Crop specific algorithms trained over ground measurements provide the best performance for GAI and fAPAR estimates from Landsat-8 observations

Abstract Estimation of Green Area Index (GAI) and fraction of Absorbed Photosynthetically Active Radiation (fAPAR) from decametric satellites was investigated in this study using a large database of ground measurements over croplands. It covers six main crop types including rice, corn, wheat and barley, sunflower, soybean and other types of crops. Ground measurements were completed using either digital hemispherical cameras, LAI-2000 or AccuPAR devices over sites representative of a decametric pixel. Sites were spread over the globe and the data collected at several growth stages concurrently to the acquisition of Landsat-8 images. Several machine learning techniques were investigated to retrieve GAI and fAPAR from the Landsat-8 top of canopy reflectance values, either using empirical or simulated calibration databases and generic or crop-type specific algorithms. Results show that using the six Landsat-8 bands together provided the best estimates of GAI and fAPAR. Machine learning techniques trained over a dataset simulated by the PROSAIL model provided less accurate estimates of GAI and fAPAR as compared to machine learning techniques trained over the ground data collected and applied over a similar dataset (best-case scenario). All machine learning techniques performed similarly when calibrated on PROSAIL simulations. The Gaussian process regression (GPR) was the best performing machine learning technique as compared to artificial neural networks (ANN), support vector machine regression (SVM) and as compared to NDVI simple model when calibrated over the ground dataset in the best-case scenario. However, the performance of the GPR trained over ground data is notably degraded when applied to crop types excluded from the calibration (worst-case scenario), and models based on simulations performs similarly of even better. Furthermore, training the GPR over specific crop types was performing slightly better than training over all the crop types together if at least 100 well distributed data points were available in the training dataset. Similar conclusions were obtained for the other machine learning techniques, with crop specific empirical models providing slightly better performances with a few exceptions. Finally, GAI was estimated by GPR with a RMSE varying between 0.45 and 1.19 and fAPAR with RMSE varying between 0.07 and 0.15 depending on the crop type.