Motivated by the lack of long-term global soil moisture products with both high spatial and temporal resolutions, a global 1 km daily spatiotemporally continuous soil moisture product (GLASS SM) was generated from 2000 to 2020 using an ensemble learning model (eXtreme Gradient Boosting – XGBoost). The model was developed by integrating multiple datasets, including albedo, land surface temperature, and leaf area index products from the Global Land Surface Satellite (GLASS) product suite, as well as the European reanalysis (ERA5-Land) soil moisture product, in situ soil moisture dataset from the International Soil Moisture Network (ISMN), and auxiliary datasets (Multi-Error-Removed Improved-Terrain (MERIT) DEM and Global gridded soil information (SoilGrids)). Given the relatively large-scale differences between point-scale in situ measurements and other datasets, the triple collocation (TC) method was adopted to select the representative soil moisture stations and their measurements for creating the training samples. To fully evaluate the model performance, three validation strategies were explored: random, site independent, and year independent. Results showed that although the XGBoost model achieved the highest accuracy on the random test samples, it was clearly a result of model overfitting. Meanwhile, training the model with representative stations selected by the TC method could considerably improve its performance for site- or year-independent test samples. The overall validation accuracy of the model trained using representative stations on the site-independent test samples, which was least likely to be overfitted, was a correlation coefficient (R) of 0.715 and root mean square error (RMSE) of 0.079 m³ m⁻³. Moreover, compared to the model developed without station filtering, the validation accuracies of the model trained with representative stations improved significantly for most stations, with the median R and unbiased RMSE (ubRMSE) of the model for each station increasing from 0.64 to 0.74 and decreasing from 0.055 to 0.052 m³ m⁻³, respectively. Further validation of the GLASS SM product across four independent soil moisture networks revealed its ability to capture the temporal dynamics of measured soil moisture (R=0.69–0.89; ubRMSE = 0.033–0.048 m³ m⁻³). Lastly, the intercomparison between the GLASS SM product and two global microwave soil moisture datasets – the 1 km Soil Moisture Active Passive/Sentinel-1 L2 Radiometer/Radar soil moisture product and the European Space Agency Climate Change Initiative combined soil moisture product at 0.25^∘ – indicated that the derived product maintained a more complete spatial coverage and exhibited high spatiotemporal consistency with those two soil moisture products. The annual average GLASS SM dataset from 2000 to 2020 can be freely downloaded from https://doi.org/10.5281/zenodo.7172664 (Zhang et al., 2022a), and the complete product at daily scale is available at http://glass.umd.edu/soil_moisture/ (last access: 12 May 2023).

Satellite vegetation index (VI) products, such as normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI), have been widely used. However, they are severely contaminated by clouds and other factors and provide false signals of the surface vegetation conditions. In this study, the new global seamless 250 m, eight-day NDVI and EVI products from 2000–2021 were developed from Moderate Resolution Imaging Spectroradiometer (MODIS) surface reflectance data using a long short-term memory (LSTM) neural network method. High-quality globally representative time series VI samples were constructed to train the model using a combination of the Savitzky-Golay filter (SG), Global LAnd Surface Satellite (GLASS) leaf area index (LAI) fitting and upper envelope methods. To evaluate the proposed method and the 250 m VI products, the MODIS VI product (MOD13Q1) was used for the inter-comparisons using four widely used VI reconstruction methods. Assuming that the MODIS VI data of high quality represents the true values, the root mean square error (RMSE) for NDVI and EVI generated by the LSTM model are 0.0734 and 0.0509, respectively.

