Early estimation of canopy nitrogen (N) levels in cotton is necessary to maintain optimal canopy growth and derive best yields. Combining remote sensing and computer vision algorithms has made it possible to extract high-resolution spectral reflectance and canopy level morphology from larger fields. An experiment was conducted to study the response of cotton to four N application rates (0 kg/ha, 56 kg/ha, 112 kg/ha, 168 kg/ha). The differences in canopy N uptake and canopy height in response to the treatments became significant (p-value < 0.05) at the squaring stage. The objective of this study was to detect N stress levels in cotton canopy as early as the squaring stage and also estimate canopy N uptake (kg N/plant) from UAV-based multispectral images. Spectral vegetation indices and morphological features (canopy height and fractional canopy cover) were estimated from calibrated orthomosaics and digital elevation models (DEM). N uptake was estimated by single-parameter regression and multi-parameter regression (LASSO and Random Forest). Unsupervised k-Means clustering was used to separate the plots into three stress levels (stressed, moderately stressed, healthy) in a high dimension space. Canopy height was estimated with an overall R2 = 0.911 and MAPE = 22.244, however the estimation errors were high before the squaring stage. Rededge (RE) band-based indices (NDRE and CIrededge) were more sensitive to canopy N status than others. The best estimates of cotton plant N uptake from single parameter regression were obtained for measure canopy height (MAPE = 30.902) and NDRE (MAPE = 47.501). The multi-parameter models further improved N uptake estimation accuracy (LASSO MAPE = 22.133; random forest MAPE = 23.804). The results from k-means clustering showed that weekly changes (Δweekly) in features had better class separation and accuracy scores than instantaneous values of features observed at squaring. Clustering based on NDRE for canopy pixels and actual height gave the optimum separation (silhouette score = 0.488) and classification accuracy (adjusted rand index = 0.417).
|