PNNL

Abstract

Leaf Area Index (LAI) is a critical indicator in agriculture, yet its accurate estimation from high resolution satellite imagery like PlanetScope (3-m) is challenging due to limited pixel-level field data. To bridge this scale gap and enable large-scale monitoring, this study introduces LAI-TransNet, a novel two-stage transfer learning framework for soybean LAI prediction from UAV-scale field data to PlanetScope imagery. Stage 1 establishes a UAV imagery LAI estimation benchmark using PROSAIL-simulated UAV reflectance (UAV-Sim) and field-measured LAI. Through training and evaluation of various models, including traditional ML (Random Forest, XGBoost, LightGBM), deep learning (CNN, MLP, Transformer), and transfer learning (CNN-TF, MLP-TF, Transformer-TF), the CNN-TF model achieved the highest accuracy (R² = 0.81, RMSE = 0.64 m²/m², rRMSE = 11.47%), serving as the UAV benchmark for cross-scale validation. Stage 2 refines the model using simulated PlanetScope data (PS-Sim) corrected with Stage 1 data (UAV-Sim), involving preprocessing via cross-domain mapping and feature enhancement. Three deep learning models (CNN, MLP, Transformer) are trained directly on this preprocessed PS-Sim. Concurrently, LAI-TransNet is developed via secondary transfer learning from the Stage 1 CNN-TF model onto the preprocessed PS-Sim. LAI predictions from all four models on real PlanetScope imagery are then compared to the Stage 1 UAV benchmark to assess consistency. LAI-TransNet yields a cross-scale R² of 0.69, demonstrating superior performance over the other deep learning models evaluated in this stage (R² values: 0.62, 0.63, 0.60). By effectively bridging UAV and satellite scales, LAI-TransNet shows superior accuracy and scalability for large-scale crop LAI monitoring, providing a robust solution for precision agriculture using PlanetScope data.