Soo-Jin Lee

Understanding how phenological, soil, and topographic factors interact to shape rice yield is essential in reclaimed paddy soils. We built an interpretable modeling framework that combines Sentinel-2 NDVI-based phenological metrics with soil, topographic, and phenology-aligned meteorological variables to describe yield variability of the floury rice cultivar Baromi2 across ten regions in South Korea. Fifteen phenology indicators were derived from Savitzky-Golay-smoothed NDVI trajectories, while 52 environmental variables were screened sequentially through Variance Inflation Factor (VIF < 5), Recursive Feature Elimination (RFE), and Permutation Importance (PI) before Random Forest regression to ensure transparency and control multicollinearity. In the phenology-only model, late-season greenness (OffsetV) explained about 62 % of yield variation (R² = 0.623). The environment-only model showed that soil Mg concentration and Altitude were the strongest drivers (R² = 0.620). Combining all predictors improved accuracy slightly (R² = 0.648; ΔR² ≈ 0.025) and removed spatial autocorrelation, suggesting that phenology chiefly mediates soil and terrain effects rather than adding large independent power. Partial dependence analysis indicated that Mg > 2 mg/kg lowered yield by roughly 10–12%, whereas moderate Altitude (0–10 m) raised it by ≈ 6 %. This harmonized VIF–RFE–PI workflow follows explainable-AI principles and shows that, under irrigated and climatically steady conditions, phenology and soil variables provide more consistent yield signals than weather aggregates. The framework supports precision-agriculture applications and can be adapted across years and cultivars to advance sustainable rice-yield management.

Soo-Jin Lee, Jae-Hyun Ryu, Geun-Ho Kwak, Lak-Yeong Choi, Dasom Jeon, Meejeong Park, Kyung-Do Lee. Korean J. Remote Sens. 2025;41:717-26. https://doi.org/10.7780/kjrs.2025.41.5.2

Geun-Ho Kwak, Jae-Hyun Ryu, Soo-Jin Lee, So Young Lee, Young-Ah Jeon, Suk Young Hong, Lak-Yeong Choi, Kyung-Do Lee. Korean J. Remote Sens. 2025;41:551-64. https://doi.org/10.7780/kjrs.2025.41.3.5

Droughts are expected to increase in both frequency and severity, exacerbated by rising global temperatures associated with climate change. These trends pose serious threats to the agricultural sector, directly impacting food production and security. Moreover, increasing drought incidence increases the risks associated with agricultural and forestry disasters, including reduced crop yields, soil degradation, and wildfires. Given these challenges, the ability to accurately monitor and predict drought conditions is critical. Effective drought forecasting plays an important role in establishing agricultural and water management policies and enabling better handling of the impacts of these events. This will enable timely and informed decisions to ensure that appropriate measures are in place to mitigate the adverse impacts of drought on ecosystems, food supplies and overall environmental health. The development and improvement of tools for drought time series forecasting is therefore essential to ongoing efforts to adapt to and mitigate the impacts of climate change. This study introduces a model designed to predict Vegetation Health Index (VHI) time series data using the Predictive Recurrent Neural Network Version 2 (PredRNN-V2). The VHI, which effectively integrates land surface temperature and vegetation status, has been widely used in drought assessment. The study focuses on South Korea, utilizing long-term weekly VHI data from NOAA for short-term prediction. The PredRNN-V2 model utilizes a network of interconnected spatio-temporal LSTM cells to learn and predict the temporal and spatial characteristics of time series images. This architecture can properly handle the complex spatial and temporal dynamics inherent in satellite-based drought data and can therefore be an effective tool for drought prediction.This research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(NRF-2022R1I1A1A01073185)

