Frost forecasting via weakly-supervised semantic segmentation of satellite imagery

Recent research on frost forecasting has employed machine learning approaches to build prediction models tailored to specific locations. Although these models have proven effective, their applications are limited to areas where meteorological observations are available. Forecasting coverage can be extended by using meteorological satellite data as predictors, enabling frost forecasting over broader regions. This can be formulated as a semantic segmentation task of detecting areas where frost is likely to occur. Using satellite images and geographical information of the target region at the forecast time as inputs, the semantic segmentation model generates a frost probability map for the target time of forecast. However, an important challenge arises from the limited availability of pixel-wise labels, as frost occurrence information is only available for pixels corresponding to frost observatories. To address this issue, we propose a weakly-supervised learning method for training the semantic segmentation model using satellite imagery with incomplete supervision. The learning objective involves accurately classifying labeled pixels while suppressing the entire frost probability map to zero when no frost is observed at any observatory within the target region. Additionally, a metric-surrogate loss is incorporated to maximize the critical success index for labeled pixels. We demonstrate the effectiveness of the proposed method for frost forecasting with varying lead times across the South Korean region using Geo-KOMPSAT-2A satellite data.