+ de critères

Testing satellite detention methods

Eric Dufrêne (CNRS) and Kamel Soudani (Université Paris Sud) take the floor

Using field observation data to test remote sensing tools: satellite detection of bud burst and leaf senescence

Summary

Plant phenology is the study of periodical events such as bud burst, flushing, senescence and leaf fall. Climate is the governing factor here and phenology is the first visual indicator of variations in climate.
Phenology is a precision science which necessitates exhaustive sampling in order to take into account the diversity of the many plant species and the variability of the climate. Field observations, while necessary, are laborious and do not guarantee large-scale spatial representation; consequently, they are difficult to extrapolate to broader scales.

Remote sensing by satellite has often been presented as an alternative. The first such observations began several decades ago thanks to daily images provided by the AVHRR radiation-detection imager aboard U.S. NOAA (National Oceanic and Atmospheric Administration) satellites. The estimates resulting from this new source of data, though subject to high levels of uncertainty, made it possible to develop the first global maps of phenological events taking place in the different terrestrial biomes.
In the last fifteen years, significant improvements have been made thanks to the Moderate Resolution Imaging Spectro-radiometer (MODIS), a payload on the TERRA and AQUA satellites (NASA, USA), which is designed to continuously track the vital signs of the terrestrial biosphere. Not only has spatial resolution improved (250 m-1 km for MODIS vs. 1 km for AVHRR), the number of spectral bands has increased (36 vs. 6 for AVHRR), GPS positioning is more precise (50 m vs. 1 to 2 km) and data quality is higher. MODIS data are used to generate (and distribute?? -ceci me semble étrange même en français: comment MODIS peut distribuer les cartes?) a variety of cartographic products at the planetary scale, notably including phenological maps with resolutions of 500 m (MCDI2Q2) and one km (MOD12Q2).

In general, phenological estimates reflect temporal dynamics based on certain indices, called vegetation indices, which are sensitive to plant biomass. It should be kept in mind that these indices can only distinguish changes in foliage, and only when those changes are well marked. The Normalized Difference Vegetation Index (NDVI) is the most commonly-used spectral index. It takes advantage of the strong contrast between the amount of near infrared (NIR) and red (RED) light waves reflected by vegetation, a difference due to plants' high absorption of light in the red wavelength. The index is noted as follows: NDVI = (NIR - RED) / (NIR + RED). The figure below shows the changes over one year in NDVI recorded in situ and the main phenological phases for foliage.
Intra-annual dynamics of NDVI on a deciduous forest. The squares corresponding to the NDVI measured. (d1, d2, d3) and (s1, s2, s3) are the phenological indicators of the theoretical curve (red). They correspond to the transition dates of bud burst and spring flushing and senescence in autumn. The blue curve detect these transitions
Intra-annual dynamics of NDVI on a deciduous forest. The squares corresponding to the NDVI measured. (d1, d2, d3) and (s1, s2, s3) are the phenological indicators of the theoretical curve (red). They correspond to the transition dates of bud burst and spring flushing and senescence in autumn. The blue curve detect these transitions © Kamel Soudani / Université Paris-Sud

Studies testing the accuracy of the estimated dating of phenological events based on data from the MODIS spectro-radiometer have mainly relied on field observations recorded within the framework of the RENECOFOR network (Soudani et al. 2008; Hmimina et al. 2013; Testa et al., submitted).
Results show that the springtime inflection point on the NDVI curve (d2) is the best marker of the observed bud burst date. Prediction error is 8.5 days with a slight positive bias of 3.5 days. However, the MOD12Q2 programme provides estimates which differ considerably from observations; uncertainty reaches 20 days with a negative bias of 17 days. Leaf bud burst and flushing occur relatively quickly (20 to 30 days between bud burst and the maximum foliage index value for beech and oak in temperate forests). The senescence phase, on the other hand, is more gradual, and also coincides with the return of unfavourable autumn weather conditions which make remote sensing less accurate. Here, prediction error is 20 days in the best case scenario (Testa et al., in press).

In short, the observation data acquired by the RENECOFOR network through long-term monitoring and with standardised protocols offer an unequalled opportunity to validate and calibrate remote sensing products. This calibration has become even more crucial since the launch of new satellites for the European Space Agency's (ESA) Sentinel Mission, which combines very high-resolution spatial and temporal remote-sensing, especially dedicated to observing the planet Earth and to continuously providing data for a variety of actors. The degree of uncertainty for the estimates furnished must be evaluated to avoid erroneous interpretations and to define the limits of the use of remote-sensing (Soudani & François, 2014).

Pour en savoir plus

Ressources