Authors: Ayodeji Adesuyi*, University of Cape Town. South Africa, Zahn Műnch, Stellenbosch University. Stellenbosch. South Africa
Topics: Geographic Information Science and Systems, Remote Sensing, Land Use and Land Cover Change
Keywords: decision trees, ensemble classifiers, land cover, NDVI, MODIS, Landsat, Python, scripting
Session Type: Paper
Start / End Time: 5:00 PM / 6:40 PM
Room: Balcony B, Marriott, Mezzanine Level
Presentation File: No File Uploaded
Automation of land cover mapping can provide a cost effective alternative to manual processing and classification of multiple satellite images. This study describes development of an automated technique for identifying agricultural land cover using a custom scripting tool developed on an ArcGIS/Python platform. The tool integrates study area selection, reprojection, classification and accuracy assessment using geo-processing tools combined with custom-built scripts in Python, incorporating the Scikit-learn Python library to integrate machine learning algorithms. Multiple ensemble classifiers were implemented in a workflow automation tool (MEAWAT). MODIS normalized difference vegetation index (NDVI) data (MOD13Q1), as well as NDVI extracted from Landsat 8 were used in the analysis. NDVI phenology was extracted for three agricultural land cover classes and used to create training data for the classification. Using MEAWAT, classification accuracies in excess of 70% were achieved for MODIS data with decision tree and ensemble classifiers, random forest, extra-tree, and Adaboost, a meta-estimator. An accuracy of 89% was achieved using a boosted random forest classifier on Landsat 8 data. It was demonstrated that a better classification output was derived using MEAWAT on higher resolution satellite imagery provided good training data are available. These findings highlight the potential of MEAWAT for large dataset land cover classification using different satellite imagery.