DO LIDAR DATA OFFER PRACTICAL SIGNIFICANCE IN LULC CLASSIFICATION OVER NAIP DATA? COMPARING MULTIPLE MACHINE LEARNING ALGORITHMS USING GEOGRAPHIC OBJECT-BASED IMAGE ANALYSIS COUPLED WITH TARGET-ORIENTED VALIDATION

Keywords: Machine learning, Multi-class classification, target-oriented validation, Support Vector Machines, Random Forests, Spatial structure
Session Type: Virtual Poster Abstract
Day: Friday
Session Start / End Time: 2/25/2022 03:40 PM (Eastern Time (US & Canada)) - 2/25/2022 05:00 PM (Eastern Time (US & Canada))
Room: Virtual 39

Abstract

Accurate mapping of complex heterogeneous landscapes using high spatial resolution imagery is necessary for better land management decisions. However, large-area mapping at this resolution remains challenging due to radiometric differences between scenes, low spectral depth of imagery, landscape heterogeneity and computational limitations. Furthermore, in most land use land cover (LULC) studies, the spatial structure of supervised training data is largely ignored, which often inflates the model's accuracy. In this research, we used the training data on which spatial structures were accounted for to evaluate the effectiveness of spatial blocking in LULC mapping. Specifically, we evaluated the classification performance of two widely used machine learning algorithms (Support vector machine [SVM] and Random Forest [RF]). We demonstrated that spatial accounting of sample data reduces the over-optimistic performance of random cross-validation while improving location accuracy. We used National Agriculture Imagery Program (NAIP) digital-ortho quadrangle quads (DOQQ: n = 137) along with LiDAR footprints (2199) over the extent of Kenedy County. We employed multi-resolution image segmentation on NAIP and LiDAR surface features using eCognition software and derived 39 and 47 features at object level on NAIP and LiDAR-derived objects, respectively. We used 3719 spatially independent sample points on eight identified classes. We found that an RF-produced thematic map had higher accuracy in a grass-dominated landscape than SVM. Although LiDAR information on NAIP improved performance metrics, we concluded that the information contribution of LiDAR over NAIP data was not practically significant in grass-dominated landscapes owing to the large computational cost associated with LiDAR data.