Prediction of rice yields at pixel scale rather than county scale can benefit crop management and scientific understanding because it is useful for monitoring how crop yields respond to various agricultural systems and environmental factors. In this study, we propose a methodology for the early prediction of rice yield at pixel scale combining a crop model and a deep learning model for different agricultural systems throughout South and North Korea. Initially, satellite-integrated crop models were applied to obtain a pixel-scale reference rice yield. Then, the pixel-scale reference rice yields were used as target labels in the deep learning model to leverage the advantages of crop models. Models of five different deep learning network architectures were employed to help determine the hybrid structure of long-short term memory (LSTM) and one-dimensional convolutional neural network (1D-CNN) layers by predicting the optimal model about two months ahead of harvest time. The suggested model showed good performance [R² = 0.859, Nash-Sutcliffe model efficiency = 0.858, root mean squared error = 0.605 Mg ha^-1], with specific spatial patterns of rice yields for South and North Korea. Analysis of the relative importance of the input variables showed the water-related index and maximum temperature in North Korea and the vegetation indices and geographic variables in South Korea to be crucial for predicting rice yields. The proposed approach successfully predicted and diagnosed rice yield at the pixel scale for inaccessible locations where reliable ground measurements are not available, especially North Korea.

Recent satellite missions have provided new perspectives by offering high spatial resolution, a variety of spectral properties, and fast revisit rates to the same regions. In this study, we examined the utility of both broadband red-edge spectral information and texture features for classifying paddy rice crops in South Korea into three different growth stages. The rice grown in South Korea can be grouped into early-maturing, medium-maturing, and medium-late-maturing cultivars, and each cultivar is known to have a minimum and maximum productivity. Therefore, the accurate classification of paddy rice crops into a certain time line enables pre-estimation of the expected rice yields. For the analysis, two seasons of RapidEye satellite image data were used. The results showed that the broadband red-edge information slightly improved the classification accuracy of the paddy rice crops, particularly when single-season image data were used. In contrast, texture information resulted in only minor improvement or even a slight decline in accuracy, although it is known to be advantageous for object-based classification. This was due to the homogeneous nature of paddy rice fields, as different rice cultivars are similar in terms of their morphology. Based on these results, we conclude that the additional spectral information such as the red-edge band is more useful than the texture features to detect different crop conditions in relatively homogeneous rice paddy environments. We therefore confirm the potential of broadband red-edge information to improve the classification of paddy rice crops. However, there is still a need to examine the relationship between textural properties and paddy rice crop parameters in greater depth.

A practical approach to continuously monitor and provide real-time solar energy prediction can help support reliable renewable energy supply and relevant energy security systems. In this study on the Korean Peninsula, contemporaneous solar radiation images obtained from the Communication, Ocean and Meteorological Satellite (COMS) Meteorological Imager (MI) system, were used to design a convolutional neural network and a long short-term memory network predictive model, ConvLSTM. This model was applied to predict one-hour ahead solar radiation and spatially map solar energy potential. The newly designed ConvLSTM model enabled reliable prediction of solar radiation, incorporating spatial changes in atmospheric conditions and capturing the temporal sequence-to-sequence variations that are likely to influence solar driven power supply and its overall stability. Results showed that the proposed ConvLSTM model successfully captured cloud-induced variations in ground level solar radiation when compared with reference images from a physical model. A comparison with ground pyranometer measurements indicated that the short-term prediction of global solar radiation by the proposed ConvLSTM had the highest accuracy [root mean square error (RMSE) = 83.458 W · m⁻², mean bias error (MBE) = 4.466 W · m⁻², coefficient of determination (R²) = 0.874] when compared with results of conventional artificial neural network (ANN) [RMSE = 94.085 W · m⁻², MBE = −6.039 W · m⁻², R² = 0.821] and random forest (RF) [RMSE = 95.262 W · m⁻², MBE = −11.576 W · m⁻², R² = 0.839] models. In addition, ConvLSTM better captured the temporal variations in predicted solar radiation, mainly due to cloud attenuation effects when compared with two selected ground stations. The study showed that contemporaneous satellite images over short-term or near real-time intervals can successfully support solar energy exploration in areas without continuous environmental monitoring systems, where satellite footprints are available to model and monitor solar energy management systems supporting real-life power grid systems.

Although data-driven methods including deep neural network (DNN) were introduced, there was not enough assessment about spatial characteristics when using limited ground observation as reference. This work aimed to interpret the feasibility of several machine learning approaches to assess the spatial distribution of solar radiation on Earth based on the Communication, Ocean, and Meteorological Satellite (COMS) Meteorological Imager (MI) geostationary satellite. Four data-driven models were selected (artificial neural network (ANN), random forest (RF), support vector regression (SVR), and DNN), to compare their accuracy and spatial estimating performance. Moreover, we used a physical model to probe the ability of data-driven methods, implementing hold-out and k-fold cross-validation approaches based on pyranometers located in South Korea. The results of analysis showed the RF had the highest accuracy in predicting performance, although the difference between RF and the second-best technique (DNN) was insignificant. Temporal variations in root mean square error (RMSE) were dependent on the number of data samples, while the physical model showed relatively less sensitivity. Nevertheless, DNN and RF showed less variability in RMSE than the others. To examine spatial estimation performance, we mapped solar radiation over South Korea for each model. The data-driven models accurately simulated the observed cloud pattern spatially, whereas the physical model failed to do because of cloud mask errors. These exhibited different spatial retrieval performances according to their own training approaches. Overall analysis showed that deeper layers of networks approaches (RF and DNN), could best simulate the challenging spatial pattern of thin clouds when using satellite multispectral data.

In this study, a new deep learning method was developed to estimate the spatiotemporal properties of the hourly aerosol optical depth (AOD) because existing physical models are limited in their abilities to separate the reflectance between aerosols and the underlying surface over land, accurately and effectively. By incorporating geostationary ocean color imagery (GOCI), multispectral bands were applied to train data-driven models to estimate the high-spatiotemporal-resolution AOD over Northeast Asia. Physical model and traditional machine learning (ML) models (the random forest (RF) and support vector regression (SVR) models) were compared with the deep neural network (DNN) model to evaluate its accuracy, implementing hold-out validation and $k$ -fold cross-validation approaches. In the statistical results of the hold-out validation, the DNN model showed the higher accuracy (root mean square error (RMSE) = 0.112, mean bias error (MBE) = 0.007, and correlation coefficient $(R) = 0.863$ ) relative to the traditional SVR (RMSE = 0.123, MBE = −0.010, and $R = 0.833$ ) and RF (RMSE = 0.125, MBE = 0.004, and $R = 0.825$ ) models. The DNN model also exhibited the best performance for most statistical metrics among the traditional SVR, RF, and selected physical models (except for the correlation coefficients and index of agreement) in the spatial and temporal cross-validation analyses. Although the DNN model was trained using the match-up dataset between the top of atmosphere (TOA) reflectance from GOCI multispectral bands and AErosol RObotic NETwork measurements, it showed high spatial and temporal generalization performance owing to its deeper and more complicated network structure. Hourly GOCI AOD data obtained using a deep learning approach with high accuracy are expected to be useful for the quantification of aerosol contents and monitoring of diurnal variations in the AOD.