Machine Learning Applications for Earth Observation

Abstract

201 panels show the dust sources identified by our approach with (lower) and without (upper) SOM class 137. The SOM had automatically determined that the sediment in the Bodélé depression was distinct from the surrounding dust sources and put it in a class all of its own, class 137. Indeed it is different, the Bodélé depression was once filled with a freshwater lake that has long since dried up (Washington et al., 2005). This has left behind diatoms that now make up the surface of the depression. The two key points being, first that the dust source of the Bodélé is distinct from the surrounding dust sources, and second, that it consists of diatoms. This is interesting as if we could devise a way of distinguishing dust sources with containing certain biological materials it would have significant applications for public health issues. Commercial and recreational fisheries within the Gulf of Mexico contribute significantly to the region's ocean economy making effective management a priority. The goal of fisheries management is to optimize the benefits of living marine resources by addressing threats to a resources' sustainability through conservation, development, and full utilization of the fishery resources to provide food, employment , income, and recreation. Therefore, it is desirable to minimize management actions that may result in negative impacts on fishermen and the coastal community. However, depending on the source of the threat, some management actions such as implementing fishing gear restrictions, time and area closures, and harvest limits, may have direct adverse impacts on fishermen. Often issues impacting fisheries populations arise from degradation or loss of habitat, requiring a different management approach. Coastal and marine habitats can be significantly and rapidly impacted by a number of anthropogenic actions and natural events such as coastal storms, development and hydrological alterations. With approximately 98% of Gulf of Mexico fisheries dependent on estuarine and near-shore habitats at some point in their life cycle (Lellis-Dibble et al., 2008), it is critical that resource managers have the ability to quickly and frequently monitor and assess habitat loss and degradation. However, inconsistencies in the approach that various agencies use for naming habitats make it difficult to develop a region-wide habitat map without standardizing the information. Unlike most habitat classification systems currently in use, the Coastal and Marine Ecological Classification Standard (CMECS) (Committee et al., 2012) has a water column component which identifies key classifiers required to characterize pelagic habitat types. The vastness and dynamic nature of the ocean's water column limit the feasibility of the frequent in situ sampling that would be necessary to monitor these classifiers and routinely produce region-wide map products. Our ultimate goal is to provide an example of how the Machine Learning classification manifests automatically to the physical classification schemes such as CMECS. Alternatives to in-situ sampling such as remote sensing classification offer a