Deep Learning based Fusion of LiDAR Point Cloud and Multispectral Imagery for Crop Classification Sensitive to Nitrogen Level

Abstract

Deep learning techniques have been evolving at a faster pace offering a common framework for developing models for various applications using remote sensing data. Availability of high resolution LiDAR point cloud and multispectral imagery offers structural and spectral features vital for identifying different crops exhibiting similar spectral similarities. Our work explores the potential of deep machine learning approach for the fusion of ground based terrestrial LiDAR point cloud and satellite-based multispectral imagery for three important horticulture crops, viz cabbage, eggplant and tomato at three different nitrogen (N) levels. The core idea is evaluating the discrimination potential of the crops considering the inherent nature of nutrient effects on the crop growth. The challenges in this study are (i) horticulture crops are relatively lower in height and lack sturdy geometric profiles, (ii) crop identification of a single crop at three N levels requires discernible self-derived features in the Deep Learning (DL) based model. Contrasting with the results obtained network (LiDAR and multispectral), the deep Convolution Neural Network (CNN) exhibits substantially higher discrimination performance on the fused dataset when sensitivity to N level is not considered. Considering the sensitivity to N level, results from using LiDAR point cloud alone are comparable with the fused dataset.