Time-series Normalized Difference Vegetation Index (NDVI) data extracted from optical satellite images is the most widely used data for predicting crop growth and yield. In the time series pattern of NDVI, variation due to the growth cycle of crops and reduction due to clouds are often seen. In particular, the rainy season in Korea is an important growing period for crops, but it is very difficult to use optical satellite images due to many clouds. In this study, a spatio-temporal gap-filling technique was developed for Sentinel-2A/B NDVI images obtained in rice paddy in Dangjin. The spatial gap-filling technique is used to correct the boundary of a missing area or a small-area missing area by obtaining the NDVI information of normal pixels at the periphery. Afterward, pixels in large areas of missing areas are temporally gap-filled by applying a Gaussian Process Regression (GPR) model to data acquired before and after the target period. The spatio-temporal gap-filling technique developed in this study showed a Root Mean Squared Error (RMSE) of less than 0.15 for the periodically composited NDVI image, and the clouds were removed and the corrected image could be confirmed with visual interpretation. In addition, for daily NDVI images with a large number of clouds, it was found that RMSE was 0.11 to 0.17 depending on the amount of clouds. The algorithm was developed as a program using Python and takes less than 5 minutes to process. In the future, it will be provided to experts who use agricultural and forestry satellite images. This algorithm will be tested on a variety of land cover areas, including mountainous areas as well as agricultural areas, and will be applied to longer time series data. Through this, we plan to analyze the applicability not only to the vegetation index but also to other biophysical variables required for crop monitoring.

The Korea Multi-Purpose Satellite (KOMPSAT) series developed according to the Master Plan for the Promotion of Space Development started with launching the first satellite in December 1999, and five satellites have been launched. Currently, KOMPSAT-3, 3A, and 5 are in operation. The KOMPSAT series equipped with high-resolution optical, synthetic aperture radar (SAR), and middle-wave infrared (MIR) sensors have been utilized in various fields such as land, environment, forestry, agriculture, and disaster over the past 25 years, and have recently been utilized in various studies such as image segmentation, object detection, and change detection in conjunction with deep learning technology. In this special issue, we would like to introduce various research results recently conducted using images from the KOMPSAT series.

Ground control points (GCPs) are essential for the precise georeferencing of drone imagery, but traditional manual measurement methods are inefficient and resource-intensive. Research on automated georeferencing methods utilizing GCP patches is necessary to address this issue. The purpose of this study is to propose and validate an automated method that improves georeferencing accuracy for drone imagery by leveraging GCP patches. The proposed method combines iterative automatic measurement and bundle adjustment to improve measurement accuracy incrementally. A Random Forest-based binary classifier was employed to incorporate only reliable measurement results, enhancing the overall georeferencing accuracy. The method’s performance was validated using datasets with significant temporal gaps, and iterative refinement was performed to progressively adjust the matching range, exterior orientation parameters, and interior orientation parameters. Experimental results show that the Random Forest classifier achieved a high Area Under the ROC Curve (AUC) of 0.9, and the final positional accuracy, based on checkpoints, reached 6 cm. Additionally, the success rate and precision of automatic measurements improved significantly through the iterative process.

Satellite data are used in precision agriculture to optimize crop management. Thus, the planting pattern (e.g., flat and ridge-furrow) and crop type should be accurately reflected in the data. The purpose of this study was to identify the spatial characteristics of errors in the surface reflectance (SR) and vegetation index (VI) obtained from the Sentinel-2 satellite. Drone data were used to evaluate the suitability of the Sentinel-2 satellite for precision agriculture applications in agricultural fields. Four VIs (normalized difference vegetation index, green normalized difference vegetation index, enhanced vegetation index, and normalized difference red edge index) were calculated. The rice paddy exhibited a homogeneous surface, whereas garlic/onion and soybean fields showed high surface heterogeneity because of the presence of ridges and furrows. The SR values of the rice paddy, measured at near-infrared (NIR) wavelength using the Sentinel-2 satellite, were saturated. The VIs derived from both satellite and drone data exhibited a correlation above 0.811 and normalized root mean square error (NRMSE) below 11.1% after bias correction. The garlic and onion fields exhibited the worst results, with a bias-corrected NRMSE for VIs ranging between 12.9% and 13.8%. The soybean field, where the vegetation covered the surface almost completely, exhibited the best relationship between the Sentinel-2 and drone data. The correlation coefficient and bias-corrected NRMSE of VIs for the combination of the two devices were above 0.969 and below 6.4%, respectively. In addition, the SR at NIR had a correlation of 0.925 and a slope of 1.157, unlike in the rice paddy. These results indicate that crop structure has a greater effect than the planting pattern. The absolute difference between the VIs measured by the satellite and drone is influenced by the degree of surface heterogeneity. The errors are more pronounced at the farm-land edges. Our study contributes to a better understating of the characteristics of Sentinel-2 data for use in agricultural fields.

In this study, a synthetic aperture radar (SAR) phase refocusing algorithm using vessel velocity was proposed and applied to three different satellite SAR images in order to enhance the recognition performance of moving vessels. Exploiting the Doppler frequency character of target velocity, it effectively calibrates azimuth defocusing and aids vessel recognition by using the well-focused image as training data. Refocusing performance on vessels exceeded that of the conventional autofocusing algorithm, while vessel recognition accuracy was enhanced after performing the refocusing.

In this study, we propose an automatic training dataset generation method to build a self-supervised matching network, using an End-to-end approach, to extract matching points between very high-resolution (VHR) satellite images. A homography matrix that transforms the scale, rotation, and translation of a single VHR remote-sensing image is applied to generate reference and sensed image patches. After adjusting the contrast and brightness of the sensed image patch, Gaussian and speckle noise are added, and shading and motion blur effects are applied to give it different characteristics from the reference image patch. Subsequently, multiple feature point extractors are combined with homographic adaptation to extract robustly detected feature points by different detectors under various geometric conditions from each image patch. The extracted feature points are optimized using the non-maximum suppression (NMS) technique. Feature point pairs with distance errors within 1 pixel between image patches are identified as matching points using the inverse homography matrix. The coordinates of these matching points, along with the homography matrix, are then employed as pseudo-labels. Training data was generated only when the automated method, applied to the VHR remote sensing database collected from various sources, extracted more than 20 matching points. As a result, training and validation datasets were generated, comprising a total of 341,820 and 44,389 image patches, respectively. The End-to-end matching network trained with the proposed dataset extracted matching points more accurately compared to other matching methods and deep learning networks. Therefore, the proposed method can automatically generate high-quality pseudo-labels that reflect the characteristics of VHR satellite images, thereby improving the training efficiency of deep learning networks.

The importance of growth and water management for open-field crops is increasing due to climate change. Although automatic irrigation systems based on soil moisture sensors are effective for water management, they have limitations in spatially representing the entire field. To supplement this, drone imagery can be utilized. In this study, we evaluated the response of outputs based on RGB, multispectral, and thermal imagery according to the growth stages and water status of Chinese cabbage. The normalized difference vegetation index (NDVI) was useful for monitoring the initial growth stage of cabbage, while the normalized difference red edge index contributed to a more detailed assessment of the cabbage’s growth status by reflecting chlorophyll content. Plant height, estimated through the crop height model, showed the growth status during the bulbing stage under different irrigation treatments more clearly than NDVI and the height of the Chinese cabbage consistently irrigated under dry weather conditions was taller. The vegetation index and plant height from drone imagery effectively detected spatial variations within the same treatment as well as growth differences between plots with and without irrigation. The crop water stress index, derived from drone thermal imagery, immediately reflected changes in Chinese cabbage water stress after irrigation and rainfall. These results are expected to contribute not only to the utilization of various products observed by drones but also to the growth and water management for open-field Chinese cabbage farming.

