Remote Sensing

The growing number of earth observation satellite missions and the corresponding abundance of image data enable us to monitor any given spot on the Earth’s surface. This makes remote sensing an ideal tool for gaining insights into large, distributed, and inaccessible areas with standardized workflows. Two major developments have boosted the availability of remote sensing data in the preceding decade:

• Within the Copernicus program funded by the EU and implemented by ESA several earth observation satellite missions (with the Sentinel satellites as flagships) were launched with a focus on continuous, global data acquisition. Thanks to an Open Data policy, all data are publicly available free of charge without any usage restrictions. Together with accompanying data products, third-party contributing missions, as well as the long-term open satellite program Landsat by NASA, the spatial and temporal coverage of freely available satellite imagery, has never been greater.

• Advances in sensor and spacecraft development have made it possible to equip very small satellites with sophisticated imaging technology. Commercial providers have launched fleets of tens to hundreds of such SmallSats (e.g. PlanetScope, ICEYE) which are able to achieve global coverage within days or weeks while delivering very high-resolution (VHR) imagery in the meter-to-submeter domain.

Spatial resolution 

It describes the smallest feature on the earth’s surface that can be detected by a satellite sensor. This effectively translates to the pixel size of a remote sensing image, which is the edge length of a pixel in meters on the ground. Free and open data has a medium resolution, with a 10 m pixel size of Sentinel-1 and Sentinel-2, and 30 m of Landsat. Commercial satellites can reach high resolution (HR), e.g. 4 m for PlanetScope, and very high resolution (VHR) , e.g. up to 0.5 m (VHR) for Pléiades. For an analysis that targets e.g. individual trees, a resolution in the HR/VHR domain is thus preferable while for monitoring larger forest plots medium resolution imagery is usually sufficient.

Temporal resolution 

It refers to the minimum time period between individual acquisitions of the same area in days or weeks. Since earth observation satellites typically fly in a polar orbit, the revisit time is shorter towards the poles and longer at the equator. The open data missions Sentinel-1/-2 and Landsat are acquiring imagery continuously and globally (the Sentinels since 2014/2015 and the Landsat program since the 1980s) and have revisit times of three to ten days. Most commercial sensors have a theoretical revisit time of one or two days, however, they do not operate in a continuous monitoring mode but are tasked to acquire imagery of a specific area. This means that image acquisition can be planned beforehand, but data availability in the archives may be sparse. An exception is the PlanetScope program, which is a fleet of commercial satellites that acquire global imagery coverage daily. For analyses that focus on long-term trends, dense imagery time series and thus continuously acquiring sensors are the preferred data sources. For assessing the status at a specific point in time, commercial data may be a valid option too, but data availability is more limited further back in time. In practice, acquiring HR/VHR commercial data for a given point in time in the past (e.g. for EUDR relevant December 2020) is conducted on a best-effort basis and trade-offs in spatial resolution and/or temporal exactness must be expected.

Imaging technique

For vegetation mapping, the most relevant analysis methods rely on data from multispectral or SAR (snythetic-aperture-radar) sensors. Multispectral sensors acquire images in different wavelengths of the electromagnetic spectrum, typically in the human visible light range red, green, and blue, as well as near-infrared, which is very sensitive towards the chlorophyll content and thus the health of vegetation. Therefore, multispectral imagery is well suited to differentiate between vegetated and non-vegetated areas and to assess the health of plants. SAR sensors on the other hand operate in the microwave domain, which makes their imagery harder to interpret in the vegetation context as it is mostly the physical structure of a plant that affects the image intensity. However, SAR sensors have the benefit that due to their operating wavelength, they are not influenced by cloud coverage – a clear advantage, especially in tropical areas.

Cost

Lastly, satellite imagery of an area can come with very different prices, mostly depending on the spatial resolution of the sensor. While medium-resolution data is free of charge (i.e. Sentinel and Landsat data), HR data typically starts at 2-3$/km², and VHR data can be as costly as 30$/km² (for archived data – roughly the double price can be assumed if a satellite is to be tasked to acquire imagery of a specific area). Monitoring large areas with high spatial resolution can thus quickly become cost-intensive.

This section covers the analyses to map potential changes in forested areas. The analysis aims to identify forest areas that have been cleared since the EUDR relevant month of December 2020 and to show the potential and limits of different satellite imagery sources.

The general approach is a change detection: Satellite imagery from two dates is quantitatively compared to detect areas that underwent land cover change between the acquisitions. Spectral vegetation indices are suitable indicators to map forest change. They are calculated by combining information from different areas of the electromagnetic spectrum that are acquired by multispectral sensors. A vegetation index represents condensed information on plant coverage and/or health within a satellite image pixel. Hence, the decrease of a vegetation index between two acquisitions indicates a loss of vegetation in that area (however if the area of interest is subject to seasons, a decrease of a vegetation index might be caused by e.g. the yearly dry period and not an actual loss of vegetation cover. It is thus important in this case to select imagery from the same seasons, but different years as data sources).  

SAR sensors cannot be used to calculate spectral vegetation indices, however their signal contains information about roughness and vertical structures within an imaged pixel. Forest loss can thus cause a change in the SAR signal.

The image above shows a comparison of deforestation analyses carried out for medium (left) and high (middle) resolution multispectral imagery as well as for medium-resolution SAR data (right). The bottom row shows areas of vegetation loss in red, which can be automatically extracted and vectorized for further analysis and distribution. The high-resolution commercial SPOT result shows great spatial detail, but the same general pattern of vegetation loss can also be extracted from free Sentinel-2 imagery. The pattern extracted from free Sentinel-1 data differs slightly because the SAR imaging technology captures completely different properties of the earth’s surface than multispectral sensors. While the resulting deforestation map may be less precise, Sentinel-1 SAR imagery is a reliable data source with global data availability in dense time series due to its weather independence. This advantage cannot be overstated for analyses in the tropics: multispectral data availability is sparse due to cloud coverage and it cannot be assumed that cloud-free multispectral imagery for a given area is available for a certain month let alone day.

Whether a detected deforested area is caused by legal silvicultural activity or whether timber extraction was performed in a sustainable manner cannot be assessed from remote sensing data alone. Hence, the true value of the deforestation analysis described above lies in the combination of the resulting maps with further (geo)data. Possible applications include:

• Intersection with plot data: If the information on ownership or conservation status is available at a polygon plot level, violations of e.g. extraction limits can be assessed.

• Combination with land cover/vegetation data: Detected vegetation changes can be further disaggregated into start and end classes (e.g. change from tropical rainforest to palm oil plantation) if land cover information is available for all analyzed dates.. Thus, the land cover information can be extracted from the combination of remote sensing data sources as well, while the complexity of this task increases with the number of desired land cover classes.

• Quantification of biomass loss: The amount of removed AGB (Above-Ground-Biomass) can be estimated by intersecting detected deforestation areas with biomass or canopy height maps (which are usually also generated by remote sensing methods). Depending on the area of interest, regional biomass/canopy maps may be available that usually have a higher resolution and accuracy than globally available datasets.

• Detection of forest degradation: While mapping forest clearance is a relatively straight-forward exercise, detecting forest degradation (caused by e.g. selective logging or groundwater depletion) is a more complex task that requires dense time-series of satellite imagery as the changes of vegetation indices, etc. can be rather small but nevertheless trackable if such a trend remains consistent over time.