AUTHOR=Okyere Frank Gyan , Cudjoe Daniel , Sadeghi-Tehran Pouria , Virlet Nicolas , Riche Andrew B. , Castle March , Greche Latifa , Simms Daniel , Mhada Manal , Mohareb Fady , Hawkesford Malcolm John 

TITLE=Modeling the spatial-spectral characteristics of plants for nutrient status identification using hyperspectral data and deep learning methods

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 14 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2023.1209500

DOI=10.3389/fpls.2023.1209500

ISSN=1664-462X

ABSTRACT=Sustainable fertilizer management in precision agriculture is essential for both economic and environment reasons. To effectively manager fertilizer input, various methods are employed to monitor and track plant nutrient status. One of such methods is hyperspectral imaging which is on the rise in recent times. It is a remote sensing tool used to monitor plant physiological changes in response to environmental conditions and nutrient availability. However, conventional hyperspectral processing mainly focuses solely on either the spectral or spatial information of plants. This study aims to develop a hybrid 3D-2D Convolution Neural Network (CNN) capable of extracting both spatial and spectral information from quinoa and cowpea plants, to identify nutrient status at different growth stages. To achieve this, a nutrient experiment with four treatments (high and low levels of nitrogen and phosphorus) was conducted in the glasshouse. Three pre-processing techniques, comprising second derivative, standard normal variate, and linear discriminant analysis were applied to selected regions of interest within the hypercube. Alongside the raw data, these pre-processed datasets were used as inputs to train the model. The proposed model uses a 3D convolution architecture with different 3D kernel sizes to extract the spectral information from the hypercube. The output is reshaped and fed into a 2D separable convolution network that extracts spatial information from the hypercube. To assess the effectiveness of the hybrid model, comparisons were made with a 3D CNN, a 2D CNN, and a Hybrid Spectral Network (HybridSN). The pre-processed methods were assessed to estimate their impact on the performance of the models. Effective wavebands were selected from the best performing method using the greedy stepwise based correlation feature selection (CFS) method. The selected wavebands were then used to retrain the models to identify the nutrient status at five selected plant growth stages. The proposed hybrid model achieved a classification accuracy of over 95% on the test dataset, demonstrating its potential for identifying nitrogen and phosphorus status in cowpea and quinoa at different growth stages.