Convolutional Neural Networks to Improve Soil-Cover Maps by Identifying Rock Outcrops in California

Abstract:

Researchers and land managers use land-cover maps to model how land-cover change impacts water and other elemental fluxes. For instance, whether land surfaces are soil or bare rock impacts hydrological models, carbon storage estimates, and susceptibility models for soil erosion and landslides. In the United States, the available Natural Resources Conservation Service (NRCS) soil maps and National Land Cover Database (NLCD) tend to over- or underestimate extents of rock soil when compared to more detailed maps. In our area of interest in the Sierra Nevada Mountains in California (USA), NRCS soil maps overestimated rock outcrops by 41% and NLCD underestimated the same by 88%. The increasing availability of high-resolution remote sensing imagery can be used along with machine learning techniques to improve regional maps of land-cover, including the distinction between soil and exposed rock surfaces. We built a convolutional neural network (CNN) to differentiate exposed bare rock from soil cover across the Sierra Nevada Mountains using National Aerial Inventory Program (NAIP) 1-m othroimagery. Exposed rock was mapped at eight sites in the Sierra Nevada Mountains, and a CNN was trained to classify these rock outcroppings. In the Sierra Nevada Mountains, the model classifies bare rock with an F₁ score of 0.95, which is significantly higher than classical methods and existing databases. This tested CNN model was then used to predict rock outcrops across the Sierra Nevada Mountains (~49,000 km²) and will be applied to the entire state of California. The results are validated with additional training sites. When applied to the entire state of California, we expect the CNN model to similarly improve classification accuracy of rock outcrops. Improved representation of rock outcrops in land cover maps will be important to improve estimates of soil erosion, infiltration for hydrological models, among other applications.