and soil-adjusted vegetation index (SAVI) from 10 RapidEye images of an agricultural area. Support vector machines (SVM) was considered as the classification algorithm. The obtained results of the experiments showed that the proposed spatio-temporal features by proving the classification accuracy of 87.93% and 75.96% respectively for NDVI and SAVI time-series can be very efficient features for crop mapping. These features also sharply improved the crops classification accuracy in comparison with other spatial and temporal features.