Improving pixel‐based classification of GRASS GIS with support vector machine

Abstract

Open source GIS software GRASS releases 8.0 to 8.4 have received some long overdue improvements in imagery handling such as the ability to reuse spectral signature files of existing classifiers, machine readable output of accuracy assessment tool and a support vector machine (SVM) classifier. Practical comparison of all three pixel‐based classifiers of GRASS GIS indicated that the maximum likelihood discriminant analysis classifier is the fastest and least accurate one, followed by a sequential maximum a posteriori classifier with reasonably fast execution time and good accuracy. The newly added SVM classifier is the slowest one but provides the highest accuracy and also shows the highest improvement potential by hyperparameter optimization. To illustrate the capabilities of the core classifiers, a showcase is presented where a single Sentinel 2 image is classified to distinguish 18 crop types using all three classifiers available in GRASS.

Assessing Forest-Change-Induced Carbon Storage Dynamics by Integrating GF-1 Image and Localized Allometric Growth Equations in Jiangning District, Nanjing, Eastern China (2017–2020)

“…For example, Huang et al combined Sentinel-2 spectral features and radar data backscattering features to classify tree species of typical plantation forests in the tropics via the random forest method [9]. Although the classical statistics-based or machine learning methods have operational speed and simplicity, the "salt and pepper phenomenon" is quite common in classification results [10]. To minimize the negative effects of pixel-based analysis methods, Object-based Image Analysis (OBIA) has been proposed and applied in forest-related remote sensing efforts because it can make full use of shallow information such as spectral, texture and geometric features and the spatial topology of features in medium-or high-resolution remote sensing imagery for classification [11].…”