The Maps

Accurate and up-to-date geospatial data are a modern requirement for resource management and conservation, land use planning, economic development, and policy making. In 2018, we launched Intelligent Geosolutions (IGS), an initiative housed within the UMaine Center for Research of Sustainable Forests (CRSF), and partnered with the UMaine Advanced Computing Group to implement a semi-automated image processing and machine learning software system we refer to as the Supervised Adaptive Multi-objective Mapper (SAMM).

Satellite remote sensing imagery from programs such as Landsat and Sentinel have the potential to enable near-real time mapping of spatial forest attributes as well as changes in landscape conditions. New imagery is now freely available every few days, but the technical expertise and computing resources required to manipulate raw imagery into usable knowledge continue to act as barriers to widespread utilization. Over the past 15 years, our applied R&D program at the University of Maine has advanced to the point that we can offer a unique, local resource for advanced geospatial services.

Typical image classification methods almost universally fail to control systematic map error in the form of over- or under-estimation of class extent. Imperfect reference data, predictor uncertainty associated with satellite imagery, algorithm bias, and analyst error can contribute to dramatic misrepresentation of class extents and spatial distributions. This systematic error impacts map use in ways that are difficult to predict or correct. SAMM utilizes an innovative set of multi-objective ML algorithms designed to fit accurate and unbiased predictive models of forest attributes including: forest type, tree species abundance (e.g., % above ground biomass), and the occurrence and intensity of canopy disturbance (e.g., % biomass change). The ML methods developed for SAMM control or eliminate systematic patterns of map error by combining the strength of support vector machines (SVMs) to model complex, nonlinear relationships with the adaptability of a multi-objective genetic algorithm (GA). The GA drives the evolution of SVMs to simultaneously increase accuracy and reduce or eliminate systematic error including over- or under-estimation of class extent. SAMM integrates our multi-objective ML algorithms into semi-automated image processing and map production workflows executed on the cloud. SAMM enables efficient, high throughput processing of raw image data into high-quality output products.