Spatiotemporal Fusion Model Research Articles

A modified flexible spatiotemporal data fusion model

Remote sensing spatiotemporal fusion models blend multi-source images of different spatial resolutions to create synthetic images with high resolution and frequency, contributing to time series research where high quality observations are not available with sufficient frequency. However, existing models are vulnerable to spatial heterogeneity and land cover changes, which are frequent in human-dominated regions. To obtain quality time series of satellite images in a human-dominated region, this study developed the Modified Flexible Spatial-temporal Data Fusion (MFSDAF) approach based on the Flexible Spatial-temporal Data Fusion (FSDAF) model by using the enhanced linear regression (ELR). Multiple experiments of various land cover change scenarios were conducted based on both actual and simulated satellite images, respectively. The proposed MFSDAF model was validated by using the correlation coefficient (r), relative root mean square error (RRMSE), and structural similarity (SSIM), and was then compared with the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) and FSDAF models. Results show that in the presence of significant land cover change, MFSDAF showed a maximum increase in r, RRMSE, and SSIM of 0.0313, 0.0109 and 0.049, respectively, compared to FSDAF, while ESTARFM performed best with less temporal difference of in the input images. In conditions of stable landscape changes, the three performance statistics indicated a small advantage of MFSDAF over FSDAF, but were 0.0286, 0.0102, 0.0317 higher than for ESTARFM, respectively. MFSDAF showed greater accuracy of capturing subtle changes and created high-precision images from both actual and simulated satellite images.

A New Approach for Surface Urban Heat Island Monitoring Based on Machine Learning Algorithm and Spatiotemporal Fusion Model

Land surface temperature (LST) is an important indicator for assessing the surface urban heat island (SUHI) effect. This paper presents a novel approach to derive LST estimates by integrating machine learning algorithm and spatiotemporal fusion model at high spatial and temporal resolution. The spatial resolutions of Landsat TM and Landsat 8 LST data were first downscaled using random forest (RF) algorithm from 120 m and 100 m, respectively, to 30 m. The resultant LST data were fused with MODerate-resolution Imaging Spectroradiometer (MODIS) LST data, by means of the Flexible Spatiotemporal Data Fusion method (FSDAF), in order to generate high spatiotemporal resolution summer daytime LST data covering the center of Chengdu city in China. The proposed new method was used to estimate the spatiotemporal variations of the summer daytime SUHI from 2009 to 2018 over Chengdu city. Results show that: (1) RF performs way better than the classical downscaling algorithm-thermal sharpening algorithm (TsHARP) for LST, and produces higher accuracy for different land covers; (2) the fused high spatiotemporal resolution summer daytime LST values were evaluated with in situ LST obtained from Chengdu Meteorological Office and the final validation results indicated that the proposed method, in generating LST dataset, can provide more details of urban thermal environment and produce higher accuracy than the traditional FSDAF; (3) significantly increasing trends of summer daytime SUHI intensity (SUHII) in the study area were observed. SUHII increased from 2.78 °C in 2009 to 4.04 °C in 2018. The highest and lowest summer daytime LST estimates were recorded over impervious surface area (ISA) and waters, respectively.

Spatiotemporal Fusion With Only Two Remote Sensing Images as Input

Spatiotemporal data fusion is an effective way of generating a dense time series with a high spatial resolution. Traditionally, the spatiotemporal fusion models, especially the popular ones such as the spatial and temporal adaptive reflectance fusion model, require at least three images as input, i.e., a coarse-resolution image on the target date and a pair of fine- and coarse-resolution images on the reference date. However, this cannot always be satisfied, as the high-quality coarse-resolution image on the reference date may be unavailable in some application scenarios. This led to efforts to achieve data fusion only using the other two images as input. In this article, we proposed an effective strategy that can be combined with any spatiotemporal fusion model to accomplish the fusion with simplified input. To confirm the validity of the method, we comprehensively compared the fusion performances under the two input modalities. In total, 38 tests were conducted with Moderate Resolution Imaging Spectroradiometer (MODIS), Landsat, and Sentinel-2 land surface reflectance products. Results suggest that by applying the proposed method, the fusion performance with only two input images is comparable or even superior to that with three input images. This article challenges the stereotype that spatiotemporal data fusion strictly needs at least three input images. The proposed method extends the application scenarios of spatiotemporal fusion, and creates opportunities to fuse sensors with barely overlapping temporal coverages, such as the Landsat 8 Operational Land Imager and the Sentinel-2 MultiSpectral Instrument.

