Chengcheng Huang

Long short‐term memory (LSTM) networks effectively capture temporal dependencies in ecosystem carbon flux prediction, but whether the temporal patterns they learn reflect genuine ecological memory or arise from architectural biases remains largely unexplored. This study develops an LSTM‐attention model for predicting gross primary production (GPP) and ecosystem respiration (Reco), the two component fluxes of net ecosystem exchange (NEE), and introduces a validation procedure for integrated gradients (IG) attributions consisting of two complementary experiments. Temporal shuffling tests whether attribution patterns depend on the sequential order of the inputs, and timestep ablation confirms functional dependence on individual input months. Applied to 43 flux tower sites covering deciduous broadleaf forests (DBF), evergreen needleleaf forests (ENF), and grasslands (GRA) across North America, the models achieved test‐set R ² of 0.75–0.90 across six configurations with high attribution stability across a 10‐member ensemble. GPP showed recency‐dominant temporal memory across all three biomes, while Reco showed biome‐dependent structure, with long‐term memory in DBF ( R ² drop of 47.68% at t = −6), recency dominance in GRA ( R ² drop of 125.26% at t = −1), and intermediate behavior in ENF. Temporal shuffling collapsed or reorganized the ecosystem‐specific patterns, and the ablation R ² drops qualitatively corresponded to the IG attribution profiles in all six configurations. The decomposition further reveals that long‐term components of NEE memory arise predominantly from Reco and short‐term components from GPP. This work provides a transferable methodology for distinguishing ecologically meaningful temporal signals from statistical artifacts in deep learning attribution analysis, with broad applicability across the geosciences.

Ecosystem carbon dynamics are governed by complex temporal dependencies and environmental interactions, yet these processes remain poorly understood and quantified across diverse biomes. Here, we developed an explainable LSTM‐Attention framework that integrates LSTM networks, attention mechanisms, and gradient‐based attribution methods to reveal temporal dependencies in ecosystem carbon flux responses by analyzing eddy covariance flux measurements from 71 sites spanning eight North American biomes. Using this approach, we identified three distinct temporal memory patterns governing carbon flux responses. Grasslands exhibit short‐term memory dominance with exponentially increasing temporal contributions from distant to recent past. Deciduous broadleaf forests and wetlands show long‐term memory dominance, with deciduous broadleaf forests displaying the strongest historical dependence (attribution values declining from 0.30 at 6 months to 0.045 at 1 month). Croplands and evergreen needleleaf forests demonstrate U‐shaped dual memory patterns. Building on these temporal patterns, we identified biome‐specific environmental drivers operating within each memory framework: wetlands primarily controlled by soil moisture, evergreen needleleaf forests by radiation, and closed shrublands by vapor pressure deficit. Beyond individual drivers, we uncovered critical nonlinear interactions that diverged from linear correlations at most sites (55 of 71 with ρ < 0.5). For instance, the carbon sink capacity of deciduous broadleaf forests depends on synchronized canopy development and photosynthetic activity, while closed shrublands show strong suppression of carbon uptake by atmospheric water deficit regardless of vegetation greenness, revealing how multiple drivers jointly regulate ecosystem functioning. Validating memory's fundamental role, ablation experiments confirmed that removing memory mechanisms degraded model prediction performance and altered environmental driver identification (Kendall's Tau < 0.5), demonstrating that temporal memory is integral to accurately modeling ecosystem carbon flux responses to environmental drivers. These findings provide mechanistic insights into temporal controls of carbon exchange across different biomes. This knowledge is critical for improving terrestrial carbon‐climate feedback representations under global change.

Estimation of regional-scale carbon budget still faces considerable uncertainties, whereby reconciling bottom-up estimates and top-down inversions are essential steps towards reducing such uncertainties. Here, we make full use of available flux tower observations and novel satellite land surface observations (e.g., soil moisture and solar-induced chlorophyll fluorescence) to constrain the spatiotemporal regimes of net ecosystem carbon exchange (NEE) for North America based on the advanced Long Short-term Memory (LSTM) networks. The LSTM-based model is capable of incorporating the memory effects of climate and environments on NEE variations, thus more accurately simulating the interannual variations (IAVs) and long-term trends of NEE. We produced gridded NEE at a spatial resolution of 0.1° × 0.1°and monthly time-step over 2001–2021 for North America.The annual total NEE during 2001–2021 is -1.74 ± 0.10 Pg C yr-1, which is much closer to the top-down inversions (-0.73 and -0.63 Pg C yr-1 for CarbonTracker2022 over the same period and the Orbiting Carbon Observatory-2 (OCO-2) v10 MIP ensemble mean over 2015–2020 respectively) than the global flux upscaling products (-3.04 and -2.75 Pg C yr-1 by FLUXCOM2020 RS+CRUJRA1.1 and RS+ERA5 respectively and -3.30 Pg C yr-1 by NIES2020). Moreover, the spatial patterns matched with the OCO-2 v10 MIP ensemble mean reasonably well, which indicated the largest carbon uptake in the Midwest Corn-Belt area during peak growing seasons and in the Southeast on an annual basis. The estimated IAV of NEE is highly correlated with that by CarbonTracker2022. The LSTM estimate captured the NEE anomalies caused by the droughts in 2011, 2012, 2017, 2020–2021, and the 2019 Midwest floods. The performances are clearly better than existing global flux upscaling products. In addition, the estimated annual NEE exhibited a significant decline at the rate of -0.008 Pg C yr-2 (p < 0.05), indicating an overall enhanced carbon sink during the recent two decades, which is in line with contemporary estimates.These results suggest that the LSTM-based NEE upscaling provides an improved estimation of North American NEE for both spatial and temporal characteristics, narrowing down the gap between bottom-up estimates and top-down inversions, which is an important step towards robust regional carbon budget estimations.

