Ki Hwan Cho

Extreme rainfall events intensified by climate change have increased the need for accurate and spatially explicit assessments of agricultural damage. Conventional flood-inundation-based analyses primarily detect the extent of water coverage and often fail to capture cultivation-infeasible damage such as sediment burial and cropland washout. To overcome this limitation, this study directly mapped severe agricultural land damage in the Sindeungcheon basin, South Korea, following the extreme rainfall event of July 19, 2025, using 10-m Sentinel-2 Level-2A imagery acquired before (July 9) and after (August 1) the event. Damaged croplands were defined as areas where more than 50% of the field surface was covered by sediment or where notable erosion and soil exposure indicated physical loss of arable land. A random forest classifier using four multispectral bands (blue, green, red, near-infrared) achieved high performance, with an overall accuracy of 99.1% and a kappa coefficient of 0.93 based on independent field-verified samples. Of the 2,743.41 ha of agricultural land in the basin, 65.56 ha (2.39%) were classified as severely damaged. Field crops experienced the largest impact (34.92 ha, 2.79%), followed by paddy fields (25.02 ha, 2.35%). Damage exhibited a clear spatial gradient, with the highest damage ratio (3.64%) occurring within 100 m of the river, highlighting the dominance of flood-driven sediment transport. Numerous micro-scale (<0.05 ha) damage patches also indicated highly localized and heterogeneous disturbance patterns. The resulting dataset provides a detailed representation of extreme-event-induced agricultural loss and addresses the key limitations of inundation-only damage assessments by identifying actual cultivation- infeasible areas. This dataset can support economic loss estimation, post-disaster recovery planning, and spatial mitigation strategies aimed at reducing agricultural vulnerability to future extreme rainfall events.

Deforestation is a global issue; however, each deforestation phenomenon occurs within its own local context. Elucidating this context in detail is important to prevent deforestation and maintain sustainable ecosystem management. In this study, we analyzed the land cover changes, forest characteristics, and modeled the forest decline over the last two decades to reveal the pattern and affecting factors of deforestation in the Honam-Jeongmaek mountain range. Forests less than 50-years-old dominate the study area, indicating they were mainly regenerated after the 1970s. Reforestation policies such as planting trees have helped forest regeneration. In the study region, as deforestation occurred, agricultural and residential areas decreased, and barren and grassland increased. We applied the Weibull regression model to determine forest survivorship and covariates. The deforestation risks are significantly different among regions; protected areas lose less forest than non-protected areas but the losses in protected areas were also significant, with approximately 5% from 2000 to 2020. Areas of higher elevation and steep slopes experience less deforestation, whereas areas closer to the mountain ridge are at greater risk. With survival analysis, it is possible to assess the risk of deforestation quantitatively and predict long-term survival of forests. The findings and methods of this study could contribute to better forest management and policymaking.

The aim of this study is to adapt and evaluate the effectiveness of a multi-temporal downscaled images technique for classifying the typical vegetation types of a reclaimed area. The areas reclaimed from estuarine tidal flats show high spatial heterogeneity in soil salinity conditions. There are three typical vegetation types for which the distribution is restricted by the soil conditions. A halophyte-dominated vegetation is located in a high saline area, grass vegetation is found in a mid- or low saline area, and reed/small-reed vegetation is situated in a low saline area. Multi-temporal satellite images were used to classify the vegetation types. Landsat images were downscaled to take into account spatial heterogeneity using cokriging. A random forest classifier was used for the classification, with downscaled Landsat and RapidEye images. Classification with RapidEye images alone demonstrated a lower level of accuracy than when combined with multi-temporal downscaled images. The results demonstrate the usefulness of a downscaling technique for mapping. This approach can provide a framework which is able to maintain low costs whilst producing richer images for the monitoring of a large and heterogeneous ecosystem.