Assessing the Ability of Vegetation Indices to Identify Shallow Subsurface Water Flow Pathways from Hyperspectral Imagery Using Machine Learning: Application
Abstract
Shallow underground water flow pathways expressed as slight depressions are common in the land surface. Under conditions of saturated overland flow, such as during heavy rain or snow melt, these areas of preferential flow might appear on the surface as very shallow flowing streams. When there is no water flowing in these ephemeral channels it can be difficult to identify them. It is especially difficult to discern the slight depressions above the subsurface water flow pathways (SWFP) when the area is covered by vegetation. Since the soil moisture content in these SWFP is often greater than the surrounding area, the vegetation growing on top of these channels shows different vigor and moisture content than the vegetation growing above the non-SWFP area. Vegetation indices (VI) are used in visible and near infrared (VNIR) hyperspectral imagery to enhance biophysical properties of vegetation, and so the brightness values between vegetation atop SWFP and the surrounding vegetation were highlighted. We performed supervised machine learning using ground-truth class labels to determine the conditional probability of a SWFP at a given pixel given either the spectral distribution or VI at that pixel. The training data estimates the probability distributions to a determined finite sampling accuracy for a binary Naïve Bayes classifier between SWFP and non-SWFP. The ground-truth data provides a test bed for understanding the ability to build SWFP classifiers using hyperspectral imagery. SWFP were distinguishable in the imagery within corn and grass fields and in areas with low-lying vegetation. However, the training data is limited to particular types of terrain and vegetation cover in the Shenandoah Valley, Virginia and this would limit the resulting classifier. Further training data could extend its use to other environments.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2017
- Bibcode:
- 2017AGUFM.H31K..01D
- Keywords:
-
- 1894 Instruments and techniques: modeling;
- HYDROLOGY;
- 1914 Data mining;
- INFORMATICS;
- 1916 Data and information discovery;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS