Deriving Canopy Nutrient Content using Hyperspectral Data and Machine Learning for Cannabis Sativa

Keywords: Hyperspectral, Machine Learning, Remote Sensing, Precision Agriculture
Abstract Type: Virtual Paper Abstract

Abstract

Cannabis Sativa has recently experienced a resurgence of commercial interest due to its various use in the pharmaceutical and construction industries. With legalization on state and federal levels in Canada and United States, commercial production is on the rise. Applying phenotyping techniques, based on Precision Agriculture (PA) practices for other crops, in Cannabis Sativa production allows growers to quantify the development and quality of the plant. Quantification of canopy nutrient levels allows targeted management and reduction in fertilizer used. Machine and deep learning methods are applied to hyperspectral data to derive the canopy nutrient content including nitrogen (N), phosphorous (P), and potassium (K). Techniques such as linear multivariate regression do not utilize the hyperspectral data to its full advantage. In this study, hyperspectral reflectance data was gathered for 150 plants using a spectroradiometer. In-situ tissue sampling was also conducted to extract canopy nutrient levels. Hyperspectral data are pre-processed to remove the atmospheric absorption effect and instrument noise. Feature extraction methods including Principal Component Analysis and narrow-band vegetation indices are used to create input data for random forest regression (RFR). For comparison, a 1-Dimensional convolution neural network, with automatic feature extraction, is applied to the raw spectral data to derive a canopy nutrients prediction model. The expected result is a recommended model that can help growers quantify and estimate canopy nutrient levels.