Journal
EARTH SCIENCE INFORMATICS
Volume 14, Issue 4, Pages 2231-2247Publisher
SPRINGER HEIDELBERG
DOI: 10.1007/s12145-021-00685-4
Keywords
Radial basis function; Sigmoid; Kernel; Random forests; Naive Bayes; Mc Nemar's test
Funding
- Deanship of Scientific Research at King Khalid University [RGP2/194/42]
Ask authors/readers for more resources
The study compared the relative effectiveness of nine different pixel- and object-based classification algorithms in generating remotely sensed LULC maps, with random forest and object-based classification identified as potential classifiers. Additionally, the results suggest that achieving higher overall accuracy in classifying a large urban area using 5 m resolution satellite dataset is difficult.
The preparation of accurate LULC is of great importance as it is used in various studies ranging from change detection to geospatial modelling. Literature offers studies comparing different classification algorithms/approaches to prepare LULC maps. However, still there is a lack of studies that can provide a comprehensive analysis on widely used classification algorithms. Hence, in the present study, nine different pixel- and object-based classification algorithms have been used to compare their relative effectiveness in generating remotely sensed LULC maps. The algorithms include maximum likelihood, neural network, support vector machine (linear, polynomial, RBF (radial basis function), sigmoid kernels), random forest (RF) and Naive Bayes for pixel-based classification and maximum likelihood algorithm for object-based classification (OBC) approach. Additionally, the study has analysed the impact of different types of satellite datasets (i.e., high resolution image and resolution merged images of same resolution) on relative effectiveness of the algorithms in classifying the satellite imageries accurately. High resolution (5 m) satellite image LISS 4 MX70, resolution merged satellite images (5 m) LISS 3 merged with LISS 4 mono and LISS 3 merged with IRS-1D are employed for classification. 27 LULC maps (9 classification algorithms * 3 images) are evaluated for comparing classification algorithms. The accuracy assessment of the images is carried out using confusion matrix and Mc Nemar's test. It has been observed that (1) the performance of all classification algorithms differs from each other and (2) resolution merged data impacts classification accuracy differently when compared to other satellite image of same spatial resolution. RF and OBC are identified as potential classifiers with majority of datasets. The results suggest that due to heterogeneity in urban land-use, it is difficult to achieve higher overall accuracy in classifying a large urban area using 5 m resolution satellite dataset. Moreover, visual examination of LULC should be considered for choosing better classification approach as pixel-based approach produces salt-pepper effect in LULC, whereas OBC produces visually smoothened LULC and identifies even smaller objects in urban landscape. The comparative evaluation of different image types reveal that RF performs better with all images, however, the performance of OBC was found to be improved with original high-resolution data.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available