Soil moisture is a critical parameter that significantly impacts the global energy balance, including the hydrologic cycle, land–atmosphere interactions, soil evaporation, and plant growth. Currently, soil moisture is typically measured by installing sensors in the ground or through satellite remote sensing, with data retrieval facilitated by reanalysis models such as the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis 5 (ERA5) and the Global Land Data Assimilation System (GLDAS). However, the suitability of these methods for capturing local-scale variabilities is insufficiently validated, particularly in regions like South Korea, where land surfaces are highly complex and heterogeneous. In contrast, artificial intelligence (AI) approaches have shown promising potential for soil moisture retrieval at the local scale but have rarely demonstrated substantial products for spatially continuous grids. This paper presents the retrieval of daily soil moisture (SM) over a 500 m grid for croplands in South Korea using random forest (RF) and automated machine learning (AutoML) models, leveraging satellite images and meteorological data. In a blind test conducted for the years 2013–2019, the AutoML-based SM model demonstrated optimal performance, achieving a root mean square error of 2.713% and a correlation coefficient of 0.940. Furthermore, the performance of the AutoML model remained consistent across all the years and months, as well as under extreme weather conditions, indicating its reliability and stability. Comparing the soil moisture data derived from our AutoML model with the reanalysis data from sources such as the European Space Agency Climate Change Initiative (ESA CCI), GLDAS, the Local Data Assimilation and Prediction System (LDAPS), and ERA5 for the South Korea region reveals that our AutoML model provides a much better representation. These experiments confirm the feasibility of AutoML-based SM retrieval, particularly for local agrometeorological applications in regions with heterogeneous land surfaces like South Korea.

Soil moisture (SM) is an indicator of the moisture status of the land surface, which is useful for monitoring extreme weather events. Representative global SM datasets include the National Aeronautics and Space Administration (NASA) Soil Moisture Active Passive (SMAP), the Global Land Data Assimilation System (GLDAS), and the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis 5 (ERA5), but due to their low spatial resolutions, none of these datasets well describe SM changes in local areas, and they tend to have a low accuracy. Machine learning (ML)-based SM predictions have demonstrated high accuracy, but obtaining semi-real-time SM information remains challenging, and the dependence of the validation accuracy on the data sampling method used, such as random or yearly sampling, has led to uncertainties. In this study, we aimed to develop an ML-based model for real-time SM estimation that can capture local-scale variabilities in SM and have reliable accuracy, regardless of the sampling method. This study was conducted in South Korea, and satellite image data, numerical weather prediction (NWP) data, and topographic data provided within one day were used as the input data. For SM modeling, 13 input variables affecting the surface SM status were selected: 10- and 20-day cumulative standardized precipitation indexes (SPI10 and SPI20), a normalized difference vegetation index (NDVI), downward shortwave radiation (DSR), air temperature (Tair), land surface temperature (LST), soil temperature (Tsoil), relative humidity (RH), latent heat flux (LE), slope, elevation, topographic ruggedness index (TRI), and aspect. Then, SM models based on random forest (RF) and automated machine learning (AutoML) were constructed, trained, and validated using random sampling and leave-one-year-out (LOYO) cross-validation. The RF- and AutoML-based SM models had significantly high accuracy rates based on comparisons with in situ SM (mean absolute error (MAE) = 2.212–4.132%; mean bias error (MBE) = −0.110–0.136%; root mean square error (RMSE) = 3.186–5.384%; correlation coefficient (CC) = 0.732–0.913), while the AutoML-based SM model tended to have a higher accuracy than the RF-based SM model, regardless of the data sampling method used. In addition, when compared to in situ SM data, the SM models demonstrated the highest accuracy, outperforming both GLDAS and ERA5 SM data and well representing changes in the dryness/wetness of the land surface according to meteorological events (heatwave, drought, and rainfall). The SM models proposed in this study can, thus, offer semi-real-time SM data, aiding in the monitoring of moisture changes in the land surface, as well as short-term meteorological disasters, like flash droughts or floods.

