Identification of dynamically homogeneous areas with time series segmentation of remote sensing data

We propose a method of segmentation of remote sensing time series data, which exploits multi-temporal information to identify objects’ boundaries. Extracting homogeneous objects with similar temporal behavior, the method analyzes large volumes of multi-temporal input data in a piecewise way and produces a consistent output segmentation layer for large territories. Segment building logic is simplified to minimize the computation time, while objects’ boundary identification accuracy remains sufficient for remote monitoring and mapping of vegetation, and specifically, agricultural crops. At the Space Research Institute of the RAS, the proposed method is currently applied for automated on-line satellite imagery analysis for recognition and mapping of (winter and spring) crops on large territories and land-use evaluation. The method successfully deals with gaps in remote sensing time series data and performs well even when input images are contaminated with speckle noise. Due to its ability to map dynamically homogeneous surface areas with partially missing data, the method provides a potential for their recovery.