The potential of Sentinel-2 spatial resolution for LAI derivation in Alpine grasslands

Kurzfassung

Leaf Area Index (LAI) is a key parameter of vegetation structure and particularly important for quantifying the exchange fluxes of water, energy, and gases between land surface and atmosphere. As high temporal resolution is crucial for monitoring such a variable vegetation parameter, so far mainly medium resolution data have been used for automated LAI derivation procedures. However, with the upcoming Sentinel 2 mission the scientific community is moving towards exploiting new generation sensors that combine high spatial and temporal resolution. These systems potentially offer advancements for ecosystem monitoring, especially when dealing with challenging landscapes such as the Alpine region with its small-scale land use pattern and temporal highly dynamic grasslands. As RapidEye data can be scaled to the Sentinel-2 spatial resolution of approximately 20 m and also has a red edge channel, analyses based on this sensor can serve for evaluating the suitability of Sentinel-2 data for vegetation assessment. Recently, Pasolli et al. (2011) applied an algorithm based on inverted radiation transfer modeling (RTM) to MODIS imagery to derive regionally adapted LAI estimates over Alpine grasslands with improved performance with respect to standard LAI products. This approach is now applied to MODIS time series (250 m spatial resolution) as well as to RapidEye images (6.5 m) acquired over the TERENO site in the Bavarian alpine upland. Additionally, intermediate spatial resolution data (19.5 m, 32.5 m, 45.5 m, and 97.5 m) are simulated based on the RapidEye imagery and used for LAI derivation. By comparing LAI estimates derived at different spatial resolutions in the complex mountainous environment the influence of spatial resolution on LAI derivations can be investigated. Four RapidEye scenes from May 9, May 25, July 16, and September 6, 2011 are available. The data have five bands in the visible, red edge and near infrared domain. During preprocessing geometric correction, orthorectification, and atmospheric correction were performed (Richter and Schläpfer, 2012). Grassland masks were derived based on a random forest land cover classification. RTMs simulate the interactions between solar radiation and canopy elements. The model is characterized in a first step by its leaf and canopy variables. For the RapidEye scale, the PROSAIL model (Jacquemoud et al. 2009) is parameterized using field data, literature values, topography information, and the RapidEye scene geometries. After calculating the reflectances of multiple canopy realizations the model is inverted, i.e. the parameter set (including LAI) which produced the most similar reflectances is selected for each pixel of the remote sensing data. Thereby, the local topography of each pixel is taken into account by selecting the appropriate viewing geometry for the model runs. For the inversion, look up tables and a cost function based on the normalized mean squared error are applied. The MODIS LAI time series was derived according to the algorithm described in Pasolli et al. (2011) using the 250 m MODIS Terra reflectance in the PROSAIL model inversion. Results were validated with in situ LAI measurements which have been collected during the four corresponding weeks of the RapidEye data acquisitions. At in total 88 locations LAI was sampled using a LICOR LAI-2000 PCA and corrected using 15 destructive LAI samples from the same sites and dates. The measured LAI range covered values from 1.5 to 7.4. The resulting high resolution LAI maps fit the in situ LAI with an average RMSE of 0.89. The spatial LAI patterns of growth stages of individual meadows within the scenes are clearly captured. The 8-day MODIS LAI time series reflects the temporal pattern of green-up, senescence, and sudden LAI reduction caused by mowing satisfactorily. Considering the MODIS resolution and the heterogeneity of the area, this algorithm achieved good results in comparison with in situ LAI (RMSE of 1.36). Although the MODIS LAI tends to overestimate the in situ LAI and does not catch the variability of all plots, this time series information could be used for a combination with high resolution time series in future work. To analyze systematically the role of the data’s spatial resolution on the LAI retrieval accuracy, synthetic datasets with intermediate resolutions were created from the RapidEye scenes by cubic convolution. The simulated reflectances were then used as input for the PROSAIL model inversion for each time step. The resulting LAI maps show mostly increasing overall RMSEs of 0.99 (19.5 m resolution), 0.92 (32.5 m), 1.01 (45.5 m), and 1.13 (97.5 m) with coarser resolutions. Although the 32.5 m resolution LAI has a slightly higher accuracy than the 19.5m scale LAI, the range of RMSEs for the individual dates is considerably smaller at the 19.5 m scale. With regard to the persistence of spatial patterns, the 19.5 m and 32.5 m resolution maps provide a sufficient level of detail for the specific landscape. Although small fields are no longer recognizable on the 45.5 m scale, the LAI values range is still captured, i.e. the distinction of rough growth stages is possible for larger meadows. The 97.5 m data show only few similar patterns compared to the original data, as do the MODIS LAI data. Based on these results, a scaling effect with regard to the recognition of spatial patterns can be observed for the synthetic 45.5 m and less resolution data. The accuracy achieved using 20 m resolution data indicates a high suitability of future Sentinel-2 data for grassland LAI mapping in Alpine environments. Further, the additional Sentinel-2 bands in the red edge and SWIR spectrum promise higher accuracies as they might increase the RTM inversion stability. The expected revisit time of Sentinel 2 could further enable the mapping of sudden LAI reductions not provided by present high spatial resolution sensors.