Investigation of the Possibilities of Nitrogen Content Determination in Soil Samples by Means of Hyperspectral Image Analysis in the Near-Infrared Region

The investigation deals with preliminary assessment of possibilities for determination the amount of nitrogen in soil samples using analysis of hyperspectral images, obtained in near-infrared spectrum region. Nitrogen content assessment is an important issue for determination of the degree of fertilization in cultivated areas, as part of precision agriculture applications. The soil samples are preliminary prepared by adding different amount of CH4N2O in a dry soil and in a soil sample with constant water content. The images of the samples are then captured using N17E hyperspectral camera (Specim LTD, Finland) in the region of 850 ÷ 1700 nm. The hyperspectral images are processed and a set of normalized average spectral characteristics for each soil sample is presented for further analysis. The performance of several mathematical models: PLS, Stochastic Gradient Descent, Decision Tree, Neural Network (Multilayer Perceptron), kNN, AdaBoost, Random Forest, Gradient Boosting and Linear Regression was tested in order to obtain qualitative assessment of nitrogen content in soil samples. The results show that for most of the models the coefficient of determination R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> is above 0.6 and for Multilayer Perceptron, Stochastic Gradient Descent and Gradient Boosting it is above 0.9.