Enhancing potato leaf protein content, carbon-based constituents, and leaf area index monitoring using radiative transfer model and deep learning

Abstract

Accurate determination of potato leaf protein content (Cp), carbon-based constituents (CBC), and leaf area index (LAI) are crucial for precise monitoring of potato growth. Dynamic monitoring of leaf Cp, CBC, and LAI can provide valuable insights for agricultural management, such as analyzing the impact of environment stress factors on potato growth throughout its lifecycle. Currently, the most commonly used method for estimating crop parameters is the vegetation spectral feature-statistical regression approach. However, leaf Cp and CBC estimation are greatly influenced by water absorptions, as they exhibited overlapping spectral features in the short-wave infrared (SWIR) region. Consequently, the accuracy of protein estimation using traditional vegetation spectral feature-statistical regression methods is limited. This study aims to propose a comprehensive approach called PCPNet (Potato Canopy and Leaf Parameter Network), which could jointly estimate potato canopy and leaf parameters including Cp, CBC, and LAI. The performance of the PCPNet was compared with traditional spectral feature-statistical regression methods in estimating Cp, CBC and LAI. A simulated dataset for pre-training was generated using the PROSPECT-PRO and SAIL radiative transfer models to represent various complex scenarios encountered in real-world potato cultivation practices. The designed PCPNet was initially pre-trained based on this simulated dataset and then re-trained using ground-based measurements from five potato growing seasons across two distinct regions in China through transfer learning techniques. The validation of potato canopy and leaf parameters was conducted based on the estimations provided by the PCPNet model, while assessing their accuracy. This study yields the following results: (1) The PCPNet-based deep learning model demonstrated markedly superior accuracy in estimating potato Cp, CBC, and LAI compared to traditional machine learning models. (2) The deep learning model pretrained with transfer learning exhibited greater estimation accuracy than the deep learning model trained from scratch. In future research, experiments should be conducted across multiple regions and crops to verify both accuracy and generalizability of this approach in remote sensing of leaf Cp, CBC, and LAI.