RSRU – Automated land-cover change detection

Meraka Institute researchers (front) Waldo Kleynhans and Dr Frans van den Bergh with Professor Corne Olivier (University of Pretoria) (middle left), Dr Konrad Wessels and Brian Salmon, and (back) Seare Araya, Albert Gazendam and Asheer Bachoo

Land-cover change often indicates land use change with major socio-economic impacts. The transformation of vegetation cover (e.g. deforestation, agricultural expansion and urbanisation) has significant impacts on hydrology, ecosystems and climate.

Researchers at the Meraka Institute of the CSIR have taken on the challenge of automated land-cover change detection by applying novel signal processing, time-series analyses and machine learning methods to the problem, using high-performance computing resources. The multidisciplinary research involves computer scientists, electronic engineers, ecologists and remote sensing specialists.

Mapping national land cover with high resolution satellite imagery takes years to complete, at great cost. The 1995/96 National Land Cover (NLC) and 2000 NLC studies were based on single-date Landsat TM and ETM imagery (30 m pixel resolution) and is frequently used for various spatial planning purposes.

Unfortunately, the 1995 and 2000 versions of the NLC were compiled using very different methods. As a result, land use and land-cover change cannot be deduced reliably from a map-to-map comparison of the two dates. An alternative or complementary approach is to use an extended time-series of daily coarse-resolution satellite images (e.g. the moderate resolution imaging spectroradiometer (MODIS) at 500 m pixel resolution), to identify areas of potential land-cover change, which can then be further investigated using high-resolution satellite images.

Despite the coarse resolution of MODIS images compared to Landsat TM and ETM imagery (500 m versus 30 m), the temporal profile of the pixels of various land cover and/or uses provides a distinctive temporal signature or signal. For example, the natural seasonal patterns of green grasslands during the rainy season versus the senescent, brown grasslands in the dry season are very different from the seasonal patterns of irrigated croplands or expanding rural settlements.

For regional applications, these change detection methods need to be sufficiently automated to process very large volumes of data and minimise time-consuming,expert human interpretation. As daily, regional coarse-resolution datasets become more accessible and computational resources become more affordable, such automated change detection systems should become more attainable.

Monitoring land-cover change in such an automated fashion has nevertheless remained an elusive goal in environmental remote sensing, as two previous NASA-funded research initiatives have been unable to develop operational systems for the MODIS sensor. It is clear that this can be achieved only through crossdisciplinary research benefiting from the advancements in machine learning, statistics, database technology, high-performance computing, data visualisation and image processing.

Project leader, Dr Konrad Wessels, summarises the team objective as “Interpret the present in the context of the past, in near-real time!” Researchers are collaborating with the Department of Electrical, Electronic and Computer Engineering of the University of Pretoria, the School of Computer Science at the University of the Witwatersrand, the ecosystems and earth observation research group of the CSIR and the CSIR Cluster Computing Centre (C4).The team was recently joined by Dr Amandine Robin, a postdoctoral student from France.

Methods previously limited to the fields of telecommunications and human language technology are now being applied to environmental monitoring, using a time-series of satellite images. The data (or signal) from each image pixel are analysed through time to identify changes within a seven-year period (2000 to present). This requires that very large image archives (in the order of two terabytes) be restructured for efficient per-pixel time-series analyses, initially a daunting feat in itself. Various time-series classification and clustering methods are then used to detect change whenever a pixel’s spectro-temporal characteristics changes to that of a different land cover type.

Adapting these techniques to work with remotely-sensed time-series data presents a new challenge to the research community. Moreover, change detection across the entire South Africa has the added challenge of scale and can only be achieved in a high-performance computing environment.

Representation of a Multi-Temporal Image Analysis

The system has been dubbed ‘HiTempo’, touting its ability to rapidly perform time series analysis of hyper-temporal satellite data. Two major challenges have emerged: Distinguishing human impacts on the land surface from interannual rainfall variability; and determining the accuracy of change detection methods, given limited regional validation data, i.e. well-documented examples of known land-use land-cover change.

The issues are being addressed partially by simulating land-cover change, such as informal settlement or cultivated field expansion, by blending the signals from near-by pixels of contrasting land use from a specific point in time. A database of known examples of land cover change is also being compiled.