Automated object-based satellite image time series classification using dynamic time warping
Abstract
Satellite image analysis can support efficient monitoring of crops by providing timely and relevant spatial and temporal information products and tools. However, there are challenges analyzing satellite image time series (SITS), such as: few available samples, irregular temporal sampling of images, and temporal variability of the phenomena of interest. The Dynamic Time Warping (DTW) method is able to address these challenges by comparing the similarity between two temporal sequences, finding their optimal alignment and providing a dissimilarity measure between the evaluated temporal sequences. In this paper, we present an automated implementation of DTW for processing Sentinel-2 SITS within an object-based image analysis framework, implemented as a ready-to-use parameter-free algorithm in eCognition software. By adapting the DTW to be object-based, we solve one of the main challenges of DTW, namely a reduction in computational time. We tested the developed tool on Sentinel-2 SITS in south-eastern Romania, using 11 images of 5192x4367 pixels. The 10m spatial resolution red, green, blue and near-infrared bands were used in the segmentation process. To speed up the segmentation, we used SLIC superpixels for an initial oversegmentation of the images, and then applied multiresolution segmentation on the resulted superpixels using the ESP2 tool with an automated selection of scale parameter. We used 30 training samples per class, with five classes of interest (wheat, maize, sunflower, forest, water), to extract the NDVI temporal pattern to be used in DTW. The tool compares the temporal pattern of an object with the training temporal patterns, classifying the object into the class to which it had the minimum DTW dissimilarity value. The overall accuracy for resulted classification was 90.27%, with kappa index of 0.87. Besides the categorical classification of SITS, the tool delivers also a map of DTW dissimilarity values, which is useful to understand which objects have a strong similarity with the assigned class (values closer to 0) or can have a doubtful assignment to one of the five classes used (the largest values). This tool represents the first operational implementation of the DTW method at the object-level. Therefore, it can benefit agricultural management and monitoring agencies at different scales towards designing and implementing sustainable agricultural practices. The availability of this kind of operationalized approaches dedicated to turning data into information are increasingly necessary as we move into imagery-dense workflows with satellite constellations providing increasing spatial and temporal resolutions.