An Iterative Algorithm for Automatic Fitting of Continuous Piecewise Linear Models

Continuous piecewise linear models constitute useful tools to extract the basic features about the patterns of growth in complex time series data. In this work, we present an iterative algorithm for continuous piecewise regression with automatic change-points estimation. The algorithm requires an initial guess about the number and positions of the change-points or hinges, which can be obtained with different methods, and then proceeds by iteratively adjusting these hinges by displacements similar to those of Newton algorithm for function root finding. The algorithm can be applied to high volumes of data, with very fast convergence in most cases, and also allows for sufficiently close hinges to be identified, thus reducing the number of change-points, and so resulting in models of low complexity. Examples of applications to feature extraction from remote sensing vegetation indices time series data are presented. Key–Words: Continuous piecewise regression, Segmented regression, Multiple change-point models, Remote sensing, NDVI, MODIS.