Monitoring Seasonal and Diurnal Surface Urban Heat Islands Variations Using Landsat-Scale Data in Hefei, China, 2000–2017

The surface urban heat island (SUHI) is one of the most common effects of the urban ecological environment, and its long-term diurnal and seasonal fine-scale analysis remains poorly understood. This article utilized a modified reconstruction model and spatiotemporal fusion model to generate twice daily (day and night) and seasonal Landsat-scale (30 m) land surface temperature (LST) data using Moderate Resolution Imaging Spectroradiometer (MODIS) and Landsat images in Hefei, China, from 2000 to 2017. The diurnal and seasonal characteristics of SUHI variations and the complex interrelationships between SUHI and landscape patterns were analyzed. The result shows that fine-scale SUHI varied diurnally and seasonally, with a great deal of spatial heterogeneity. The fine-scale SUHI phenomenon was obvious in spring and summer during the daytime (average annual intensity of the heat island is 5.05 °C in summer, 3.26 °C in spring, 2.08 °C in autumn, and 1.90 °C in winter), and appeared at nighttime throughout the year (average annual intensity of the heat island is 2.85 °C in summer, 2.75 °C in spring, 2.69 °C in autumn, and 2.30 °C in winter). The change characteristics of landscape pattern indices in different SUHI zones reveal that construction land has an amplifying effect for SUHI, while areas dominated by vegetation and water weaken the effects of SUHI. The mitigation effect of vegetation on SUHI during the day is better than that at night.

An Enhanced Deep Convolutional Model for Spatiotemporal Image Fusion

Earth observation data with high spatiotemporal resolution are critical for dynamic monitoring and prediction in geoscience applications, however, due to some technique and budget limitations, it is not easy to acquire satellite images with both high spatial and high temporal resolutions. Spatiotemporal image fusion techniques provide a feasible and economical solution for generating dense-time data with high spatial resolution, pushing the limits of current satellite observation systems. Among existing various fusion algorithms, deeplearningbased models reveal a promising prospect with higher accuracy and robustness. This paper refined and improved the existing deep convolutional spatiotemporal fusion network (DCSTFN) to further boost model prediction accuracy and enhance image quality. The contributions of this paper are twofold. First, the fusion result is improved considerably with brand-new network architecture and a novel compound loss function. Experiments conducted in two different areas demonstrate these improvements by comparing them with existing algorithms. The enhanced DCSTFN model shows superior performance with higher accuracy, vision quality, and robustness. Second, the advantages and disadvantages of existing deeplearningbased spatiotemporal fusion models are comparatively discussed and a network design guide for spatiotemporal fusion is provided as a reference for future research. Those comparisons and guidelines are summarized based on numbers of actual experiments and have promising potentials to be applied for other image sources with customized spatiotemporal fusion networks.

Extending the SMAP 9-km soil moisture product using a spatio-temporal fusion model

SMAP satellite has provided us the first 9-km global soil moisture (SM) product, which is retrieved from the combined L-band radiometer and radar observations with a balance between accuracy and resolution. However, SMAP's radar failed on July 7, 2015, making the continuous production of the 9-km SM impossible, which has considerably affected the application of SMAP in hydrological monitoring. This study was aimed at extending the SMAP 9-km SM by developing a non-local filter based spatio-temporal fusion model (STFM). With the auxiliary of the historical 9-kmand 36-km products, the STFM was used to downscale the daily 36-km product to 9-km. Two-year 9-km product was estimated using STFM in the study. It shows that the estimated product has the detailed information retention of 9-km product and the comparable accuracy with 36-km product, which makes it feasible to improve the application potential of the current SMAP SM products.

Multi-Temporal Effects of Urban Forms and Functions on Urban Heat Islands Based on Local Climate Zone Classification.

