Developing Land Cover and Land Cover Change Mapping Methods for the UK: Concept and Preliminary Results
Rowland, Clare; Morton, Daniel; Wood, Claire
CEH Lancaster, UNITED KINGDOM
Land cover and land-use change are key causes and consequences of environmental change. The accurate monitoring of land cover change is a key issue nationally and internationally. Land cover monitoring in the UK has been achieved via the UK’s Land Cover Map (LCM) series. LCMs have been produced for 1990, 2000 and 2007. The LCM series are a set of image-classification based land cover maps produced mainly from Landsat data, but with some LISS-III and SPOT-4/5 data. Over the last 20 years there have been developments in technical capabilities and user requirements, which have resulted in changes to the land cover classes and the spatial framework of the LCMs. Consequently, each of the Land Cover Maps has been produced with a different spatial framework. LCM1990 (originally called GB1990) was a per-pixel classification, whilst the second UK LCM was a segment-based classification, with segments derived from satellite data. The most recent map, LCM2007, is an object-based land cover map of broad habitats for the UK containing around 10 million objects. LCM2007 is based on the generalisation of OS MasterMap, to provide the spatial framework for Great Britain, and Ordnance Survey of Northern Ireland, to provide the Northern Ireland framework. The three LCMs are valuable data sets, but their differing spatial frameworks, and in the case of LCM1990 slightly different land cover classes, prevent their use for land cover change mapping. Additionally, a number of studies have show that post-classification change detection performs poorly as a system for change detection. Consequently, a new system is required that enables both stock and change to be monitored. In this paper we discuss the concept and preliminary method development for the UK Land Cover Information System, which is intended to eventually provide land cover and land cover change information.
The paper outlines the current concept for the Land Cover Information System and discusses the testing of some of the key underlying concepts at study areas in Norfolk and the Cairngorms area of Scotland. The two sites are strongly contrasting, with the Norfolk site dominated by agriculture, with some fenlands, whilst the Cairngorms is dominated by semi-natural upland vegetation types. One of the methods being tested for suitability in the final system is the use of a network of fixed training areas across the UK. The advantage of using fixed training areas, such as those used for the MODIS land cover products, is that they minimise the manual intervention required in the classification process. We present results for both study areas showing the accuracy of classifications from manually-chosen and fixed training areas. The impact on accuracy of different classifiers, including Random forest and Maximum Likelihood, is also assessed. Some preliminary results from the change detection are also presented. The discussion focuses on issues affecting the development of databases of training and validation areas, plus the choice of classification algorithm, and the suitability of Sentinel data for UK land cover mapping needs.