DEEP LEARNING TRAINING WITH UNBALANCE SAMPLE DISTRIBUTION FOR REMOTE SENSING IMAGE SEGMENTATION

We developed an interpretable machine learning model to predict sub-seasonal and near-future wildfire-burned area over African and South American regions. We found strong time-lagged controls (up to 6–8 months) of local climate wetness on burned areas. A skillful use of such time-lagged controls in machine learning models results in highly accurate predictions of wildfire-burned areas; this will also help develop relevant early-warning and management systems for tropical wildfires.

Wildfire is a devastating Earth system process that burns about 500 million hectares of land each year. It wipes out vegetation including trees, shrubs, and grasses and causes large losses of economic assets. However, modeling the spatial distribution and temporal changes of wildfire activities at a global scale is challenging. This study built a machine-learning-based wildfire surrogate model within an existing Earth system model and achieved high accuracy.

We produce the 30 m annual China land cover dataset (CLCD), with an accuracy reaching 79.31 %. Trends and patterns of land cover changes during 1985 and 2019 were revealed, such as expansion of impervious surface (+148.71 %) and water (+18.39 %), decrease in cropland (−4.85 %) and increase in forest (+4.34 %). The CLCD generally reflected the rapid urbanization and a series of ecological projects in China and revealed the anthropogenic implications on LC under the condition of climate change.

Daily time series and full space-covered sub-pixel snow cover area data are urgently needed for climate and reanalysis studies. Due to the fact that observations from optical satellite sensors are affected by clouds, this study attempts to capture dynamic characteristics of snow cover at a fine spatiotemporal resolution (daily; 6.25 km) accurately by using passive microwave data. We demonstrate the potential to use the passive microwave and the MODIS data to map the fractional snow cover area.