Busan is the 6th largest port city in the world, where nitrogen dioxide (NO₂) emissions from transportation and port industries are significant. This study aims to assess the NO₂ products of the Geostationary Environment Monitoring Spectrometer (GEMS) over Busan using ground-based instruments (i.e., surface in-situ network and Pandora). The GEMS vertical column densities of NO₂ showed reasonable consistency in the spatiotemporal variations, comparable to the previous studies. The GEMS data showed a consistent seasonal trend of NO₂ with the Korea Ministry of Environment network and Pandora in 2022, which is higher in winter and lower in summer. These agreements prove the capability of the GEMS data to monitor the air quality in Busan. The correlation coefficient and the mean bias error between the GEMS and Pandora NO₂ over Busan in 2022 were 0.53 and 0.023 DU, respectively. The GEMS NO₂ data were also positively correlated with the ground-based in-situ network with a correlation coefficient of 0.42. However, due to the significant spatiotemporal variabilities of the NO₂, the GEMS footprint size can hardly resolve small-scale variabilities such as the emissions from the road and point sources. In addition, relative biases of the GEMS NO₂ retrievals to the Pandora data showed seasonal variabilities, which is attributable to the air mass factor estimation of the GEMS. Further studies with more measurement locations for longer periods of data can better contribute to assessing the GEMS NO₂ data. Reliable GEMS data can further help us understand the Asian air quality with the diurnal variabilities.

Wildfires have grown in scale and intensity over recent decades, profoundly impacting global carbon cycles. For instance, the 2019–2020 Australian wildfire burned 18.6 million hectares of forest and released 715 million tonnes of carbon dioxide. Assessing vegetation recovery after wildfire at such scales is challenging due to conventional satellite products’ spatial and temporal resolution limitations. This study addresses these limitations by fusing two datasets, MODerate Resolution Imaging Spectroradiometer (MODIS) MCD43A4 product and Landsat 8/9 Nadir Bidirectional Reflectance Distribution (NBAR) using an enhanced Flexible Spatiotemporal Data Fusion (FSDAF) algorithm that incorporates sub-pixel class fraction change information (SFSDAF). Before the fusion process, preprocessing involved detecting cloudcontaminated pixels using the Function of Mask (Fmask) algorithm and the Bidirectional Reflectance Distribution Function (BRDF) correction of Landsat. After the fusion process, Neighborhood Similar Pixel Interpolator (NSPI) was employed for the gap-filling process. The fusion images produced high-resolution spatiotemporal vegetation indices with a 30-meter spatial resolution and daily temporal coverage. By comparing these datasets with flux tower measurements, the study examined vegetation recovery in Tumbarumba, southeastern Australia. Findings revealed that the fusion dataset captured fine-scale spatial variability and temporal dynamics, providing detailed insights into localized recovery patterns and seasonal changes. The results highlighted rapid vegetation greenness recovery, though flux tower data indicated slower photosynthesis and carbon sequestration recovery. This study emphasizes the importance of highresolution imagery for accurate recovery monitoring and highlights the necessity of integrating multiple datasets for a comprehensive understanding.

This study evaluated the performance of LiDAR sensors by mounting the DJI Zenmuse L1 and L2 on a DJI Matrice 350 RTK drone to capture ground point cloud data over flat terrain and forested areas. Scans were conducted at altitudes of 50 m and 80 m, with 70 cm × 70 cm aerial targets placed on open ground and under forest vegetation to measure point density. We assessed the multiple reflections in the captured LiDAR data and reviewed absolute accuracy using ground control points. Results showed that the L2 sensor achieved higher point density than the L1 sensor in open areas and performed well at low altitudes (50 m). For simple terrain like flat ground, adequate data was obtained with only the first or second reflections. In forested areas, the L2 sensor effectively captured ground data beneath dense vegetation, whereas the L1 sensor struggled to accurately detect ground under these conditions. In terms of absolute accuracy, the L2 sensor showed less noise and higher precision on flat, stable terrain than the L1. On more complex terrain, the L2 enabled accurate 3D modeling at higher point density. Overall, the flexible use of both L1 and L2 sensors is expected to improve disaster site analysis and response across varied landscapes.