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

Abstract

Remote sensing (RS) of biophysical variables plays a vital role in providing the information necessary for understanding spatio-temporal dynamics in ecosystems. The hybrid approach to retrieve biophysical variables from RS by combining Machine Learning (ML) algorithms with surrogate data generated by Radiative Transfer Models (RTM). The susceptibility of the ill-posed solutions to noise currently constrains further application of hybrid approaches. Here, we explored how noise affects the performance of ML algorithms for biophysical trait retrieval. We focused on synthetic Sentinel-2 (S2) data generated using the PROSAIL RTM and four commonly applied ML algorithms: Gaussian Processes (GPR), Random Forests (RFR), and Artificial Neural Networks (ANN) and Multi-task Neural Networks (MTN). After identifying which biophysical variables can be retrieved from S2 using a Global Sensitivity Analysis, we evaluated the performance loss of each algorithm using the Mean Absolute Percentage Error (MAPE) with increasing noise levels. We found that, for S2 data, Carotenoid concentrations are uniquely dependent on band 2, Chlorophyll is almost exclusively dependent on the visible ranges, and Leaf Area Index, water, and dry matter contents are mostly dependent on infrared bands. Without added noise, GPR was the best algorithm (<0.05%), followed by the MTN (<3%) and ANN (<5%), with the RFR performing very poorly (<50%). The addition of noise critically affected the performance of all algorithms (>20%) even at low levels of added noise (≈5%). Overall, both neural networks performed significantly better than GPR and RFR when noise was added with the MTN being slightly better when compared to the ANN. Our results imply that the performance of the commonly used algorithms in hybrid-RTM inversion are pervasively sensitive to noise. The implication is that more advanced models or approaches are necessary to minimize the impact of noise to improve near real-time and accurate RS monitoring of biophysical trait retrieval.

Gaussian processes retrieval of crop traits in Google Earth Engine based on Sentinel-2 top-of-atmosphere data

“…GPR ( Rasmussen and Williams, 2006 ) was chosen as core machine learning algorithm in the hybrid retrieval scheme since it has proven good to excellent performances in numerous studies (e.g., Brede et al, 2020 ; Mateo-Sanchis et al, 2021 ; de Sá et al, 2021 ; Xie et al, 2021 ; Verrelst et al, 2012a , 2020 ; Zhou et al, 2018 ). In the context of recent vegetation traits mapping activities from EO data, these Bayesian non-parametric approaches are to be found among the preferred regression models ( Camps-Valls et al, 2018 ).…”

Prototyping Crop Traits Retrieval Models for CHIME: Dimensionality Reduction Strategies Applied to PRISMA Data

“…The rationale is that simulated data is overly perfect as opposed to image data where noise is always present for multiple reasons, e.g., due to sensor electronics and optics or poor geometric, radiometric, or atmospheric corrections. Adding noise to the synthetic training data may also support accounting for variability present on the surface, e.g., due to sub-pixel heterogeneity [ 19 , 26 , 92 ]. It must also be remarked, however, that the optimized sampling through AL techniques largely surpasses the need for adding noise, as was observed in recent active learning studies [ 12 , 21 ].…”

“…This holds, in particular, true for the estimation of leaf-level traits, where additional work is needed to provide optimized retrieval models. Ideally, the in situ data set covers a broad range of vegetation types collected during multiple phenological stages in combination with spectral data and corresponding uncertainty information of the measurements [ 14 , 24 , 26 ]. A further critical issue when employing AL is the optimal timing at which learning should be stopped, i.e., the stopping criterion [ 94 ].…”

“…An overview and elaborated taxonomy of variable retrieval methods from Earth observation data is provided by Verrelst et al [ 16 , 17 ]. From the main families of retrieval methods, i.e., (1) parametric regressions, (2) nonparametric regressions, (3) physically-based methods, and (4) hybrid approaches, the last method evolved as the most appealing in operational contexts [ 18 – 26 ]. Hybrid strategies blend the physics described by radiative transfer models (RTM) and use the efficiency of machine learning regression algorithms (MLRAs) in a synergistic way to infer the traits of interest.…”

Quantifying mangrove leaf area index from Sentinel-2 imagery using hybrid models and active learning

“…Despite the advantages of hybrid methods (see Verrelst et al 2019), the inversion of RTMs using real-world spectral data, like our Sentinel-2 observations, remains ill-posed; multiple configurations of leaves and/or canopy variables can produce identical or similar spectral responses (Verrelst et al 2016a). This problem is further amplified when the number of bands is limited or by the presence of noise (de Sá et al 2021). Commonly, illposedness is minimized through a reasonable pre-selection of expected biophysical trait range of values, but this has an obvious implication on the generality and scalability of the trained models (Verrelst et al 2019(Verrelst et al , 2015.…”