Evaluating the Impact of Spatial Heterogeneity on the Prosail Model and Lai Inversion

The PROSAIL (PROSPECT + SAIL) model has been utilized in the retrieval of many vegetation biophysical and biochemical variables such as the leaf area index (LAI) and the leaf chlorophyll content (LCC). However, less attention is paid to the impact of spatial heterogeneity on the PROSAIL model. This paper uses a linear-mixing model to simulate within-pixel spatial heterogeneity. Simulated spectra are resampled to Sentinel-2's bands, and several common spectral indices are calculated. Results show that spectral indices calculated from mixed spectra can deviate up to 10% compared to those calculated from pure spectra when spatial heterogeneity comes to 0.5 (measured in Coefficient of Variation, CV). We further use four widely used machine learning methods for LAI inversion, i.e., Supporting Vector Regression (SVR), Random Forest (RF), Gradient Boosting Decision Tree (GBDT) and Neural Network (NN). Concerning R-square and RMSE, all models trained with pure spectra show performance decline when tested on mixed spectra, which can be partially mitigated by using mixed spectra for model training.