Urban forms and functions have critical impacts on urban heat islands (UHIs). The concept of a “local climate zone” (LCZ) provides a standard and objective protocol for characterizing urban forms and functions, which has been used to link urban settings with UHIs. However, only a few structure types and surface cover properties are included under the same climate background or only one or two time scales are considered with a high spatial resolution. This study assesses multi-temporal land surface temperature (LST) characteristics across 18 different LCZ types in Beijing, China, from July 2017 to June 2018. A geographic information system-based method is employed to classify LCZs based on five morphological and coverage indicators derived from a city street map and Landsat images, and a spatiotemporal fusion model is adopted to generate hourly 100-m LSTs by blending Landsat, Moderate Resolution Imaging Spectroradiometer (MODIS), and FengYun-2F LSTs. Then, annual and diurnal cycle parameters and heat island and cool island (HI or CI) frequency are linked to LCZs at annual, seasonal, monthly, and diurnal scales. Results indicate that: (1) the warmest zones are compact and mid and low-rise built-up areas, while the coolest zones are water and vegetated types; (2) compact and open high-rise built-up areas and vegetated types have seasonal thermal patterns but with different causes; (3) diurnal temperature ranges are the highest for compact and large low-rise settings but the lowest for water and dense or scattered trees; and (4) HIs are the most frequent summertime and daytime events, while CIs occur primarily during winter days, making them more or less frequent for open or compact and high- or low-rise built-up areas. Overall, the distinguishable LSTs or UHIs between LCZs are closely associated with the structure and coverage properties. Factors such as geolocation, climate, and layout also interfere with the thermal behavior. This study provides comprehensive information on how different urban forms and functions are related to LST variations at different time scales, which supports urban thermal regulation through urban design.

An Improved Spatiotemporal Fusion Approach Based on Multiple Endmember Spectral Mixture Analysis.

High spatial and temporal resolution remotely sensed data is of great significance for the extraction of land use/cover information and the quantitative inversion of biophysical parameters. However, due to the limitation of sensor performance and the influence of rain cloud weather, it is difficult to obtain remote sensing images with both high spatial and temporal resolution. The spatiotemporal fusion model is a crucial method to solve this problem. The spatial and temporal adaptive reflectivity fusion model (STARFM) and its improved models are the most widely used spatiotemporal adaptive fusion models. However, the existing spatiotemporal adaptive reflectivity fusion model and its improved models have great uncertainty in selecting neighboring similar pixels, especially in spatially heterogeneous areas. Therefore, it is difficult to effectively search and determine neighboring spectrally similar pixels in STARFM-like models, resulting in a decrease of imagery fusion accuracy. In this research, we modify the procedure of neighboring similar pixel selection of ESTARFM method and propose an improved ESTARFM method (I-ESTARFM). Based on the land cover endmember types and its fraction values obtained by spectral mixing analysis, the neighboring similar pixels can be effectively selected. The experimental results indicate that the I-ESTARFM method selects neighboring spectrally similar pixels more accurately than STARFM and ESTARFM models. Compared with the STARFM and ESTARFM, the correlation coefficients of the image fused by the I-ESTARFM with that of the actual image are increased and the mean square error is decreased, especially in spatially heterogeneous areas. The uncertainty of spectral similar neighborhood pixel selection is reduced and the precision of spatial-temporal fusion is improved.

Comparison of Spatiotemporal Fusion Models for Producing High Spatiotemporal Resolution Normalized Difference Vegetation Index Time Series Data Sets

It has a great significance to combine multi-source with different spatial resolution and temporal resolution to produce high spatiotemporal resolution Normalized Difference Vegetation Index (NDVI) time series data sets. In this study, four spatiotemporal fusion models were analyzed and compared with each other. The models included the spatial and temporal adaptive reflectance model (STARFM), the enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM), the flexible spatiotemporal data fusion model (FSDAF), and a spatiotemporal vegetation index image fusion model (STVIFM). The objective of is to: 1) compare four fusion models using Landsat-MODIS NDVI image from the Banan district, Chongqing Province; 2) analyze the prediction accuracy quantitatively and visually. Results indicate that STVIFM would be more suitable to produce NDVI time series data sets.

Assimilating multi-source remotely sensed data into a light use efficiency model for net primary productivity estimation

High spatiotemporal resolution satellite data are necessary for the retrieval of vegetation indexes, such as Normalized Difference Vegetation Index (NDVI), to be assimilated into the Carnegie-Ames-Stanford Approach (CASA) model for net primary productivity (NPP) estimation, especially in the growing season. However, current remotely sensed data cannot accurately monitor vegetation changes at high spatiotemporal resolution. To consider both temporal and spatial information, spatiotemporal fusion models have been developed to obtain the temporal information from high temporal resolution data (e.g., MODIS) together with the spatial information from high spatial resolution data (e.g., Landsat). In this paper, synthetic NDVI images with the spatial resolution of Landsat data and the temporal resolution of MODIS data were first produced using spatiotemporal fusion models. Next, phenological features were extracted from synthetic NDVI time series data to improve land cover classification accuracy. Finally, we evaluated the approach of assimilating the synthetic NDVI and land cover classification map into the CASA model for synthetic NPP estimation. The results revealed that the accuracy of the synthetic NPP was better than NPP estimation from non-fusion NDVI data, and improving the land cover classification accuracy could improve the accuracy of the synthetic NPP estimation. Furthermore, the monthly synthetic NPP showed a significant exponential relationship with the temperature, rainfall, and solar radiation of the current and previous month.

