Novel spectral indices and transfer learning model in estimat moisture status across winter wheat and summer maize

Abstract

Timely and accurate estimation of crop moisture status is important for understanding crop growth and development. It also provides guidance for irrigation strategies and precision management. The development of UAV remote sensing technology makes it possible to quickly and accurately estimate crop moisture status. This study focuses on evaluating the performance of multispectral (MS) and hyperspectral (HS) data from UAVs. It explores their potential for predicting the Fuel Moisture Content (FMC) of winter wheat and maize. Additionally, it examines the robustness of FMC prediction across different crop types. Spectral data were collected using UAV platforms equipped with MS and HS sensors. The data were gathered during the flowering and filling stages of winter wheat, and the tasseling and silking stages of summer maize. Eleven common MS indices were constructed. Additionally, nine novel HS indices were developed by analyzing 300 hyperspectral bands. FMC prediction models were built using four algorithms: Random Forest (RF), Gaussian Process (GP), Multilayer Perceptron (MLP), and Bidirectional Recurrent Neural Network (BRNN). For the transfer learning prediction, the summer maize dataset was used as a calibration set, and a certain amount of winter wheat data was added. The results showed that for the novel two-band hyperspectral indices, the FMC of both winter wheat and maize was most sensitive to bands in the red wavelength range. For the novel three-band indices, a wider range of bands showed a stronger relationship with FMC in both crops. New samples from the grain-filling stage of winter wheat MS and HS integrated datasets were added to the silking stage dataset of summer maize. The BRNN model achieved the best transfer performance when 16 winter wheat samples were added (R² = 0.638, RMSE = 0.033). These results suggest that the BRNN model, combined with transfer learning, can provide robust FMC predictions across different crop types. This approach supports the monitoring of crop water status.