Surface air temperature (T_a), as an important climate variable, has been used in a wide range of fields such as ecology, hydrology, climatology, epidemiology, and environmental science. However, ground measurements are limited by poor spatial representation and inconsistency, and reanalysis and meteorological forcing datasets suffer from coarse spatial resolution and inaccuracy. Previous studies using satellite data have mainly estimated T_a under clear-sky conditions or with limited temporal and spatial coverage. In this study, an all-sky daily mean land T_a product at a 1 km spatial resolution over mainland China for 2003–2019 has been generated mainly from the Moderate Resolution Imaging Spectroradiometer (MODIS) products and the Global Land Data Assimilation System (GLDAS) dataset. Three T_a estimation models based on random forest were trained using ground measurements from 2384 stations for three different clear-sky and cloudy-sky conditions. The random sample validation results showed that the R² and root-mean-square error (RMSE) values of the three models ranged from 0.984 to 0.986 and from 1.342 to 1.440 K, respectively. We examined the spatiotemporal patterns and land cover type dependences of model accuracy. Two cross-validation (CV) strategies of leave-time-out (LTO) CV and leave-location-out (LLO) CV were also used to evaluate the models. Finally, we developed the all-sky T_a dataset from 2003 to 2009 and compared it with the China Land Data Assimilation System (CLDAS) dataset at a 0.0625^∘ spatial resolution, the China Meteorological Forcing Data (CMFD) dataset at a 0.1^∘ spatial resolution, and the GLDAS dataset at a 0.25^∘ spatial resolution. Validation accuracy of our product in 2010 was significantly better than other datasets, with R² and RMSE values of 0.992 and 1.010 K, respectively. In summary, the developed all-sky daily mean land T_a dataset has achieved satisfactory accuracy and high spatial resolution simultaneously, which fills the current dataset gap in this field and plays an important role in the studies of climate change and the hydrological cycle. This dataset is currently freely available at https://doi.org/10.5281/zenodo.4399453 (Chen et al., 2021b) and the University of Maryland (http://glass.umd.edu/Ta_China/, last access: 24 August 2021). A sub-dataset that covers Beijing generated from this dataset is also publicly available at https://doi.org/10.5281/zenodo.4405123 (Chen et al., 2021a).

Land surface temperature (LST) serves as a crucial variable in characterizing climatological, agricultural, ecological, and hydrological processes. Thermal infrared (TIR) remote sensing provides high temporal and spatial resolutions for obtaining LST information. Nevertheless, TIR-based satellite LST products frequently exhibit missing values due to cloud interference. Prior research on estimating all-weather instantaneous LST has predominantly concentrated on regional or continental scales. This study involved generating a global all-weather instantaneous and daily mean LST product spanning from 2000 to 2020 using XGBoost. Multisource data, including Moderate-Resolution Imaging Spectroradiometer (MODIS) top-of-atmosphere (TOA) observations, surface radiation products, and reanalysis data, were employed. Validation using an independent dataset of 77 individual stations demonstrated the high accuracy of our products, yielding root mean squared errors (RMSEs) of 2.787 K (instantaneous) and 2.175 K (daily). The RMSE for clear-sky conditions was 2.614 K for the instantaneous product, which is slightly lower than the cloudy-sky RMSE of 2.931 K. Our instantaneous and daily mean LST products exhibit higher accuracy compared to the MODIS official LST product (instantaneous RMSE = 3.583 K; daily 3.105 K) and the land component of the fifth generation of the European ReAnalysis (ERA5-Land) LST product (instantaneous RMSE = 4.048 K; daily 2.988 K). Significant improvements are observed in our LST product, notably at high latitudes, compared to the official MODIS LST product. The LST dataset from 2000 to 2020 at the monthly scale, the daily mean LST on the first day of 2010 can be freely downloaded from https://doi.org/10.5281/zenodo.4292068 (Li et al., 2024), and the complete product will be available at https://glass-product.bnu.edu.cn/ (last access: 22 August 2024).

