Exploring the transferability of wheat nitrogen status estimation with multisource data and Evolutionary Algorithm-Deep Learning (EA-DL) framework

Abstract

Accurate and transferable wheat nitrogen status estimation is very important to plant phenotyping and smart agricultural management. The goal of this study was to establish a wheat nitrogen status estimation model across all growth stages by combining proximal sensing and meteorological data. From 2010–2020, nine multi-nitrogen rates field trials were conducted at five sites involving different wheat varieties. Proximal sensing data were acquired from a Crop Circle sensor at key growth stages and meteorological data were aggregated from planting to the corresponding sensing date. Deep neural network (DNN) and long short-term memory (LSTM) were adopted to estimate above-ground biomass, plant nitrogen uptake, plant nitrogen concentration, and the nitrogen nutrition index. Random forest (RF) was used as a benchmark regression model. Multi-task learning (MTL) based on DNN was conducted to estimate the four nitrogen indicators simultaneously. A genetic algorithm (GA) was tested to optimize the hyperparameters, connection weights, and loss function weights (for MTL) of neural networks separately. The results revealed that DNN (R2 =0.83–0.96) and MTL (R2 =0.81–0.96) achieved an overall comparable high accuracy with RF (R2 =0.83–0.97), whereas LSTM (R2 =0.76–0.93) did not improve the nitrogen status estimation in our dataset. This study presented a concise and efficient framework dedicated to exploring the transferability of phenotypic predictions and provided insights into understanding crop growth and nitrogen dynamics in response to environmental conditions.