Innovating land cover monitoring by complementing annual land cover maps with near real-time observations

Most operational global land cover datasets, such as the Copernicus Global Land Service (CGLS) Land Cover product at 100m resolution, National Aeronautics and Space Administration (NASA) MCD12Q1 dataset at 500 m resolution, or the European Space Agency (ESA) Climate Change Initiative (CCI) land cover dataset at 300 m resolution, focus on classifying the Earth surface at a yearly interval. The yearly interval allows to feed the classifier with both temporal (e.g. phenological cycles) and spectral information, increasing the accuracy of the classification and facilitating the separation of a rich set of land cover classes. For example, it allows for the distinction between evergreen and deciduous trees by capturing seasonal foliage changes, as well as the identification of croplands versus grasslands through the detection of harvest and planting/sowing events. 

While the annual land cover products are an invaluable source of information for a suite of applications, including monitoring and modelling the Earth’s surface, their fixed legend system does not always serve the needs of all users. In addition, the yearly update interval does not support the near-real time monitoring of fast changes (e.g. forest loss) or seasonal change (e.g. cropland rotations or changing water levels in ephemeral water bodies). Timely information on land cover and land cover change is however crucial to understand anthropogenic pressures on our environment, the state of ecosystems, and their evolution.  

This prototype aims to develop a methodology to map a set of basic, independent diagnostic categories and characteristics describing the properties of the Earth’s surface. These categories, called Land Surface Categories (LSC), include the presence of snow, water, woody vegetation. Our mapping will be done at 10m resolution, with a near real-time update frequency based on Sentinel-2 data. 

The near-real time mapping of LSC complements the information provided in the yearly land cover maps. For instance, the continuous, near-real time information on LSC has the potential to capture both land cover dynamics (e.g. forest loss) and to get information on seasonal dynamics (e.g. seasonal presence of water, snow and ice). The dynamic nature of the LSC may additionally support downstream applications, such as the production of land cover and land cover change products. For instance, temporal information on the presence of bare and herbaceous vegetation may support the classification of cropland and grassland on an annual basis. Other applications may extend to mapping of dynamics (e.g. the mapping of snow, water, or vegetation cycles) or the generation of user-defined classification systems (e.g., the mapping of ephemeral versus persistent water bodies based on user-defined thresholds).  

Yet, given that restricted or no temporal information is available for the near real-time updates, the LSC classes should be tailored to match surface properties that are directly observable from single scenes. From a single vegetated scene, it is for instance difficult to distinguish grassland and cropland. To make this differentiation, temporal information on harvest and sowing/planting events should be available. Hence, pixels that are covered by cropland (land use/land cover class) will be mapped as a combination of bare/sparsely vegetated and herbaceous vegetation in the LSC. 

To map the LSC, we are training deep learning architectures that are fed with a set of Sentinel-2 features and localizing features (such as latitude and longitude). Deep learning models have the capacity to capture spatial context, but require a large number of training samples, which is time consuming to annotate from single scenes. To facilitate the collection of LSC annotations, a novel annotation tool is being developed that supports active learning. 

The main challenge while developing the product is to maximize temporal consistency while keeping the sensitivity to change. During several months, the spectral signal of herbaceous and woody vegetation is for instance similar, leading to confusion between these classes. This results in false seasonal fluctuations in woody cover. Hence, the method should find a balance between reducing these fluctuations induced by confusion and maximizing the capacity to find near real-time changes in woody cover due to harvest, planting, etc. 

In line image credit: VITO

Cover image credit: European Union, Copernicus Sentinel-2 imagery 

Would you be interested to follow the results of EvoLand and the development of this candidate prototype? Sign up for our newsletter (here below or via this form), follow us on LinkedIn, Mastodon or X (Twitter). 

Questions about the article or the project? E-mail us at contact@evo-land.eu.  

== 

This article is part of a series providing more details on all 11 EvoLand candidate prototypes.

Previous article in the series: Detecting forest disturbance with Earth Observation.