Abstract
This paper introduces the Spatial Imprecision Adjustment (SIA) method, a neural-network-based post-processing framework designed to enhance the predictive accuracy of geospatial deep learning models trained on imprecise labels, a common challenge in socioeconomic survey data. SIA addresses the challenges of imprecision by applying a series of post-hoc adjustments to model estimates based on a neighborhood of observations, without requiring re-training of the deep learning components of the model. In cases where the exact location at which a measurement was taken is unknown (i.e. household income), the SIA approach (a) samples multiple potential candidates in an adaptable-size buffer region, (b) extracts activations from the fully connected (FC) layers of convolutional-based models for each candidate; and (c) applies a Random Forest (RF) model to each candidate’s activations to generate a single prediction of the target variable. Through ablation studies, we evaluate the contribution of various architectural components of this approach, such as the inclusion of spatial statistics or coordinates, demonstrating their impact on improving model performance. Additionally, we investigate the transferability of the model across diverse geographies and evaluate the limitations of the method when applied to heterogeneous landscapes. Across our tests, the method shows significant improvements (up to 20% in some cases) in prediction accuracy, with the largest improvements being found in urban regions.
Author supplied keywords
Cite
CITATION STYLE
Baier, H., & Runfola, D. (2025). Addressing spatial imprecision in deep learning for satellite imagery-based socioeconomic predictions. GIScience and Remote Sensing, 62(1). https://doi.org/10.1080/15481603.2025.2540537
Register to see more suggestions
Mendeley helps you to discover research relevant for your work.