Land subsidence is a geological process that occurs at a slow and persistent pace. However, due to the limited archive of high-resolution SAR images, the single-sensor Multi-temporal Interferometric Synthetic Aperture Radar (MTInSAR) technology is unable to satisfy the demands of long-term and high-resolution deformation monitoring. To address this issue, we employed a Small Baseline Subset of InSAR technology, using ALOS-2 and Sentinel-1 data, to detect the deformation characteristics of the Beijing Subway Network (BSN). The two datasets produced consistent results regarding the evolution process and spatial distribution. Subsidence monitoring can be sustained using other datasets when high-resolution images are insufficient. To investigate the differences in deformation characteristics at different stages, we focused on the newly opened section of Beijing Subway Line 17 (L17). Our findings revealed that the southern portion of L17 posed a significant risk of instability with subsidence occurring at Jiahuihu Station, Ciqu Station, and Ciqubei Station. The deformation at these stations was 12.8 mm, 18.7 mm, and 5.6 mm, respectively. Before and after the opening of L17, their deformation rates changed. Specifically, the deformation rate at Jiahuihu Station decreased from -9.4 mm/year to -5.5 mm/year, the deformation rate at Ciqu Station decreased from -14.1 mm/year to -7.5 mm/year, and the deformation rate at Ciqubei Station decreased from -4.9 mm/year to -1.5 mm/year. These results indicate that, after the subway's opening, the deformation rates at these stations are decreasing. The deformation along the subway will intensify during construction and then gradually stabilize once it is operational.

Southwestern North America (SWNA) continuously experienced megadroughts and large wildfires in 2020 and 2021. Here, we quantified their impact on the terrestrial carbon budget using net biome production (NBP) estimates from an ensemble of atmospheric inversions assimilating in-situ CO₂ and Carbon Observatory–2 (OCO-2) satellite XCO₂ retrievals (OCO-2 v10 MIP Extension), two satellite-based gross primary production (GPP) datasets, and two fire CO₂ emission datasets. We found that the 2020–2021 drought and associated wildfires in SWNA led to a large CO₂ loss, an ensemble mean of 95.07 TgC estimated by the satellite inversions using both nadir and glint XCO₂ retrievals (LNLG) within the OCO-2 v10 MIP, greater than 80% of SWNA’s annual total carbon sink. Moreover, the carbon loss in 2020 was mainly contributed by fire emissions while in 2021 mainly contributed by drought impacts on terrestrial carbon uptake. In addition, the satellite inversions indicated the huge carbon loss was mainly contributed by fire emissions from forests and grasslands along with carbon uptake reductions due to drought impacts on grasslands and shrublands. This study provides a process understanding of how some droughts and following wildfires affect the terrestrial carbon budget on a regional scale.

Upscaling flux tower measurements based on machine learning (ML) algorithms is an essential approach for large‐scale net ecosystem CO₂ exchange (NEE) estimation, but existing ML upscaling methods face some challenges, particularly in capturing NEE interannual variations (IAVs) that may relate to lagged effects. With the capacity to characterize temporal memory effects, the Long Short‐Term Memory (LSTM) networks are expected to help solve this problem. Here we explored the potential of LSTM for predicting NEE across various ecosystems using flux tower data over 82 sites in North America. The LSTM model with differentiated plant function types (PFTs) demonstrates the capability to explain 79.19% (R² = 0.79) of the monthly variations in NEE within the testing set, with RMSE and Mean Absolute Error values of 0.89 and 0.57 g C m⁻² d⁻¹ respectively (r = 0.89, p < 0.001). Moreover, the LSTM model performed robustly in predicting cross‐site variability, with 67.19% of the sites that can be predicted by both LSTM models with and without distinguished PFTs showing improved predictive ability. Most importantly, the IAV of predicted NEE highly correlated with that in flux observations (r = 0.81, p < 0.001), clearly outperforming that by the random forest model (r = −0.21, p = 0.011). Among all nine PFTs, solar‐induced chlorophyll fluorescence, downward shortwave radiation, and leaf area index are the most important variables for explaining NEE variations, collectively accounting for approximately 54.01% in total. This study highlights the great potential of LSTM for improving carbon flux upscaling with multi‐source remote sensing data.