Spatial Transfer Learning with Simple MLP

TL;DR

This work tackles spatial prediction under non-stationarity and data scarcity by introducing a neural-network–based transfer-learning framework that pre-trains on large external spatial data and finetunes on a target dataset. The core method combines a Wendland-based radial-basis embedding (yielding a 139-dimensional feature) with a deep fully connected network, trained in two phases to transfer learned representations to the target domain. In simulations over stationary and non-stationary processes, the approach consistently outperforms target-only models and Kriging at small target sample sizes, and converges toward baseline performance as target data increases, marking a first step toward spatial transfer learning. The findings suggest practical value for improving spatial predictions when local data are limited, with avenues for exploring deeper or graph-based architectures in future work.

Abstract

First step to investigate the potential of transfer learning applied to the field of spatial statistics

Figures (3)

• Figure 1: Simulation data
• Figure 2: Stationary process MSE
• Figure 3: Non-stationary process MSE