A scalable method for the estimation of spatial disaggregation models.

Gaining information about detailed processes using aggregation information is a frequent challenge in research involving geospatial data, with examples in different fields of knowledge such as agronomy, soil science, meteorology, public health, epidemiology, and others. Analyses using aggregated data lead to distorted conclusions since they disregard local patterns, and such a problem has motivated different approaches for reconstructing the information in a finer resolution from the aggregated data. However, most existing methods focus on the particular case where the volume of data does not exceed the amount of memory available for computations, a situation that has become increasingly less frequent with the fast pace of data generation nowadays. In practice, this problem limits either spatial resolution or coverage of applications, thus precluding their use in a more general context. In this paper, we address the problem of disaggregation of spatial data with huge datasets by proposing a scalable method to estimate the parameters of a well-established model. We propose an iterative scheme for model estimation and prove its convergence to a critical point of the likelihood function derived. To test the method, we provide a controlled simulation and a real example for sugarcane production in Brazil. In the simulation, the results indicate a successful reconstruction of 1 million pixels from 90 block areas. In the real example, the results had a compatible match with the agronomic literature, indicating a reasonable prediction of sugarcane production in a 100 m spatial resolution (i.e., approx. 5 × 1 0 8 pixels) from 5,565 block-areas. Compared to the most similar previous work, scalability allowed us to use a nearly 100 times higher resolution, which corresponds to 10,000 times more pixels. With our methods, we expect to assist researchers from different fields in disaggregating spatial information to larger areas or higher resolutions. • Disaggregating spatial data on large spatial domains and fine spatial resolutions is intensive. • We present a scalable method to estimate the parameters of a disaggregation model. • The proposed method is iterative, so we provide a proof of convergence. • We illustrate its application with a simulation and a real example. • Scalability allowed us to downscale crop yield to a high-resolution in Brazil. [ABSTRACT FROM AUTHOR]