Predictive modeling of sugarbeet quality using vegetative index, statistical, and artificial neural network methods.

摘要

Sugarbeet productivity is measured by the amount of biomass produced and recoverable sucrose contained within the plant root. Earlier studies have focused on root yield with little or no regard to sucrose concentration. Economic benefits are realized by both the processor and producer when processing a sugarbeet with high sucrose concentration in the root biomass. This study investigated the possibility of using statistical, Artificial Neural Network (ANN) and vegetative canopy models to predict whole-field sucrose concentration in sugarbeet fields using canopy spectral reflectance and sugarbeet field production data. The data consisted of five years of Landsat 5 and Landsat 7 Thematic Mapper (TM) multispectral images, field location, and field production data sets from 2003 to 2007. Fields were planted to the Beta and Hilleshog sugarbeet varieties within the Southern Minnesota Beet Sugar Cooperative's (SMBSC) growing region. Vegetative index models used in this research consisted of the Normalized Difference Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), and an index that is similar to the GNDVI, the Mid-infrared band 5 NDVI (M5NDVI). This index used band 5 in place of band 4 as was used in the conventional NDVI. Multiple Linear Regression (MLR) models were also created from late season canopy, field location, and sugarbeet quality information to both discover the predictive potential of conventional MLR models and provide a baseline comparison tool for the ANN model predictions. Statistical analysis using Pearson's correlation coefficient was applied to the ANN and conventional MLR model sucrose concentration predictions to determine the level of correlation between the predicted and actual sucrose concentration. Simple linear regression analysis performed on ANN and MLR model techniques produced statistically significant correlations between late season canopy spectral characteristics, range location information, and sucrose concentration predictions for all years except for some within the Hilleshog variety models. These statistical results revealed the M5NDVI index to have a stronger correlation with modified sucrose concentration than the other indices in all but one year. Although there were some low correlations with some of the modeling techniques applied to the Hilleshog varieties, the statistical results suggest the predictive ability of ANN, MLR, and canopy vegetative index modeling techniques can be used to classify whole-field sucrose concentration from canopy spectral and field production data prior to the start of the main harvest campaign for harvest timing considerations. Keywords: Artificial Neural Network, Precision Crop Management, Prediction, Regression analysis, Sugarbeet Canopy, Sugarbeet Quality, Vegetation Indices.