A large scale lulc map, especially a national scale, is necessary for many environmental fields. Global lulc maps are available but it seems to be uncertain for a specific application such as national estimation of forest in order to compare with field survey statistical data. Therefore, a suitable machine-based approach of lulc production with free satellite data is essential. Up to date, there are several good approaches of large scale lulc mapping developed but has not been a comprehensive review yet. Admittedly, the accuracy of a land use land cover map can be affected by many factors such as input data source (optical/multispectral satellites, hyperspectral, very high-resolution optical images, radar/SAR, or fusion from several sources), classification methods and reference data. Can anyone suggest a suitable method for this task?
Thanks in advance and welcome any discussions.
Hi!
Random forest!
For a better classification accuracy you may need to develop a hybrid classification method. You may look the following paper to develop such approach
Article Multi-temporal land use classification using hybrid approach
I would suggest to take full advantage of Google earth engine, the implementation of Random forest you can find below
https://developers.google.com/earth-engine/classification
Hope these link might be helpful!
All the best
It is seem to easy to perform what you looking for from use Landsat imagery with supervised approach SVM that I prefer under object oriented
regards
Dear Duong Pc ,
If you are working within ENVI software, you can go for SVM (Support Vector Machine). ENVI provide many kernel types (Linear, Polynomial, Radial Basis Function, and Sigmoid) you can try to play with and evaluate how accurate they compare to each other. A good reference on the relative performance of SVM kernel type functions would be the work of Ustuner et al. (2015).
Article Application of Support Vector Machines for Landuse Classific...
Also, as all the previous answers have outlined, random forest classification is likely to provide good results. They originate from the work of breiman (2001), see https://link.springer.com/article/10.1023/A:1010933404324.
This classifier is not available in ENVI (as far as I know), but there is an IDL implementation called imageRF available.
Article ImageRF - A user-oriented implementation for remote sensing ...
Outside ENVI, you can try R-based solutions (https://rspatial.org/raster/rs/5-supclassification.html), R-QGIS combination (http://amsantac.co/blog/en/2015/11/28/classification-r.html), Python based solutions (https://ceholden.github.io/open-geo-tutorial/python/chapter_5_classification.html or).
There are also stand-alone toolboxes running Random Forests (and several others machine learning algorithms), such as ESA-SNAP toolbox software (https://step.esa.int/main/toolboxes/snap/), EnMAP Toolbox (http://www.enmap.org/?q=enmapbox), WEKA (https://www.cs.waikato.ac.nz/ml/weka).
Hope this helps.
- Badges
- Science topic
- Similar topics
- Phytochemistry
- Extracts