Xiaobin Guan

Accurate estimation of global gross primary productivity (GPP) is fundamental for understanding terrestrial carbon cycling. Eddy covariance (EC) flux observations provide reliable site-scale GPP estimates, but the spatially sparse distribution limits their applicability at large scales. Satellite-based solar-induced chlorophyll fluorescence (SIF) has emerged as a promising proxy for large-scale GPP estimation; however, current satellite SIF observations also suffer from limited spatiotemporal coverage, and uncertainties remain in the SIF–GPP conversion. Moreover, conventional machine learning models trained solely on EC observations often exhibit limited spatial generalization due to the scarcity of spatially representative training samples.To address these challenges, this study proposes a satellite–ground jointly constrained framework that integrates EC flux measurements and satellite SIF observations using transfer learning and multi-task learning techniques to exploit the complementary strengths of both data sources for global GPP estimation. First, for TROPOMI SIF data that has global spatial coverage but short temporal records, SIF is treated as a source domain to pre-train the model, which is then fine-tuned using long-term EC-derived GPP data as a target domain. This transfer learning-based model (SIFTML) demonstrates improved spatial generalization compared to models trained solely on SIF or EC data, effectively reducing systematic underestimation and overestimation at high and low GPP levels, respectively, while remaining insensitive to the magnitude scaling of source-domain SIF inputs.Second, for the spatially sparse and track-like distributed OCO-2 SIF observations, a multi-task learning framework based on a mixture-of-experts architecture is developed. A physically constrained loss function derived from the SIF–GPP relationship is introduced to simultaneously achieve seamless SIF reconstruction and high-accuracy GPP estimation by jointly leveraging SIF and EC constraints. Results indicate that the multi-task model outperforms traditional single-task approaches in both GPP estimation and SIF reconstruction.Overall, this study provides a new paradigm for long-term, high-accuracy global GPP estimation by alleviating limitations associated with the spatiotemporal coverage of ground EC and satellite SIF observations, as well as the uncertainties in SIF–GPP conversion, thereby offering improved support for global carbon cycle research.

Satellite solar-induced chlorophyll fluorescence (SIF) provides a new opportunity to quantify vegetation photosynthesis at large scales, but individual satellite observations are usually constrained by their respective limitations regarding spatial continuity or temporal coverage. Although various reconstructed SIF products have already been generated, their spatiotemporal variability remains highly dependent on selected explanatory variables, and complementary information from multiple sensors has not been fully exploited. Here, we developed a multi-source SIF fusion framework to integrate the respective strengths of the Global Ozone Monitoring Experiment-2 (GOME-2), Orbiting Carbon Observatory-2 (OCO-2), and TROPOspheric Monitoring Instrument (TROPOMI) products, to generate a Fusion-based global continuous SIF time series with high quality (FSIF). Specifically, a statistical spatiotemporal fusion (STF) approach is first employed to preliminarily merge the spatial resolution and timespan advantages across GOME-2 and TROPOMI SIF observations, and followed by a Light Gradient Boosting Machine (LightGBM) model to extract the information from auxiliary data and OCO-2 SIF. FSIF covers the full period observed by the three sensors from 2008 to 2022, with a 0.05-degree and 8-day resolution. The spatial distribution of FSIF is consistent with TROPOMI SIF (R²=0.91, p2 difference of 0.09, which benefits from the superiority of OCO-2 SIF in GPP representation. Extensive factorial experiments also declare the importance of fusing each sensor and the significance of coupling STF- machine learning (ML) models.

Fine-resolution estimation of gross primary production (GPP) is essential for advancing our knowledge of ecosystem carbon cycling. However, most existing global GPP products are constrained by coarse temporal resolutions (typically ≥ 8 days), limiting their capacity to capture short-term variations in ecosystem productivity. This paper presents a global new GPP dataset for 2001–2020, based on a modified radiation scalar two-leaf LUE (RTL-LUE) model. The RTL-LUE GPP dataset is generated at an hourly temporal resolution and a spatial resolution of 0.1°, driven by the hourly climate data from ERA5-land, GLASS leaf area index, MODIS land cover, and NOAA atmospheric CO₂ concentration. During 2001–2020, this dataset provides a slightly lower global total GPP (124.77 PgC/yr) than the 8-day TL-LUE GPP dataset (126.92 PgC/yr), primarily due to its ability to capture short-term extreme stresses more effectively. Notably, this dataset reveals that annual GPP variability can reach up to approximately 0.10 g C/m² /h at hourly scales. The RTL-LUE GPP demonstrates robust performance at 184 towers, thereby improving insights into the temporal dynamics of carbon fluxes.

