USING GIS AND MACHINE-LEARNING TO RECOGNIZE SUBTLE LANDSCAPE FEATURES ASSOCIATED WITH GLACIAL LAKE AGASSIZ

The Red River Valley was formed as a result of glacial lake Agassiz. Transported silt and clay deposited in the bottom of the ancient lake resulted in a vast flat region near the center of the basin. Significant topography exists near basin margins where a series of shoreline complex deposits are located. To assist in the recognition of features such as shorelines and offshore bars, we used GIS and machine-learning algorithms applied to digital elevation models (DEM) of the region.

We utilized a topographic position index (TPI), slope, aspect, and several additional surfaces created from the DEM using ESRI ArcGIS as independent variables in the models. Polygons were made on the features which are both likely and unlikely to be offshore bars to “teach” the algorithms how to identify the offshore bar features. After that, we generated random points within each polygon, and extracted the raster values associated with TPI, aspect, slope, and curvature.

The resulting prediction surface is capable of detecting subtle characteristics of offshore bars and shorelines, which cannot be easily identified in the original DEM especially in lower elevation regions. The original DEM can be displayed with hillshade effect which aids in identification of these features, however, other topographic features are also exaggerated as well, making it difficult to recognize subtle features.

Model results indicate that we can adequately predict offshore bars and shorelines (R² = 0.84, Estimate of Error Rate = 31.6%, AUC = 0.99). Moreover, it can identify those features from roads. However, the model occasionally detected rivers as offshore bars, although a-priori knowledge of the area allows us to discriminate these after model completion. Since both are related to the flow of water, further training of the models is warranted. Recognition of these subtle feature may prove useful for the study of past landscapes and environments.