Comparative performance of Sentinel-2 MSI and Landsat-8 OLI data in canopy cover prediction using Random Forest model: Comparing model performance and tuning parameters

Quantifying canopy cover using Random Forest (RF) model's appropriate tuning parameters value and sensor based predictor variables is always challenging, especially in fragmented dry deciduous forests. Therefore, this study was designed to compare the performances of Sentinel-2 and Landsat-8 based models using the RF model for predicting canopy cover with assessing variables' relative importance and correlation. Sentinel-2 and Landsat-8 based bands and spectral indices were used as predictor variables. We compared different mtry, ntree and bag fraction values of the RF model. R-square (R-2) and root mean square error (RMSE) were used for comparing the model performance. The results showed that the lowest RMSE value was associated with the default value (predictors/3) or more than the default value of mtry, with bag fraction 0.3-0.7 for Sentinel-2 and 0.3-0.4 for Landsat-8. Model accuracy has increased and stabilized with increase of ntree, and received the lowest RMSE to ntree of more than 1000. Except for SWIR indices based model of Landsat-8, all other Landsat-8 based model's accuracy was lesser compared to Sentinel-2 based models. Model accuracy of Sentinel-2 based full model (except red edge indices) was marginally better (R-2 = 0.899, RMSE = 6.883 %) than Landsat-8 based full model (R-2 = 0.886, RMSE = 7.089 %). But with the incorporation of red edge indices, full model RMSE had decreased further from 6.883 % to 6.747 %, and R-2 had increased from 0.899 to 0.918. The full model of Sentinel-2 tended to spread variable importance among more variables, but the full model of Landsat-8 slightly tends to concentrate variable importance with fewer variables. However, SWIR bands and indices were the most important predictor variables and highly correlated with canopy cover. These findings can solve the parameter value choice of RF model, and the use of the Sentinel-2 based model will be superior to Landsat-8 based model. (c) 2023 COSPAR. Published by Elsevier B.V. All rights reserved.