Estimating Nitrogen Content in Corn Leaves Using Convolutional Neural Networks, Machine Learning and Hyperspectral Imaging — 7a — Rishik Aggarwal, David Clay, Graig Reicks, Thomas F. Burks, Jianwei Qin, Moon Kim, Swarnabha Roy, Amee Parmar

Corn (Zea mays L.) nitrogen (N) is an important factor influencing its qualitative and quantitative outputs. However, excessive use of N-fertilizer may leach and contaminate ground water as well as lead to emissions in the environment of greenhouse gases such as dinitrogen monoxide (N2O). This is why it is essential to monitor the N-status of corn plants for better yield and optimal N-fertilizer applications. The objective of the study was to use portable HyperSpectral Imaging (HSI) system [395 -1005 nm; 348 spectral bands] with convolutional neural network (CNN) and machine learning (ML) classifiers to differentiate corn leaves among three N-status: 0 N lb./acre, 80 N lb./acre (~90 N kg/ha) and 120 N lb./acre (134.5 N kg/ha). For this, two experimental fields (irrigated and dryland) were chosen each with 12 plots distributed as per Randomized Complete Block Design (RCBD) experiment design. Leaf samples were collected and scanned using the HSI system on a weekly basis starting at V6 growth stage. Principal Component Analysis (PCA) was used to determine the five most discriminating bands: 659.1 nm, 706.9 nm, 717.5 nm, 855.4 nm and, 934.9 nm. The selected bands were used with CNN to extract spatial features which were then used to train Random Forest (RF) and Support Vector Machine (SVM) classifiers. It was found that for the irrigated field, RF could detect N-status of corn leaves with an overall accuracy of 92.34% while SVM resulted in an overall accuracy of 88.79%. Similarly, for the dryland, RF and SVM reached accuracies of 87.22% and 65.69% respectively. The results demonstrate that HIS with CNN and ML can estimate the N-status of corn leaves before visual symptoms appear. Therefore, our future work will integrate this algorithm on a UAV-based HSI system for generating N-status map of corn field for precision fertilizer applications.

South Dakota State University
Pappu Kumar Yadav