Maximising automation in land cover monitoring with change detection

With the availability of free moderate spatial resolution Landsat satellite data land cover mapping systems are moving away from classifying single date cloud-free images to classifying data time-series. This requires the ability to handle large volumes of data, which in turn requires high levels of automation in data pre-processing, image classification and change detection. This paper reports on the progress made towards the development of a more automated land cover monitoring system for South Africa. We firstly employed a local installation of the Web-enabled Landsat Data (WELD) system to serve as the data “backbone” for pre-processing and storing large amounts of Landsat data sensed over South Africa. A system was developed to rapidly update land cover maps for previous mapped areas using highly-automated, training data generation, scalable random forest classification, accuracy assessment, change detection and rapid, online operator validation. The technology is aimed at assisting government and industry to provide land cover data at a much higher update frequency to address ever-increasing demands for land cover products and services.