Multi-sensor Fusion Workflow for Accurate Classification and Mapping of Sugarcane Crops

Abstract

This study aims to assess the classification accuracy of a novel mapping workflow for sugarcane crops identification that combines light detection and ranging (LiDAR) point clouds and remotely-sensed orthoimages. The combined input data of plant height LiDAR point clouds and multispectral orthoimages were processed using a technique called object-based image analysis (OBIA). The use of multi-source inputs makes the mapping workflow unique and is expected to yield higher accuracy compared to the existing techniques. The multi-source inputs are passed through five phases: data collection, data fusion, image segmentation, accuracy validation, and mapping. Data regarding sugarcane crops were randomly collected in ten sampling sites in the study area. Five out of the ten sampling sites were designated as training sites and the remaining five as validation sites. Normalized digital surface model (nDSM) was created using the LiDAR data. The nDSM was paired with Orthophoto and segmented for feature extraction in OBIA by developing a rule-set in eCognition software. A rule-set was created to classify and to segment sugarcane using nDSM and Orthophoto from the training and validation area sites. A machine learning algorithm called support vector machine (SVM) was used to classify entities in the image. The SVM was constructed using the nDSM. The height parameter nDSM was applied, and the overall accuracy assessment was 98.74% with Kappa index agreement (KIA) 97.47%, while the overall accuracy assessment of sugarcane in the five validation sites were 94.23%, 80.28%, 94.50%, 93.59%, and 93.22%. The results suggest that the mapping workflow of sugarcane crops employing OBIA, LiDAR data, and Orthoimages is attainable. The techniques and process used in this study are potentially useful for the classification and mapping of sugarcane crops.