A Case Study Comparing Machine Learning and Vegetation Indices for Assessing Corn Nitrogen Status in an Agricultural Field in Minnesota
Compact hyperspectral sensors compatible with UAV platforms are becoming more readily available. These sensors provide reflectance in narrow spectral bands while covering a wide range of the electromagnetic spectrum. However, because of the narrow spectral bands and wide spectral range, hyperspectral data analysis can benefit greatly from data mining and machine learning techniques to leverage its power. In this study, rainfed corn was grown during the 2017 growing season using four nitrogen treatments (between 0 and 200 kg N/ha using 67 kg/ha increment steps) where 200 kg/ha represented the economically optimum nitrogen rate (EONR). This design generated four ordinal classes of N deficiency (dN) that were confirmed using chlorophyll readings (SPAD), leaf nitrogen content, and soil nitrate all collected at V5 corn growth stage, on the same day the hyperspectral data was collected. Hyperspectral images were collected using a line scanner sensor on board a hexacopter UAV platform. Following radiometric correction, image segmentation, spectral extraction and preprocessing, eight machine learning algorithms were compared for their accuracy in determining the four classes of nitrogen deficiency levels based on the DEONR (dN =dEONR). These algorithms are the logistic regression (LR), support vector machine (SVM), random forest (RF), gradient boosting (GB), naïve Bayes (NB), decision tree (DT) and Multi-Layer Perceptron (MLP). A confusion matrix based on a 30% test set (unseen by the models) was used to determine the performance of these classifiers and to illustrate the type and magnitude of errors for each classification method. Narrow-bands and simulated broad-band vegetation indices (VIs) were also derived from the hyperspectral data and compared to machine learning algorithms for the prediction of dN. Our findings confirm the challenge of using VIs to assess early season (V5) corn nitrogen status. VIs correctly estimated N stress on only 50% of plant samples, and caused nearly 30% of plants to be under-fertilized. On the other hand, hyperspectral machine learning significantly improved the assessment of corn nitrogen stress at V5, and achieved more than 90% classification accuracy. Particularly, SVM, LR, MLP and GB showed promising results. These findings illustrate the huge potential for UAV-compatible hyperspectral sensors to improve in-season corn nitrogen management for timely variable rate sidedress application.