CLDynamicLandCover

The CLDynamicLandCover dataset is a dynamic land use dataset for Central Chile, covering latitudes -31.61 to -43.50. It provides land cover maps on a five-year scale from 1990 to 2018, created by a semi-automatic algorithm initially developed for the coastal sector of south-central Chile. This method was later expanded to cover areas from the southern Coquimbo region to the southern Los Lagos region.

The dataset uses a supervised classification approach with satellite data (Landsat and SRTM) to represent reflectivity and topography, and auxiliary cartographic information. It includes training points for 15 land cover classes, photointerpreted and georeferenced in 2018, to create a spectral signature for each class. Predictive variables are selected using machine learning algorithms and expert criteria.

The 2018 endmember is used to estimate the Jeffries-Matusita Distance for points from other years (1990, 1999, 2004, 2009, 2013), determining whether a pixel's land cover class changed or remained constant. A separability cutoff threshold optimizes class discrimination between years. Points below the threshold are retained (indicating no change), and those above are eliminated (indicating change). These points train a Random Forest classification model to generate land cover maps for each year.

Land trajectories were corrected based on ecological transitions and economic cost criteria, with a new classification for harvested plantations. The product shows high accuracy, with values ranging from 0.894 to 0.950 and Kappa coefficients from 0.877 to 0.943 across the years.