Earth Observation satellite data sets play an integral role in supporting large-scale ecological applications, whether it be with information on land cover and use, landscape-level estimates of vegetative productivity and health, or multi-temporal views of changing environments and land surface dynamics. The manner in which these and other geospatial data sets are employed has begun to shift, however, in response to the rapidly growing archive of open-access satellite data, the rise of cloud computing, and the increasing popularity of machine learning-based analyses. We demonstrate the novel application of these technologies and methods to the large-area mapping of wetland extent in Alberta, Canada, where spatially-exhaustive, consistent, up-to-date wetland maps are of critical importance for regional monitoring and management, but are currently lacking. By employing cloud computing to support the processing of large satellite data sets, high-quality topographic information, and open-access machine-learning boosted regression tree modeling, we test a scalable, repeatable approach to producing a much-needed information product on wetland location. The relative importance of optical, radar, and topographic explanatory variables to our wetland occurrence models is also examined.
Results/Conclusions
Each of our models performed well according to the area under the receiver-operator characteristic curve (> 0.8), but including all input variables produced the best model. In our study area, the topographic wetness index was the explanatory variable of greatest importance. Both optical and radar satellite data improved topographically-based model performance, but stronger improvements were observed with optical inputs (e.g., normalized difference wetness index). Conversion to binary wetland presence maps also produced good results (e.g., overall accuracies > 0.7). The scalability of our workflow was demonstrated by the production of a wetland probability-of-occurrence map covering the forested, boreal region of Alberta, and area covering over 380,000 km2. This work illustrates the successful integration of increasingly popular open-access satellite data streams, cloud computing services, and a powerful, machine-learning algorithm within an approach that promises to support long-term consistent and comprehensive wetland mapping, monitoring, and management within Alberta, Canada.