Improvement of leaf nitrogen content inference in Valencia-orange trees applying spectral analysis algorithms in UAV mounted-sensor images

• Autor(a) principal: Oscoa, Lucas Prado
• Data de Publicação: 2019
• Outros Autores: Marques Ramos, Ana Paula, Saito Moriya, Erika Akemi [UNESP], Souza, Mauricio de, Marcato Junior, Jose, Matsubara, Edson Takashi, Imai, Nilton Nobuhiro [UNESP], Creste, Jose Eduardo
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1016/j.jag.2019.101907; http://hdl.handle.net/11449/196196
• topic: Spectral band simulation; Multispectral images; Precision agriculture; Plant nutrition

Nitrogen is one of the main required nutrients for the production of citrus plants. Farmers have used the chemical analysis of leaf tissue to determine the amount of nitrogen needed in a crop. However, its possible to directly classify the leaf nitrogen content (LNC) using remote sensing data. But, the accuracy of this methodology is yet low and is unknown how to enhance it. We propose a new approach to estimate the LNC in Valencia orange trees applying spectral analysis algorithms in multispectral images of high spatial resolution. Here we show an accuracy upper than 87% in determining the LNC in Valencia orange tree. Previous research, that also used multispectral images of high spatial resolution, obtained an accuracy lower than 65%. A total of 320 spectral measurements were obtained with a field spectroradiometer and the multispectral images were acquired with a Parrot Sequoia camera mounted in an Unmanned Aerial Vehicle (UAV). We calculated the mean values of 10 spectral measurements and created 32 spectral signatures with different nitrogen content. Each spectral signature was assigned for three LNC classes; low ( <= 27 g.kg(-1)), medium ( > 27 and <= 29 g.kg(-1)) and high ( > 29 g.kg(-1)). A band simulation was performed to Parrot Sequoia images for each spectral signature. We adopted 7 spectral analysis algorithms to determine the LNC: Constrained Energy Minimization; Linear Spectral Unmixing; Mixture Tuned Matched Filtering; Minimum Distance; Orthogonal Subspace Projection; Spectral Angle Mapper (SAM) and; Spectral Information Divergence. All these algorithms were trained using the simulated spectral signatures as input data. We used the 32 spectral signatures as training data and approximately 30,000 pixels as testing data, corresponding to the identified nitrogen content in orange-trees. The performance of the algorithms was evaluated with a confusion matrix and Receiver Operating Characteristic curves. The SAM algorithm presented the highest accuracy (overall of 87.6% with a kappa coefficient of 0.75) to determine LNC in orange trees. The proposed methodology may reduce the number of leaf tissue analysis and also optimize the monitoring process of orange orchards.