Abstract:Reliable estimates of Gross Primary Productivity (GPP), crucial for evaluating climate change initiatives, are currently only available from sparsely distributed eddy covariance tower sites. This limitation hampers access to reliable GPP quantification at regional to global scales. Prior machine learning studies on upscaling \textit{in situ} GPP to global wall-to-wall maps at sub-daily time steps faced limitations such as lack of input features at higher temporal resolutions and significant missing values. This research explored a novel upscaling solution using Temporal Fusion Transformer (TFT) without relying on past GPP time series. Model development was supplemented by Random Forest Regressor (RFR) and XGBoost, followed by the hybrid model of TFT and tree algorithms. The best preforming model yielded to model performance of 0.704 NSE and 3.54 RMSE. Another contribution of the study was the breakdown analysis of encoder feature importance based on time and flux tower sites. Such analysis enhanced the interpretability of the multi-head attention layer as well as the visual understanding of temporal dynamics of influential features.

What problem does this paper attempt to address?

### Problems the paper attempts to solve This paper aims to address the challenges of estimating global hourly GPP (Gross Primary Productivity). Specifically, the research mainly focuses on the following aspects: 1. **Sparse geographical distribution**: Currently, reliable GPP measurement data only come from sparsely - distributed eddy covariance tower sites (flux tower sites), which limits the reliable quantification of GPP on regional to global scales. 2. **Insufficient time resolution**: Previous studies faced the problems of lacking high - time - resolution input features and significant missing values when scaling up in - situ GPP data to global wall - to - wall maps. In particular, GPP estimation on sub - daily or hourly time scales is crucial for understanding important ecosystem - climate interactions. 3. **Dependence on historical GPP**: Existing models usually rely on past GPP time series, which are unavailable in many locations, especially in places without eddy covariance towers. Therefore, a model that does not depend on past GPP data needs to be developed. 4. **Improving model interpretability**: By analyzing the importance of encoder features, enhance the interpretability of the multi - head attention layer and improve the visual understanding of the temporal dynamics of influencing features. ### Research methods and contributions To solve the above problems, this study introduced a new Temporal Fusion Transformer (TFT) model for upscaling global hourly GPP. Specific contributions include: 1. **Application of the TFT model**: Use the TFT model to process time - series data of different time periods, predict previously unseen entities, and handle heterogeneous inputs (such as time - varying features and static metadata). The TFT combines LSTM and self - attention mechanisms, can learn short - term and long - term historical patterns, and provides strong interpretability. 2. **Model performance evaluation**: Verified the effectiveness of the TFT model by comparing it with other machine - learning models (such as Random Forest Regressor (RFR) and XGBoost). The best model achieved an NSE (Nash - Sutcliffe Efficiency) of 0.704 and an RMSE (Root Mean Square Error) of 3.54. 3. **Feature importance analysis**: By analyzing the importance of encoder features, enhanced the interpretability of the multi - head attention layer and improved the understanding of the temporal dynamics of influencing features. 4. **Analysis of eco - regional differences**: Evaluate the model performance according to different vegetation types (IGBP classification), revealing the differences in model performance in different ecological regions, providing a basis for further optimizing the model. Through these efforts, the research not only improved the accuracy of global hourly GPP estimation but also provided valuable insights for understanding and improving carbon cycle models.

Upscaling Global Hourly GPP with Temporal Fusion Transformer (TFT)

Deep Hierarchical Temporal Data Fusion Improves Yield Estimation under Extreme Climate Stress

Explainable forecasting of global horizontal irradiance over multiple time steps using temporal fusion transformer

Resolution-Agnostic Transformer-based Climate Downscaling

A Downscaling Framework for Urban Nighttime Light Based on Multi-Factor Geographically Neural Network Weighted Regression

Uniform upscaling techniques for eddy covariance FLUXes (UFLUX)

Efficient Subseasonal Weather Forecast using Teleconnection-informed Transformers

ENSO Forecasts with Spatiotemporal Fusion Transformer Network

UFLUX-GPP: A Cost-Effective Framework for Quantifying Daily Terrestrial Ecosystem Carbon Uptake Using Satellite Data

Extreme Precipitation Seasonal Forecast Using a Transformer Neural Network

A Novel Deep Learning Approach to the Statistical Downscaling of Temperatures for Monitoring Climate Change

Short-Term Electricity Load Forecasting Using the Temporal Fusion Transformer: Effect of Grid Hierarchies and Data Sources

Comparison of Machine Learning Methods to Up-Scale Gross Primary Production

Successful Precipitation Downscaling Through an Innovative Transformer-Based Model

Quantitative assessment of the scale conversion from instantaneous to daily GPP under various sky conditions based on MODIS local overpassing time

The Improved U-STFM: A Deep Learning-Based Nonlinear Spatial-Temporal Fusion Model for Land Surface Temperature Downscaling

Fast prediction of spatial temperature distributions in urban areas with WRF and Temporal Fusion Transformers

Fusion of multiple models for improving gross primary production estimation with eddy covariance data based on machine learning

Application of Temporal Fusion Transformer for Day-Ahead PV Power Forecasting

MTTF: a Multimodal Transformer for Temperature Forecasting

A new global time-series GPP production: DFRF-GPP