As the performance of supervised classification using convolutional neural networks (CNNs) are affected significantly by training patches, it is necessary to analyze the effects of the information content of training patches in patch-based classification. The objective of this study is to quantitatively investigate the effects of class purity of a training patch on performance of crop classification. Here, class purity that refers to a degree of compositional homogeneity of classes within a training patch is considered as a primary factor for the quantification of information conveyed by training patches. New quantitative indices for class homogeneity and variations of local class homogeneity over the study area are presented to characterize the spatial homogeneity of the study area. Crop classification using 2D-CNN was conducted in two regions (Anbandegi in Korea and Illinois in United States) with distinctive spatial distributions of crops and class homogeneity over the area to highlight the effect of class purity of a training patch. In the Anbandegi region with high class homogeneity, superior classification accuracy was obtained when using large size training patches with high class purity (7.1%p improvement in overall accuracy over classification with the smallest patch size and the lowest class purity). Training patches with high class purity could yield a better identification of homogenous crop parcels. In contrast, using small size training patches with low class purity yielded the highest classification accuracy in the Illinois region with low class homogeneity (19.8%p improvement in overall accuracy over classification with the largest patch size and the highest class purity). Training patches with low class purity could provide useful information for the identification of diverse crop parcels. The results indicate that training samples in patch-based classification should be selected based on the class purity that reflects the local class homogeneity of the study area.

Constructing optical image time series for cropland monitoring requires a cloud removal method that accurately restores cloud regions and eliminates discontinuity around cloud boundaries. This paper describes a two-stage hybrid machine learning-based cloud removal method that combines Gaussian process regression (GPR)-based predictions with image blending for seamless optical image reconstruction. GPR is employed in the first stage to generate initial prediction results by quantifying temporal relationships between multi-temporal images. GPR predictive uncertainty is particularly combined with prediction values to utilize uncertainty-weighted predictions as the input for the next stage. In the second stage, Poisson blending is applied to eliminate discontinuity in GPR-based predictions. The benefits of this method are illustrated through cloud removal experiments using Sentinel-2 images with synthetic cloud masks over two cropland sites. The proposed method was able to maintain the structural features and quality of the underlying reflectance in cloud regions and outperformed two existing hybrid cloud removal methods for all spectral bands. Furthermore, it demonstrated the best performance in predicting several vegetation indices in cloud regions. These experimental results indicate the benefits of the proposed cloud removal method for reconstructing cloud-contaminated optical imagery.

Soyeon Park, Geun-Ho Kwak, Eui Ho Hwang, No-Wook Park. Korean J. Remote Sens. 2024;40:881-93. https://doi.org/10.7780/kjrs.2024.40.6.1.1

Continuous crop monitoring often requires a time-series set of satellite images. Since satellite images have a trade-off in spatial and temporal resolution, spatiotemporal image fusion (STIF) has been applied to construct time-series images at a consistent scale. With the increased availability of high spatial resolution images, it is necessary to develop a new STIF model that can effectively reflect the properties of high spatial resolution satellite images for small-scale crop field monitoring. This paper proposes an advanced STIF model using a single image pair, called high spatial resolution image fusion using object-based weighting (HIFOW), for blending high spatial resolution satellite images. The four-step weighted-function approach of HIFOW includes (1) temporal relationship modeling, (2) object extraction using image segmentation, (3) weighting based on object information, and (4) residual correction to quantify temporal variability between the base and prediction dates and also represent both spectral patterns at the prediction date and spatial details of fine-scale images. The specific procedures tailored for blending fine-scale images are the extraction of object-based change and structural information and their application to weight determination. The potential of HIFOW was evaluated from the experiments on agricultural sites using Sentinel-2 and RapidEye images. HIFOW was compared with three existing STIF models, including the spatial and temporal adaptive reflectance fusion model (STARFM), flexible spatiotemporal data fusion (FSDAF), and Fit-FC. Experimental results revealed that the HIFOW prediction could restore detailed spatial patterns within crop fields and clear crop boundaries with less spectral distortion, which was not represented in the prediction results of the other three models. Consequently, HIFOW achieved the best prediction performance in terms of accuracy and structural similarity for all the spectral bands. Other than the reflectance prediction, HIFOW also yielded superior prediction performance for blending normalized difference vegetation index images. These findings indicate that HIFOW could be a potential solution for constructing high spatial resolution time-series images in small-scale croplands.

The construction of high spatial and temporal image time series through spatiotemporal image fusion is vital for environmental monitoring and time-series analysis. However, most existing spatiotemporal image fusion models were originally developed for mid- and coarse spatial resolution satellite images, limiting their applicability to high spatial resolution images. To address this limitation, it is critical to identify models that are robust across different fusion strategies when applied to high spatial resolution images. This study evaluates the robustness of three representative weight function-based models, the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM), Flexible Spatiotemporal Data Fusion (FSDAF), and Fit-FC, using PlanetScope and Sentinel-2 images from agricultural sites with varying degrees of spatial heterogeneity. The analysis focuses on three key factors related to input data characteristics that influence fusion performance: differences in image acquisition times, differences in spatial resolution ratio, and the impact of radiometric normalization. A relative robustness index (RRI), defined as the coefficient of variation of multi-band prediction results across experimental scenarios, is introduced to facilitate the quantitative comparison of model robustness. Experimental results show that model robustness depends on both landscape heterogeneity and input data characteristics. Fit-FC was the most robust among the three models, particularly in heterogeneous landscapes. These findings provide practical insights for selecting robust models for spatiotemporal fusion in real-world applications using multi-sensor high spatial resolution images.

Cloud removal is often required to construct time-series sets of optical images for environmental monitoring. In regression-based cloud removal, the selection of an appropriate regression model and the impact analysis of the input images significantly affect the prediction performance. This study evaluates the potential of Gaussian process (GP) regression for cloud removal and also analyzes the effects of cloud-free optical images and spectral bands on prediction performance. Unlike other machine learning-based regression models, GP regression provides uncertainty information and automatically optimizes hyperparameters. An experiment using Sentinel-2 multi-spectral images was conducted for cloud removal in the two agricultural regions. The prediction performance of GP regression was compared with that of random forest (RF) regression. Various combinations of input images and multi-spectral bands were considered for quantitative evaluations. The experimental results showed that using multi-temporal images with multi-spectral bands as inputs achieved the best prediction accuracy. Highly correlated adjacent multi-spectral bands and temporally correlated multi-temporal images resulted in an improved prediction accuracy. The prediction performance of GP regression was significantly improved in predicting the near-infrared band compared to that of RF regression. Estimating the distribution function of input data in GP regression could reflect the variations in the considered spectral band with a broader range. In particular, GP regression was superior to RF regression for reproducing structural patterns at both sites in terms of structural similarity. In addition, uncertainty information provided by GP regression showed a reasonable similarity to prediction errors for some sub-areas, indicating that uncertainty estimates may be used to measure the prediction result quality. These findings suggest that GP regression could be beneficial for cloud removal and optical image reconstruction. In addition, the impact analysis results of the input images provide guidelines for selecting optimal images for regression-based cloud removal.