Inaccurate parameter estimation is a significant source of uncertainty in complex terrestrial biosphere models. Model parameters may have large spatial variability, even within a vegetation type. Model uncertainty from parameters can be significantly reduced by model–data fusion (MDF), which, however, is difficult to implement over a large region with traditional methods due to the high computational cost. This study proposed a hybrid modeling approach that couples a terrestrial biosphere model with a data-driven machine learning method, which is able to consider both satellite information and the physical mechanisms. We developed a two-step framework to estimate the essential parameters of the revised Integrated Biosphere Simulator (IBIS) pixel by pixel using the satellite-derived leaf area index (LAI) and gross primary productivity (GPP) products as “true values.” The first step was to estimate the optimal parameters for each sample using a modified adaptive surrogate modeling algorithm (MASM). We applied the Gaussian process regression algorithm (GPR) as a surrogate model to learn the relationship between model parameters and errors. In our second step, we built an extreme gradient boosting (XGBoost) model between the optimized parameters and local environmental variables. The trained XGBoost model was then used to predict optimal parameters spatially across the deciduous forests in the eastern United States. The results showed that the parameters were highly variable spatially and quite different from the default values over forests, and the simulation errors of the GPP and LAI could be markedly reduced with the optimized parameters. The effectiveness of the optimized model in estimating GPP, ecosystem respiration (ER), and net ecosystem exchange (NEE) were also tested through site validation. The optimized model reduced the root mean square error (RMSE) from 7.03 to 6.22 gC m⁻² d⁻¹ for GPP, 2.65 to 2.11 gC m⁻² d⁻¹ for ER, and 4.45 to 4.38 gC m⁻² d⁻¹ for NEE. The mean annual GPP, ER, and NEE of the region from 2000 to 2019 were 5.79, 4.60, and −1.19 Pg yr⁻¹, respectively. The strategy used in this study requires only a few hundred model runs to calibrate regional parameters and is readily applicable to other complex terrestrial biosphere models with different spatial resolutions. Our study also emphasizes the necessity of pixel-level parameter calibration and the value of remote sensing products for per-pixel parameter optimization.

The fraction of absorbed photosynthetically active radiation (FAPAR) is a critical land surface variable for carbon cycle modeling and ecological monitoring. Several global FAPAR products have been released and have become widely used; however, spatiotemporal inconsistency remains a large issue for the current products, and their spatial resolutions and accuracies can hardly meet the user requirements. An effective solution to improve the spatiotemporal continuity and accuracy of FAPAR products is to take better advantage of the temporal information in the satellite data using deep learning approaches. In this study, the latest version (V6) of the FAPAR product with a 250 m resolution was generated from Moderate Resolution Imaging Spectroradiometer (MODIS) surface reflectance data and other information, as part of the Global LAnd Surface Satellite (GLASS) product suite. In addition, it was aggregated to multiple coarser resolutions (up to 0.25^∘ and monthly). Three existing global FAPAR products (MODIS Collection 6; GLASS V5; and PRoject for On-Board Autonomy–Vegetation, PROBA-V, V1) were used to generate the time-series training samples, which were used to develop a bidirectional long short-term memory (Bi-LSTM) model. Direct validation using high-resolution FAPAR maps from the Validation of Land European Remote sensing Instrument (VALERI) and ImagineS networks revealed that the GLASS V6 FAPAR product has a higher accuracy than PROBA-V, MODIS, and GLASS V5, with an R² value of 0.80 and root-mean-square errors (RMSEs) of 0.10–0.11 at the 250 m, 500 m, and 3 km scales, and a higher percentage (72 %) of retrievals for meeting the accuracy requirement of 0.1. Global spatial evaluation and temporal comparison at the AmeriFlux and National Ecological Observatory Network (NEON) sites revealed that the GLASS V6 FAPAR has a greater spatiotemporal continuity and reflects the variations in the vegetation better than the GLASS V5 FAPAR. The higher quality of the GLASS V6 FAPAR is attributed to the ability of the Bi-LSTM model, which involves high-quality training samples and combines the strengths of the existing FAPAR products, as well as the temporal and spectral information from the MODIS surface reflectance data and other information. The 250 m 8 d GLASS V6 FAPAR product for 2020 is freely available at https://doi.org/10.5281/zenodo.6405564 and https://doi.org/10.5281/zenodo.6430925 (Ma, 2022a, b) as well as at the University of Maryland for 2000–2021 (http://glass.umd.edu/FAPAR/MODIS/250m, last access 1 November 2022).