The demonstration of the non-supervised techniques based on five real use cases (proposed during the RepreSent project).

This use case concentrates on timely and accurate detection of human induced forest disturbances in boreal forests. Such disturbances include typical forest management operations such as forest clearcutting and thinning, as well as various kinds of naturally occurring damages, such as damage because of windstorm or heavy snow load. Change detection is more reliable when the signal is analysed over a longer period to capture changes more reliably, however timeliness of detection is another important feature that should be pursued.

In the project, forest management activities will be monitored in dedicated areas and we will utilize satellite images acquired during one calendar year.

Several candidate test sites of size 50×50 km2 will be selected in Central Finland using information available to the consortium from Forestry Centres and forest companies.

Detection of forest disturbance will be performed using optical (Sentinel-2) and SAR (Sentinel-1) images, as well as their combination. The case of forest disturbance monitoring is very distinct from mapping forest variables use-case (included in use-case 5 Forest biomass estimation), where mostly continuous variables are predicted using regression models.

In this figure is presented one of the study sites in Finland, with multitemporal Sentinel-1 images and reference data on snow damage.

The purpose of this use case is to automate the generation of land cover maps. In fact, currently, these maps are generated by non-automatic methods, employing a great amount of resource in terms of manual work, or semi-automatic methods, which still need significant human intervention to correct and/or integrate the results obtained, using photointerpretation.

In order to have a pretrained model with general feature extractors, the model pre-training phase needs to take in input a wide range of areas, in terms of land cover use, acquisition time and even geographical areas. Particularly important is the last point, since, in the future, the interest in land cover classification outside Europe, possibly worldwide, will increase, and it will make it possible that the learned feature extractors will also able to characterize different type of crops, buildings or streets, each of which could appear different depending on the location.

In this figure is presented a sample of the landcover map for the Niassa-Selous park, that covers a large area between Mozambique and Tanzania. Specifically, on the left is shown a case of landcover change between the year 2000-2016 and 2016-2019, while on the right the pure landcover map, composed by many different classes, reported in the legend below.

Exploiting the whole ESA Sentinel-1 mission archive, in the next future through Copernicus European Ground Motion Service (EGMS) billions of ground displacement measurements will be available over Europe in correspondence of each single infrastructure, building, rural areas and terrain with scares vegetation.

In the project, the great amount of interferometric stacks that Sentinel-1 is acquiring will be exploited in a non‑supervised way. Thanks to the fact that e‑geos already possesses a very advanced DIFSAR processing chainit will be relatively easy to obtain huge amounts of ground displacement measurements on quite wide areas and relatively long time periods.

In this figure, the redder and bluer dots represent the points with greater displacement. As per the analysis, it is observed the building displacement anomaly. In this case, the building collapsed on Sept. 24th, 2016.

At any given time, it is estimated that more than 60% of land surfaces worldwide are covered with clouds. Many optical images contain thin or thick clouds, and typically only those with less than e.g. 10% cloud cover are used in EO processing chains, leaving a great proportion of unused images. Deep learning methods that outperform classical cloud detection and removal methods are mostly supervised. This use case leverages unsupervised approaches for thin cloud removal and cloud detection, to increase the usage percentage of optical images, in particular Sentinel-2, and facilitate land cover monitoring methods.

In this figure is presented one of the cloud removal dataset used in use case 4, globally distributed. Example images are from the test set, in several locations, showing a cloudy image, cloud-free image and result of a supervised cloud removal method.

This use case concentrates on estimating forestry variables (structural parameters) using satellite image data (Sentinel-1 and Sentinel-2) and a limited set of in-situ measurements (forest plots). 

These variables include forest height, growing stock volume, forest above ground biomass, diameter at breast height, basal area, tree species (plant functional type). This satellite image-based forest inventory information enables production of forest biomass and carbon estimates in boreal forest.

In this figure is preseented one of the study sites in Finland for forest variable estimation (here height and biomass), using a Sentinel-1 time series and a Sentinel-2 image (summer).

Extension of RepreSent for Scaling-up (RepreSent CCN) will considers three use cases for which a number of users have shown their interest. These use cases are:

  • USE CASE 1: Multi-temporal and multi-sensor forest mapping
  • USE CASE 2: Building anomaly detection
  • USE CASE 3: Cloud detection