REAL-TIME AND SEAMLESS MONITORING OF GROUND-LEVEL PM2.5 USING SATELLITE REMOTE SENSING

Abstract. Satellite remote sensing has been reported to be a promising approach for the monitoring of atmospheric PM2.5. However, the satellite-based monitoring of ground-level PM2.5 is still challenging. First, the previously used polar-orbiting satellite observations, which can be usually acquired only once per day, are hard to monitor PM2.5 in real time. Second, many data gaps exist in satellitederived PM2.5 due to the cloud contamination. In this paper, the hourly geostationary satellite (i.e., Harawari-8) observations were adopted for the real-time monitoring of PM2.5 in a deep learning architecture. On this basis, the satellite-derived PM2.5 in conjunction with ground PM2.5 measurements are incorporated into a spatio-temporal fusion model to fill the data gaps. Using Wuhan Urban Agglomeration as an example, we have successfully derived the real-time and seamless PM2.5 distributions. The results demonstrate that Harawari-8 satellite-based deep learning model achieves a satisfactory performance (out-of-sample cross-validation R2 = 0.80, RMSE = 17.49 μg/m3) for the estimation of PM2.5. The missing data in satellite-derive PM2.5 are accurately recovered, with R2 between recoveries and ground measurements of 0.75. Overall, this study has inherently provided an effective strategy for the realtime and seamless monitoring of ground-level PM2.5.

Dynamic monitoring of the Poyang Lake wetland by integrating Landsat and MODIS observations

The spatial and temporal adaptive reflectance fusion models (STARFM) have limited practical applications, because they often enforce the invalid assumption that land cover change does not occur between prior/posterior and target dates. To deal with this challenge, we proposed a spatiotemporal adaptive fusion model for NDVI products (STAFFN), to better blend highly resolved spatial and temporal information from multiple sensors. Compared with existing spatiotemporal fusion models, the proposed model integrates an initial prediction into a hierarchical selection strategy of similar pixels, and can capture landscape changes very well. Experiments using spatial details and temporal abundance comparison among MODIS, Landsat, and fusion results show that the predicted data can accurately capture temporal changes while preserving fine-spatial-resolution details. Model comparison also shows that STAFFNs produce consistently lower biases than STARFMs and the flexible spatiotemporal data fusion models (FSDAFs). A synthetic NDVI product (342 scenes in total) was produced with STAFFNs having a 16-day revisit frequency at 30-m spatial resolution from 2000 to 2014. With this product, we further provided a 15-year spatiotemporal change monitoring map of the Poyang Lake wetland. Results show that the water area in the dry season tended to lose 38.3 km2 yr−1 in coverage over the past 15 years, decreasing by 18.24% of the lake area between 2001 and 2014. The wetland vegetation group tended to increase in coverage, increasing by 10.08% of the lake area in the past 15 years. Our study indicates the STAFFN model can be reasonably applied in monitoring wetland dynamics, and can be easily adapted for the use with other ecosystems.

Spatiotemporal Satellite Image Fusion Using Deep Convolutional Neural Networks

We propose a novel spatiotemporal fusion method based on deep convolutional neural networks (CNNs) under the application background of massive remote sensing data. In the training stage, we build two five-layer CNNs to deal with the problems of complicated correspondence and large spatial resolution gaps between MODIS and Landsat images. Specifically, we first learn a nonlinear mapping CNN between MODIS and low-spatial-resolution (LSR) Landsat images and then learn a super-resolution CNN between LSR Landsat and original Landsat images. In the prediction stage, instead of directly taking the outputs of CNNs as the fusion result, we design a fusion model consisting of high-pass modulation and a weighting strategy to make full use of the information in prior images. Specifically, we first map the input MODIS images to transitional images via the learned nonlinear mapping CNN and further improve the transitional images to LSR Landsat images via the fusion model; then, via the learned SR CNN, the LSR Landsat images are supersolved to transitional images, which are further improved to Landsat images via the fusion model. Compared with the previous learning-based fusion methods, mainly referring to the sparse-representation-based methods, our CNNs-based spatiotemporal method has the following advantages: 1) automatically extracting effective image features; 2) learning an end-to-end mapping between MODIS and LSR Landsat images; and 3) generating more favorable fusion results. To examine the performance of the proposed fusion method, we conduct experiments on two representative Landsat-MODIS datasets by comparing with the sparse-representation-based spatiotemporal fusion model. The quantitative evaluations on all possible prediction dates and the comparison of fusion results on one key date in both visual effect and quantitative evaluations demonstrate that the proposed method can generate more accurate fusion results.

