High-resolution multispectral imagery and LiDAR point cloud fusion for the discrimination and biophysical characterisation of vegetable crops at different levels of nitrogen

Abstract

High-resolution remote sensing data has expanded the scope, precision, and scale of remote sensing applications in agriculture. Availability of spatial information at actionable field units is vital for using remote sensing data in agriculture. Crop discrimination and biophysical characterisation sensitive to nutrient levels have not been addressed at the patch level. This work investigates the synergetic application of high-resolution satellite imagery and terrestrial LiDAR point cloud for object-level discrimination and biophysical characterisation of a few crops at different nitrogen (N) levels. To this end, cabbage, eggplant, and tomato at three levels of N were grown on the experimental fields of the University of Agricultural Sciences, Bengaluru, India, in 2017. Fusing the multispectral imagery (WorldView-III) and LiDAR point cloud (terrestrial laser scanner) at the feature level, object-level supervised classification and estimation of two critical biophysical parameters (crown area and biomass) were performed using the support vector machine (SVM) and Random Forests (RF) algorithms with reference to different N levels. Results suggest discrimination of vegetable crops with high accuracy (92%), about 20% higher than the individual sensors, from the fused imagery sensitive to N levels. The quality of retrievals indicates a contrasting pattern wherein the accuracy of the crown area is high with the LiDAR point cloud at various N levels. For the biomass, there is no perceptible differentiation of N levels within a crop. The accuracy of crop classification with reference to N levels is similar from both RF and SVM algorithms. However, RF algorithm offered substantially higher classification results when the N status is ignored. In contrast, the quality of biophysical modelling is very high and is similar from both the algorithms. Weather conditions and sub-field level environment-induced variations in the crop growth likely are the factors responsible for the reduced sensitivity of remote sensing data to crop N levels at the patch level.