Field spectroscopy of canopy nitrogen concentration in temperate grasslands using a convolutional neural network

Abstract

Listen

Translate article

As an essential feature of plant autotrophy, Nitrogen (N) is the major nutrient affecting plant growth in terrestrial ecosystems, thus is of not only fundamental scientific interest, but also a crucial factor in crop productivity. Timely non-destructive monitoring of canopy nitrogen concentration (N%) demands fast and highly accurate estimation, which is often quantified using spectroscopic analyses in the 400—2500 nm spectral region. However, extracting a set of useful spectral absorption features from canopy spectra to determine N% remains challenging due to confounding canopy architecture. Deep Learning as a statistical learning technique is useful to extract biochemical information from canopy spectra. We evaluated the performance of a one-dimensional convolutional neural network (1D-CNN) and compared it with two state-of-the-art methods: partial least squares regression (PLSR) and gaussian process regression (GPR). We utilized a large and diverse in-field multi-season (autumn, winter, spring and summer) spectral database (n = 7014) over 8 years (2009–2016) of dairy and hill country farms across New Zealand to develop season specific and spectral-region specific (VNIR and/or SWIR) 1D-CNN models. Results on the independent validation dataset (not used to train the model) showed that the 1D-CNN model provided higher accuracy (R2 = 0.72; nRMSE% = 14) than PLSR (R2 = 0.54; nRMSE% = 19) and GPR (with R2 = 0.62; nRMSE% = 16). Season specific models based on 1D-CNN indicated apparent differences (14 ≤ nRMSE ≤19 for the test dataset), while the performance of all seasons combined model was remained higher for the test dataset (nRMSE% = 14). The full spectral range model showed higher accuracy than the spectral region-specific models (VNIR and SWIR alone) (15.8 ≤ nRMSE ≤18.5). Additionally, predictions derived using 1D-CNN were more precise (less uncertain) with <0.12 mean standard deviation (uncertainty intervals) than PLSR (0.31) and GPR (0.16). This study demonstrated the potential of 1D-CNN as an alternative to conventional techniques to determine the N% from canopy hyperspectral spectra.