Assessment of Workflow Feature Selection on Forest LAI Prediction with Sentinel-2A MSI, Landsat 7 ETM+ and Landsat 8 OLI

Abstract

The European Space Agency (ESA)’s Sentinel-2A (S2A) mission is providing time series that allow the characterisation of dynamic vegetation, especially when combined with the National Aeronautics and Space Administration (NASA)/United States Geological Survey (USGS) Landsat 7 (L7) and Landsat 8 (L8) missions. Hybrid retrieval workflows combining non-parametric Machine Learning Regression Algorithms (MLRAs) and vegetation Radiative Transfer Models (RTMs) were proposed as fast and accurate methods to infer biophysical parameters such as Leaf Area Index (LAI) from these data streams. However, the exact design of optimal retrieval workflows is rarely discussed. In this study, the impact of five retrieval workflow features on LAI prediction performance of MultiSpectral Instrument (MSI), Enhanced Thematic Mapper Plus (ETM+) and Operational Land Imager (OLI) observations was analysed over a Dutch beech forest site for a one-year period. The retrieval workflow features were the (1) addition of prior knowledge of leaf chemistry (two alternatives), (2) the choice of RTM (two alternatives), (3) the addition of Gaussian noise to RTM produced training data (four and five alternatives), (4) possibility of using Sun Zenith Angle (SZA) as an additional MLRA training feature (two alternatives), and (5) the choice of MLRA (six alternatives). The features were varied in a full grid resulting in 960 inversion models in order to find the overall impact on performance as well as possible interactions among the features. A combination of a Terrestrial Laser Scanning (TLS) time series with litter-trap derived LAI served as independent validation. The addition of absolute noise had the most significant impact on prediction performance. It improved the median prediction Root Mean Square Error (RMSE) by 1.08 m2 m−2 when 5% noise was added compared to inversions with 0% absolute noise. The choice of the MLRA was second most important in terms of median prediction performance, which differed by 0.52 m2 m−2 between the best and worst model. The best inversion model achieved an RMSE of 0.91 m2 m−2 and explained 84.9% of the variance of the reference time series. The results underline the need to explicitly describe the used noise model in future studies. Similar studies should be conducted in other study areas, both forest and crop systems, in order to test the noise model as an integral part of hybrid retrieval workflows.

Exploring the Impact of Noise on Hybrid Inversion of PROSAIL RTM on Sentinel-2 Data

“…The results of our sensitivity analysis provided clear insights into which biophysical variables can be measured using S2 data (see Figure 5 and Table 3), and furthermore, provided insights into the effects of their interaction with spectral responses. Our results on the band correlation (Table 3) and sensitivity analysis ( Figure 5) are in line with previous results [52], and in particular, when compared with recent results by Brede et al [27] also on S2 data, and Gu et al [49] on Landsat TM data. Similarly, both studies found that C ab was associated with the visible range in Landsat TM bands, and that C m , C w and LAI was associated with NIR and SWIR bands (see Figure 5).…”

“…Model performance was severely affected even at low levels of noise (Figures 8 and 9). This is of particular concern given that every data has noise, and previous research has shown that the Sen2Cor atmospheric correction algorithm produces between 5% and 10% noise [27,52]. Moreover, and in contrast to results obtained on pure data, both neural networks generally performed better than GPR when noise was added, which indicates that further developments with GPR are needed to be able to deal effectively with noisy observations.…”

“…These effects are decomposed into first-order, interaction and total-order sensitivity indices that represent the variability associated with single features, interactions between features, and the sum of all interactions, respectively. This allowed us to explore which S2 bands are associated with which biophysical trait and if there are significant confounding factors between them, and is an analysis comparable to what was performed by Gu et al [49] and Brede et al [27]. Since the EFAST provides an estimate of the sensitivity, a conservative number of samples necessary to provide an accurate estimate must be identified first by calculating the index with increasing number of samples until the estimate stabilizes.…”

“…This problem is further amplified when the number of bands is limited or by the presence of noise [24,25]. Furthermore, the presence of noise in satellite data caused by factors such as the atmosphere makes the challenge increasingly complex [25][26][27]. Active learning methods which integrate new samples based on uncertainty or diversity criteria to improve the accuracy of the model in the context of Earth observation regression problems [28], have been shown to enhance retrieval accuracy, which may be due to mitigation of the ill-posedness within hybrid approaches [10], in particular when field data is available [29].…”

“…Thus, there is a need to identify which algorithm generally performs better irrespective of sensor and location. While various explorations exist [26,27,30], this question has so far not been addressed. Therefore, the aim of this study is to identify the best Hybrid approach for estimating biophysical parameters of vegetation, irrespective of geographic location or time, and with a reasonable robustness in treating the effects of noise.…”