Spatial portability of random forest models to estimate site-specific air temperature for prediction of emergence dates of the Asian Corn Borer in North Korea

Abstract

Listen

Translate article

The air temperature estimates derived from satellite data products would facilitate site-specific prediction of pest outbreaks. Here we propose an approach to estimate air temperature with spatial portability using both land surface temperature (LST) and atmospheric profile (AP) products based on MODIS satellites. In the approach, quantitative and qualitative variables were used as inputs to assess and improve the spatial portability of random forest (RF) models. Daily temperature data at sites in the Korean peninsula were collected from a weather database operated by the Korean Meteorological Administration. Sets of input variables were defined to represent temperature data from the LST (LT) and AP (AT) products, geographical properties (GE), data quality and cloud conditions (QC), land cover type (LC), and auxiliary properties (AX), respectively. The combinations of these sets were used as inputs to the RF models, which were cross-validated using satellite data and weather observation data in South Korea from 2013 to 2019. Concordance Correlation Coefficient (CCC) and Spatial Portability Index (SPI) were determined to compare the accuracy of the RF models across the sites of interest. The RF model that uses AT, LT, and QC variable sets as inputs (RFALQ) tended to have high values of CCC among test sites in South Korea, which ranged from 0.91 to 0.99. RFALQ also had higher values of SPI at sites in North Korea than the other models in which either AT or LT was excluded. Furthermore, the emergence date of Asian Corn Borer (Ostrinia furnacalis) was predicted with high values of CCC (0.94–0.97) across the study sites when RFALQ was used to prepare the input data to an insect phenology model. Our results suggest that both LST and AP products would contribute to higher spatial portability of air temperature estimation. This also hints that the temperature estimation model with high spatial portability would facilitate the pest models, which can be used to reduce the risk of food insecurity in developing countries with sparse weather station networks.