Affect recognition from facial movements and body gestures by hierarchical deep spatio-temporal features and fusion strategy

Affect presentation is periodic and multi-modal, such as through facial movements, body gestures, and so on. Studies have shown that temporal selection and multi-modal combinations may benefit affect recognition. In this article, we therefore propose a spatio-temporal fusion model that extracts spatio-temporal hierarchical features based on select expressive components. In addition, a multi-modal hierarchical fusion strategy is presented. Our model learns the spatio-temporal hierarchical features from videos by a proposed deep network, which combines a convolutional neural networks (CNN), bilateral long short-term memory recurrent neural networks (BLSTM-RNN) with principal component analysis (PCA). Our approach handles each video as a “video sentence.” It first obtains a skeleton with the temporal selection process and then segments key words with a certain sliding window. Finally, it obtains the features with a deep network comprised of a video-skeleton and video-words. Our model combines the feature level and decision level fusion for fusing the multi-modal information. Experimental results showed that our model improved the multi-modal affect recognition accuracy rate from 95.13% in existing literature to 99.57% on a face and body (FABO) database, our results have been increased by 4.44%, and it obtained a macro average accuracy (MAA) up to 99.71%.

A Rigorously-Weighted Spatiotemporal Fusion Model with Uncertainty Analysis

Interest has been growing with regard to the use of remote sensing data characterized by a fine spatial resolution and frequent coverage for the monitoring of land surface dynamics. However, current satellite sensors are fundamentally limited by a trade-off between their spatial and temporal resolutions. Spatiotemporal fusion thus provides a feasible solution to overcome this limitation, and many blending algorithms have been developed. Among them, the popular spatial and temporal adaptive reflectance fusion model (STARFM) is based on a weighted function; however, it uses an ad hoc approach to estimate the weights of surrounding similar pixels. Additionally, an uncertainty analysis of the predicted result is not provided in the STARFM or any other fusion algorithm. This paper proposes a rigorously-weighted spatiotemporal fusion model (RWSTFM) based on geostatistics to blend the surface reflectances of Moderate Resolution Imaging Spectroradiometer (MODIS) and Landsat-5 Thematic Mapper (TM) imagery. The RWSTFM, which is based on ordinary kriging, derives the weights in terms of a fitted semivariance-distance relationship and calculates the estimation variance, which is a measure of the prediction uncertainty. The RWSTFM was tested using three datasets and compared with two commonly-used spatiotemporal reflectance fusion algorithms: the STARFM and the flexible spatiotemporal data fusion (FSDAF) method. The fusion results show that the proposed RWSTFM consistently outperformed the other algorithms both visually and quantitatively. Additionally, more than 70% of the squared error was accounted for by the estimation variance of the RWSTFM for all three of the datasets.

Comparison and assessment of NDVI time series for seasonal wetland classification

ABSTRACTSatellite-based wetland mapping faces challenges due to the high spatial heterogeneity and dynamic characteristics of seasonal wetlands. Although normalized difference vegetation index (NDVI) time series (NTS) shows great potential in land cover mapping and crop classification, the effectiveness of various NTS with different spatial and temporal resolution has not been evaluated for seasonal wetland classification. To address this issue, we conducted comparisons of those NTS, including the moderate-resolution imaging spectroradiometer (MODIS) NTS with 500 m resolution, NTS fused with MODIS and Landsat data (MOD_LC8-NTS), and HJ-1 NDVI compositions (HJ-1-NTS) with finer resolution, for wetland classification of Poyang Lake. Results showed the following: (1) the NTS with finer resolution was more effective in the classification of seasonal wetlands than that of the MODIS-NTS with 500-m resolution and (2) generally, the HJ-1-NTS performed better than that of the fused NTS, with an overall accuracy of 88.12% for HJ-1-NTS and 83.09% for the MOD_LC8-NTS. Future work should focus on the construction of satellite image time series oriented to highly dynamic characteristics of seasonal wetlands. This study will provide useful guidance for seasonal wetland classification, and benefit the improvements of spatiotemporal fusion models.

