Identifying the correct location of channel heads remains a challenging aspect in
hydrogeomorphic analysis. Though field mapping is a reliable method, this may become
infeasible for large basins. High resolution remote sensing data provides another way to
predict channel heads. Existing literatures have suggested use of digital elevation models
(DEM) to extract channel heads by applying an area or slope-area thresholding method.
However, channel initiation process is more complex and depends on other factors like
topographic curvature, land use land cover etc. In this study, we have used machine learning
models to extract channel heads from freely available 1 arc second SRTM DEM data for a
basin in the Lesser Himalaya. We have used upstream area, local slope and local curvature as
features in our models. Actual channel heads were digitized manually from high-resolution
(1-2 m) IKONOS imagery available on Google Earth. Decision tree model generated the best
results with a F1-score of 0.45 and correctly predicted around 78% of the channel heads from
the test set along with a high number of false positives. Future work will be applying this
method on available high-resolution Lidar-derived DEM, with more field-mapped channel
heads.