A hybrid water cloud model for leaf area index estimation using machine learning and physically based model derived Sentinel-2 vegetation descriptors

Abstract

Monitoring crop health effectively requires robust and continuous estimation of Leaf Area Index (LAI). This study reports on the development of a hybrid Water Cloud Model (WCM) in which machine learning and physically based model derived Sentinel-2 vegetation descriptors are used to estimate LAI. Among the vegetation descriptors used in the WCM, Sentinel-2 optical vegetation indices (VIs) performed best, with LAI derived using WDVI and NDVI showing good correlations with measured LAI (R ≈ 0.7–0.8). In contrast, LAI derived from the WCM using Sentinel-1 SAR parameters as vegetation descriptors yielded a moderate correlation with measured LAI (R ≈ 0.6). Though Sentinel-2 vegetation indices perform well in LAI prediction, they are limited by cloud-induced optical data gaps. To address these limitations, Sentinel-2 VIs were predicted from Sentinel-1 SAR data using machine learning algorithms and used as vegetation descriptors in the hybrid WCM. Different machine learning algorithms were employed to derive the Sentinel-2 vegetation descriptors from Sentinel-1 data. The LAI estimated from the hybrid WCM using the machine learning algorithm derived vegetation descriptors achieved reliable correlations with measured LAI (R ≈ 0.7–0.8). Alternatively, the Sentinel-2 VIs were also derived from the physically based model PROSAIL to be utilized as vegetation descriptors in the hybrid WCM for LAI prediction. The physically based model derived indices provided comparable accuracy (R ≈ 0.7–0.77) as the machine learning approach for LAI estimation using the hybrid WCM model. These results demonstrate, for the first time, that machine learning-predicted and physically based model simulated optical indices can serve as effective substitutes for actual Sentinel-2 vegetation descriptors in the WCM, enabling consistent, all-weather LAI estimation and monitoring when Sentinel-2 data is not available due to persistent cloud cover and adverse weather conditions. These approaches provide a scalable framework for reliable crop growth assessment and precision agriculture applications.