Land Cover Side Event, GEO-XII

GEO-XII Side Event:

Land Cover – Harmonized Pathways Towards Policy Needs

9 November 2015, Mexico City

KEY POINTS

Growing demand and expectations

  • Growing demand can provide some justification for what is needed, including a greater range of sensors and observations, improved data processing capabilities, new algorithm development, better access, and more regularityand consistency in acquisition and data product generation.
  • Growing expectations have implications for how the LC community (and to some extent the RS community) is perceived. The use of LC information in legally binding situations and for mandatory reporting linked to treaty compliance issues bring with it stringent requirements for accuracy, accountability, transparency and reliability, among other things.

Need for more levels of classification and types of input data sources

  • Finer, multi-level classification legends would provide more local and national relevance than a simple, single-level global scheme and (if appropriately defined) also allow for aggregation into broader classes with global relevance. This would enable better cross-mapping between a variety of global legends and thus support both global and sub-global needs. However, since accuracy tends to decrease as the number of classes increases this can be challenging. The FAO’s LCCS3 was cited as one successful hierarchical classification system.
  • Increase the number and types of input data sources (“data fusion”). Current processing can have limited accuracy due to the limited amount of information in the input data stream. For example, even perfect extraction of classes from a single remote sensing source may not be able to provide the degree of discrimination needed to generate multiple class levels because the spatial or spectral resolution may be insufficient (that is, the information simply does not exist in the data stream).
  • Use more non-remote sensing data sources.Integrating socio-economic, physiographic, or other types of non-remote sensing information with the remote sensing data stream can increase accuracy. This would enable more levels of classification, finer class resolution, improved accuracy, and broader user relevance.
  • Use more sources of remote sensing data. Combining multiple remote sensing data sources, particularly different types of sensors such as radar and optical, and multiple sources of each, would improve accuracy, increase discrimination capability (eg between types of crops or forests) and significantly improve inter- and intra-annual temporal sampling. Hyperspectral data, though not discussed (and currently not globally available) has great potential to increase accuracy.

New technologies and approaches

  • Consider data cube approach. This can greatly facilitate use of the data. Although a large change to the approach that most data providers currently it offers many advantages and its use should be further explored (note that CEOS is exploring the approach with Colombia and Kenya). The data cube approach builds a rich time series of pre-calibrated, pre-corrected, ortho-rectified satellite data from multiple satellite sensors into an analytical tool able to look at any area and show how land cover has changed over time. Using a data cube approach means that the scientist does not have to be an expert at remote sensing, but can still generate products from the rich data available from this source. Australia has demonstrated the ability to do this at Landsat resolution over the last 40 years for the entire continent and this could be developed as a global tool, greatly speeding up dynamic landcover analysis while significantly reducing costs.
  • Separate the RS image preparation component from the classification component. The image preparation component may be roughly 80% of the processing work. Separating it from classification would thus allow different classification approaches to be used on the same RS dataset without starting from scratch, thus greatly decreasing cost. Also, the two components need not be done by the same organization; this provides additional benefits, including enabling national “ownership” (an important user need)
  • A “single-class” approach has some advantages. For example, focus on “surface water” and then do not attempt to discriminate between natural/artificial ponds, lakes, rivers, deltas, estuaries etc. Or focus on buildings without wrestling with difficult to define classes such as ‘urban’. Other single-theme classes (sometimes also called ‘continuous-fields’) have enjoyed success, such as forest cover and grassland datasets.
  • Fusing local and regional maps into global datasets. “Hybrid” datasets are a way of improving accuracy and providing deeper and broader thematic content.

Land Cover Portal

  • A land cover data portal would increase access to land cover datasets and potentially provide a variety of useful services. However, by itself it would not be sufficient to address some of the key limitations of the current situation.

User needs

  • Non-technical challenges and the need for more dialog. Many of the challenges to more harmonization are not technical ones—they are social. More dialog among the players would lead to improvement.
  • Change is more useful than state. This implies a need for regular, consistent datasets.
  • Need appropriate infrastructure to support operationalization (of user needs).
  • National governments need control over the dataset and its generation. Governments want control over the information they provide to an international treaty, for example—they do not want to use information developed by an outside source with unknown details or of unknown consistency, even if it is “better”. This has very important implications. Ownership often goes hand in hand with capacity-building.
  • Regularity. To understand change repeat datasets are needed. To monitor change these must be provided on a regular, reliable basis with sufficient frequency.A growing demand is change detection, not just status; note that the temporal frequency can be more important than spatial resolution.
  • Consistency. In addition to being available on a regular basis, to understand land cover change the datasets need to be generated using a consistent methodology.
  • Accuracy. Obviously, sufficient accuracy is needed. This has implications for validation.
  • Legend must meet user needs. This is obvious, but the implication is that a single legend will not meet the needs of all users. However, a finer-grained, multi-level class structure would facilitate cross-mapping at higher levels and so support generation of a variety of global maps with different legends that meet a variety of user needs.Pathfinder project(s). Often the best way to foster interaction, engagement, dialog, and progress is to find a “do-able” project and get started on it. This may be a good way to address some of the needs that UNSD has, for example.