A rigorous statistical framework for spatio-temporal pollution prediction and estimation of its long-term impact on health.

In the United Kingdom, air pollution is linked to around 40000 premature deaths each year, but estimating its health effects is challenging in a spatio-temporal study. The challenges include spatial misalignment between the pollution and disease data; uncertainty in the estimated pollution surface; and complex residual spatio-temporal autocorrelation in the disease data. This article develops a two-stage model that addresses these issues. The first stage is a spatio-temporal fusion model linking modeled and measured pollution data, while the second stage links these predictions to the disease data. The methodology is motivated by a new five-year study investigating the effects of multiple pollutants on respiratory hospitalizations in England between 2007 and 2011, using pollution and disease data relating to local and unitary authorities on a monthly time scale.

A hierarchical spatiotemporal adaptive fusion model using one image pair

ABSTRACTImage fusion techniques that blend multi-sensor characteristics to generate synthetic data with fine resolutions have generated great interest within the remote sensing community. Over the past decade, although many advances have been made in the spatiotemporal fusion models, there still remain several shortcomings in existing methods. In this article, a hierarchical spatiotemporal adaptive fusion model (HSTAFM) is proposed for producing daily synthetic fine-resolution fusions. The suggested model uses only one prior or posterior image pair, especially with the aim being to predict arbitrary temporal changes. The proposed model is implemented in two stages. First, the coarse-resolution image is enhanced through super-resolution based on sparse representation; second, a pre-selection of temporal change is performed. It then adopts a two-level strategy to select similar pixels, and blends multi-sensor features adaptively to generate the final synthetic data. The results of tests using both simulated and actual observed data show that the model can accurately capture both seasonal phenology change and land-cover-type change. Comparisons between HSTAFM and other developed models also demonstrate our proposed model produces consistently lower biases.

Bayesian Method for Building Frequent Landsat-Like NDVI Datasets by Integrating MODIS and Landsat NDVI

Studies related to vegetation dynamics in heterogeneous landscapes often require Normalized Difference Vegetation Index (NDVI) datasets with both high spatial resolution and frequent coverage, which cannot be satisfied by a single sensor due to technical limitations. In this study, we propose a new method called NDVI-Bayesian Spatiotemporal Fusion Model (NDVI-BSFM) for accurately and effectively building frequent high spatial resolution Landsat-like NDVI datasets by integrating Moderate Resolution Imaging Spectroradiometer (MODIS) and Landsat NDVI. Experimental comparisons with the results obtained using other popular methods (i.e., the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM), the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM), and the Flexible Spatiotemporal DAta Fusion (FSDAF) method) showed that our proposed method has the following advantages: (1) it can obtain more accurate estimates; (2) it can retain more spatial detail; (3) its prediction accuracy is less dependent on the quality of the MODIS NDVI on the specific prediction date; and (4) it produces smoother NDVI time series profiles. All of these advantages demonstrate the strengths and the robustness of the proposed NDVI-BSFM in providing reliable high spatial and temporal resolution NDVI datasets to support other land surface process studies.

Comparison of Spatiotemporal Fusion Models: A Review

Simultaneously capturing spatial and temporal dynamics is always a challenge for the remote sensing community. Spatiotemporal fusion has gained wide interest in various applications for its superiority in integrating both fine spatial resolution and frequent temporal coverage. Though many advances have been made in spatiotemporal fusion model development and applications in the past decade, a unified comparison among existing fusion models is still limited. In this research, we classify the models into three categories: transformation-based, reconstruction-based, and learning-based models. The objective of this study is to (i) compare four fusion models (STARFM, ESTARFM, ISTAFM, and SPSTFM) under a one Landsat-MODIS (L-M) pair prediction mode and two L-M pair prediction mode using time-series datasets from the Coleambally irrigation area and Poyang Lake wetland; (ii) quantitatively assess prediction accuracy considering spatiotemporal comparability, landscape heterogeneity, and model parameter selection; and (iii) discuss the advantages and disadvantages of the three categories of spatiotemporal fusion models.