A Method for Retrieving Maize Fractional Vegetation Cover by Combining 3-D Radiative Transfer Model and Transfer Learning

Abstract

Fractional vegetation cover (FVC) is an essential parameter of vegetation canopy. Understanding its dynamics is vital for agricultural monitoring and climate change response. The physically based method for retrieving FVC from remote sensing data has great potential due to the theoretical basis of the radiative transfer model (RTM). However, the method is limited when applied to satellite imagery due to its uncertainty in simulating canopy reflectance. This article proposes a method that combines three-dimensional (3-D) RTM and convolutional neural network based transfer learning (CNN–TL) to address the inconsistency between simulated and satellite reflectance, improving maize FVC retrieval accuracy. First, 3-D RTM was employed to generate canopy reflectance datasets of maize at various growth stages. Second, CNN–TL is used to eliminate the discrepancy between 3-D RTM simulated reflectance and satellite reflectance, and the retrieval accuracy of CNN–TL is compared with random forest regression (RF) and CNN. Finally, the feasibility of the method was validated using time-series of measured data from multiple samples covering different growth stages of maize from 2021 to 2023. The results showed that, when retrieving maize FVC on GF-1, HJ-2, and Sentinel-2, CNN–TL performed the best (RMSE = 0.117, 0.063, and 0.081) compared to RF (RMSE = 0.186, 0.226, and 0.184) and CNN (RMSE = 0.133, 0.117, and 0.098). The spatial distribution of FVC maps remains highly consistent across all three satellites, indicating the exceptional performance of CNN–TL. These results contribute to the development of physically based methods for FVC retrieval and serve as a reference for multisource satellite studies.

Keywords

• Convolutional neural network (CNN)
• fractional vegetation cover (FVC)
• multispectral remote sensing
• three-dimensional radiative transfer model (3-D RTM)
• transfer learning (TL)