Abstract

Satellite-derived agricultural drought indices can provide a complementary perspective of terrestrial vegetation trends. In addition, their integration for drought assessments under future climates is beneficial for providing more comprehensive assessments. However, satellite-derived drought indices are only available for the Earth observation era. In this study, we aim to improve the agricultural drought assessments under future climate change by applying deep learning (DL) to predict satellite-derived vegetation indices from a regional climate simulation. The simulation is produced by the Terrestrial Systems Modelling Platform (TSMP) and performed in a free evolution mode over Europe. TSMP simulations incorporate variables from underground to the top of the atmosphere (Ground to Atmosphere G2A) and are widely used for research studies related to water cycle and climate change. We leverage these simulations for long-term forecasting and DL to map the forecast variables into Normalized Difference Vegetation Index (NDVI) and Brightness Temperature (BT) images that are not part of the simulation model. These predicted images are then used to derive different vegetation and agricultural drought indices, namely NDVI anomaly (NDVIa), BT anomaly (BTa), Vegetation Condition Index (VCI), Thermal Condition Index (TCI), and Vegetation Health Index (VHI). The developed DL model could be integrated with data assimilation and used for down-stream tasks, i.e., for estimating NDVI and BT for periods where no satellite data are available and for modelling the impact of extreme events on vegetation responses with different climate change scenarios. Moreover, our study could be used as a complimentary evaluation framework for TSMP-based climate change simulations. To ensure reliability and to assess the model’s applicability to different seasons and regions, we provide an analysis of model biases and uncertainties across different regions over the Pan-Europe domain. We further provide an analysis about the contribution of the input variables from the TSMP model components to ensure a better understanding of the model prediction. A comprehensive evaluation on the long-term TSMP simulation using reference remote sensing data showed sufficiently good agreements between the model predictions and observations. While model performance varies on the test set between different climate regions, it achieves a mean absolute error (MAE) of 0.027 and 1.90 K with coefficient of determination (R2) scores of 0.88 and 0.92 for NDVI and BT, respectively, at 0.11° resolution for sub-seasonal predictions. In summary, we demonstrate the feasibility of using DL on a TSMP simulation to synthesize NDVI and BT satellite images, which can be used for agricultural drought forecasting.

Model Architecture

model design

The model design follows a U-Net architecture with focal modulation blocks. The input for deep learning is a data cube representing a specific week of TSMP data and the output are NDVI and BT corresponding to the same week. The predicted NDVI and BT are used to derive different agricultural drought indices.

focal modulation block

An illustration of the function Focal Modulation Block. It consists of 3 main parts: focal contextualization, gated aggregation, and interaction. First, the query, value and gates are obtained by projecting Xk with linear layers. Then, a stack of depth-wise 2D convolutions followed by a global pooling is used on the value to derive contextual features around pixels. Gates Gk are used to adaptive aggregate features into a modulator. Finally, the interaction between queried pixels are modulator is performed and projected by a final linear layer to compute the output. The shown is examples gates with the magnitude of modulator pixel-wise for the first block encoder.

Results

NDVI anomaly prediction

Prediction of weekly spatially averaged NDVI anomaly over the PRUDENCE regions.

NDVI anomaly prediction

Prediction of weekly spatially averaged NDVI anomaly over the PRUDENCE regions.

BT anomaly prediction

Prediction of weekly spatially averaged BT anomaly over the PRUDENCE regions.

BT anomaly prediction

Prediction of weekly spatially averaged BT anomaly over the PRUDENCE regions.

VCI prediction

Prediction of weekly spatially averaged VCI over the PRUDENCE regions.

VCI prediction

Prediction of weekly spatially averaged VCI over the PRUDENCE regions.

TCI prediction

Prediction of weekly spatially averaged TCI over the PRUDENCE regions.

TCI prediction

Prediction of weekly spatially averaged TCI over the PRUDENCE regions.

VHI prediction

Prediction of weekly spatially averaged VHI over the PRUDENCE regions.

VHI prediction

Prediction of weekly spatially averaged VHI over the PRUDENCE regions.

uncertainty analysis 1

An analysis of uncertainty and model generalization for different times of the year. (a) NDVI mean bias. (b) NDVI mean Pearson Correlation. (c) BT mean bias. (d) BT mean Pearson Correlation.

uncertainty analysis 2

An analysis of uncertainty and model generalization for different time of the year over each PRUDENCE region. (a) NDVI bias distribution. (b) BT bias distribution. Shown are the probability density functions. (c) NDVI mean absolute error. (d) BT mean absolute error.

Agricultural Drought Assessment

VHI seasonal prediction

Comparison between the seasonal predicted Vegetation Health Index (VHI) and NOAA observations over Pan-Europe domain at 0.11° resolution.

Comparison between the predicted drought frequency and NOAA observations over Pan-Europe domain at 0.11° resolution. Frequency represents the percent of weeks with Severe-to-Exceptional drought events (VHI < 26).

Global impacts of input variables

channel attention

Channel attentions for TSMP input variables. The activations are shown for both NDVI (top) and BT (bottom) with respect to all weeks in validation and test sets. Each set of attentions is grouped according to one model component in TSMP.

Long-term VHI over Europe

Poster

BibTeX

	      
	@Article{gmd-17-2987-2024,
	AUTHOR = {Shams Eddin, M. H. and Gall, J.},
	TITLE = {Focal-TSMP: deep learning for vegetation health prediction and agricultural drought assessment from a regional climate simulation},
	JOURNAL = {Geoscientific Model Development},
	VOLUME = {17},
	YEAR = {2024},
	NUMBER = {7},
	PAGES = {2987--3023},
	URL = {https://gmd.copernicus.org/articles/17/2987/2024/},
	DOI = {10.5194/gmd-17-2987-2024}
	}
-->