Soil moisture (SM) is a crucial hydrologic factor that affects the global cycle of energy, carbon, and water, as well as plant growth and crop yield; therefore, an accurate estimate of SM is important for both the global environment and agriculture. Satellite-based SM data have been provided by the National Aeronautics and Space Administration (NASA)’s Soil Moisture Active Passive (SMAP) and the European Space Agency (ESA)’s Soil Moisture and Ocean Salinity (SMOS) satellite missions, but these data are based on passive microwave sensors, which have limited spatial resolution. Thus, detailed observations and analyses of the local distribution of SM are limited. The recent emergence of deep learning techniques, such as rectified linear unit (ReLU) and dropout, has produced effective solutions to complex problems. Deep neural networks (DNNs) have been used to accurately estimate hydrologic factors, such as SM and evapotranspiration, but studies of SM estimates derived from the joint use of DNN and high-resolution satellite data, such as Sentinel-1 and Sentinel-2, are lacking. In this study, we aim to estimate high-resolution SM at 30 m resolution, which is important for local-scale SM monitoring in croplands. We used a variety of input data, such as radar factors, optical factors, and vegetation indices, which can be extracted from Sentinel-1 and -2, terrain information (e.g., elevation), and crop information (e.g., cover type and month), and developed an integrated SM model across various crop surfaces by using these input data and DNN (which can learn the complexity and nonlinearity of the various data). The study was performed in the agricultural areas of Manitoba and Saskatchewan, Canada, and the in situ SM data for these areas were obtained from the Agriculture and Agri-Food Canada (AAFC) Real-time In Situ Soil Monitoring for Agriculture (RISMA) network. We conducted various experiments with several hyperparameters that affected the performance of the DNN-based model and ultimately obtained a high-performing SM model. The optimal SM model had a root-mean-square error (RMSE) of 0.0416 m³/m³ and a correlation coefficient (CC) of 0.9226. This model’s estimates showed better agreement with in situ SM than the SMAP 9 km SM. The accuracy of the model was high when the daily precipitation was zero or very low and also during the vegetation growth stage. However, its accuracy decreased when precipitation or the vitality of the vegetation were high. This suggests that precipitation affects surface erosion and water layer formation, and vegetation adds complexity to the SM estimate. Nevertheless, the distribution of SM estimated by our model generally reflected the local soil characteristics. This work will aid in drought and flood prevention and mitigation, and serve as a tool for assessing the potential growth of crops according to SM conditions.

Various drought indices have been used for agricultural drought monitoring, such as Standardized Precipitation Index (SPI), Standardized Precipitation Evapotranspiration Index (SPEI), Palmer Drought Severity Index (PDSI), Soil Water Deficit Index (SWDI), Normalized Difference Vegetation Index (NDVI), Vegetation Health Index (VHI), Vegetation Drought Response Index (VegDRI), and Scaled Drought Condition Index (SDCI). They incorporate such factors as rainfall, land surface temperature (LST), potential evapotranspiration (PET), soil moisture content (SM), and vegetation index to express the meteorological and agricultural aspects of drought. However, these five factors should be combined more comprehensively and reasonably to explain better the dryness/wetness of land surface and the association with crop yield. This study aims to develop the Integrated Crop Drought Index (ICDI) by combining the weather factors (rainfall and LST), hydrological factors (PET and SM), and a vegetation factor (enhanced vegetation index (EVI)) to better express the wet/dry state of land surface and healthy/unhealthy state of vegetation together. The study area was the State of Illinois, a key region of the U.S. Corn Belt, and the quantification and analysis of the droughts were conducted on a county scale for 2004–2019. The performance of the ICDI was evaluated through the comparisons with SDCI and VegDRI, which are the representative drought index in terms of the composite of the dryness and vegetation elements. The ICDI properly expressed both the dry and wet trend of the land surface and described the state of the agricultural drought accompanied by yield damage. The ICDI had higher positive correlations with the corn yields than SDCI and VegDRI during the crucial growth period from June to August for 2004–2019, which means that the ICDI could reflect the agricultural drought well in terms of the dryness/wetness of land surface and the association with crop yield. Future work should examine the other factors for ICDI, such as locality, crop type, and the anthropogenic impacts, on drought. It is expected that the ICDI can be a viable option for agricultural drought monitoring and yield management.