Vegetation photosynthesis is a key Earth system process that can fix carbon dioxide in the atmosphere. Mountainous areas usually have high productivity and extensive vegetation cover, but their study requires a higher spatiotemporal resolution due to the complex climate and vegetation variations with altitude. In this study, we analyzed the variations and climatic responses of vegetation gross primary productivity (GPP) in northwestern Hubei, China, at a 30 m spatial resolution from 2001 to 2020, based on the fusion of multi-source remote sensing data. A GPP estimation framework based on the CASA model was applied, and spatiotemporal fusion of Landsat and MODIS data was achieved using the STNLFFM algorithm. The results indicate that GPP exhibits higher values in the mountainous regions of west Shennongjia, compared to the eastern plain regions, with a generally increasing trend with increasing elevation. GPP has shown an overall increasing trend over the past 20 years, with almost 90% of the high-elevation regions showing an increasing trend, and the low-elevation regions showing an opposite trend. The relationship between GPP and climate factors is greatly impacted by the temporal scale, with the most pronounced correlation at a seasonal scale. The impact of temperature has been generally stable over the past 20 years across different altitudes, while the relationship with precipitation has exhibited an overall decreasing trend with the increase of altitude. Precipitation and temperature correlations show opposing variations in different months and elevations, which can be mainly attributed to the varied climatic conditions in the different elevations.

The data quality issue induced by atmospheric and other disturbances can significantly impede the application of remote sensing vegetation indices (VIs). Despite the development of numerous VI reconstruction techniques, two major challenges remain, i.e., the reliance on quality flag inputs and parameter settings. Quality flags are usually necessary as inputs for the different methods to improve the reconstruction accuracy, but mislabeling can be common in the quality flag data, which can directly introduce uncertainties. Furthermore, constant parameter schemes are usually assigned during the reconstruction applications, but the optimal parameters for all the models can show great spatial heterogeneity. Accordingly, in this paper, we propose a parameter-free and flag-free adaptive time-series method based on a variational reconstruction framework (PF-Free) to address the afore-mentioned issues, which can be applied without any parameter or flag inputs. PF-Free makes full use of the time-series temporal smoothness and inter-annual similarity to label the data after time-series rearrangement, and the parameters are adaptively selected using an improved generalized cross-validation (GCV) technique. Simulation and real-data experiments all demonstrate that PF-Free can achieve better and more stable reconstruction results, compared to other comparative methods. The adaptive quality flags can denote the data quality robustly and accurately, while guaranteeing better reconstruction performance, as long as there is any mislabeling in the original flags. Moreover, the adaptive parameter selection strategy considers the great spatial heterogeneity in the optimal parameters on a pixel-by-pixel basis, leading to more stable reconstruction outcomes under complex conditions. Further experiments also prove the effectiveness of PF-Free in processing data without quality flags or severely contaminated data, using Advanced Very High-Resolution Radiometer (AVHRR) data and Moderate Resolution Imaging Spectroradiometer (MODIS) daily normalized difference vegetation index (NDVI) data. This work provides a practical reconstruction method for VI time series, which is both flexible and convenient, without requiring any parameter or flag input, which we believe will advance VI reconstruction applications.

Reconstruction of normalized difference vegetation index (NDVI) time series plays an imperative part in the inference of vegetation dynamics. However, it is challenging for the existing methods to achieve a good balance between accuracy and efficiency. In this article, we novelly combine deep learning with a high-precision spatiotemporal adaptive tensor completion (ST-Tensor) method and propose an end-to-end NDVI time-series reconstruction network (NIT-Net). The ST-Tensor method is first used to generate high-quality seamless NDVI data as the label data to construct sample pairs, along with the original degraded observations. A handcrafted time-series processing network is further employed for effective and rapid reconstruction of the NDVI time series. Considering the temporal continuity and spatial correlation of NDVI time-series data, we combine long short-term memory with a convolution (LSTM-Conv) structure and utilize residual learning and dense connection strategies to mine the spatiotemporal features in depth. Multidimensional gradient constraints are introduced in the loss function to retain critical information. The experiments conducted on moderate resolution imaging spectroradiometer NDVI data show that the NIT-Net framework is superior to most of the comparison methods. The mean correlation coefficient between the reconstruction results of NIT-Net and ST-Tensor can reach 0.9955, while NIT-Net achieves a more than 14 times speed-up on a CPU, compared with ST-Tensor, and a 115 times speed-up on a GPU, which fully demonstrates its efficient performance and great practical application value.