Using Artificial Neural Networks to upscale soil erosion model results from local to regional scale. An example from Crete, Greece.

Soil erosion constitutes an increasing threat to soil productivity and food security. This work describes the potential of using Artificial Neural Networks (ANN) for upscaling soil loss outputs from medium to low scale. The Revised Universal Soil Loss Equation (RUSLE) model was implemented to calculate soil loss rates in two scales in Crete, Greece. Specifically, the RUSLE model was applied in six (6) watersheds across the island using medium spatial resolution satellite images (5m), namely Planetscope. These results were used to feed an ANN model to upscale the mesoscale outputs (5m) to regional outputs (30m-island level). The ANN system was trained using spatial environmental parameters such as the Normalized Difference Vegetation Index, Digital Elevation Model, and topographical slope angle. This "optimized" soil loss derivative later made it possible to compare it with the corresponding final derivative of Crete (regional spatial scale), which emerged from the straightforward processing of RUSLE model with the more "coarse" and generalized data as estimated from the  Landsat-8 satellite images (30m). The statistics revealed that the detailed and high-quality soil loss data, as derived from the upscaling process, provide more precise and reliable results.