The LEMONTREE science-themed meeting held on July 13th, 2023, delved into the critical topic of soil moisture stress. We heard from Giulia Mengoli, Jisu Han, Olya Skulovich and Wenli Zhao and the session was chaired by Pierre Gentine. This blog highlights the key findings and advancements discussed during the meeting, shedding light on the intricate relationship between soil moisture stress and plant functionality. From understanding carbon assimilation to exploring biophysical and biochemical responses, and unveiling long-term consistent soil moisture datasets, these discussions contribute to the broader understanding of ecosystem resilience and the role of soil moisture stress. So, let’s dive into the insights shared at the LEMONTREE meeting.

A global function of climate aridity accounts for soil moisture stress on carbon assimilation.

By Giulia Mengoli (Imperial College London PhD student, supervised by Iain Colin Prentice)

Giulia’s talk centred around the crucial relationship between carbon uptake and water availability in plants. The current P model struggles to accurately simulate plants’ carbon assimilation in dry areas, resulting in an overestimation of plant productivity (GPP) due to the lack of soil moisture considerations. To address this limitation, Giulia has developed a more comprehensive approach.

Her research is focused on incorporating two key factors into the model: the assumption of a variable maximum GPP under well-watered conditions and a variable critical threshold of water availability. By analysing the GPP ratio predicted from the model compared with observed GPP in well-watered conditions across various sites and time steps, she has found that the response of GPP to water stress could be represented with segmented curves. These curves indicate a breaking point, which represents the maximum level of GPP before a linear decline towards the wilting point.

The analysis included a collection of maximum and threshold values for all 67 sites. The aridity index, calculated as the ratio of annual potential evapotranspiration to annual precipitation, played a crucial role in determining how the maximum level and the threshold of GPP vary between sites. Systematic variations according to different aridity values were observed in both the maximum level and the threshold of GPP, with decreasing values from humid to less humid sites. By performing a nonlinear regression analysis, it is possible to obtain quantitative functions that reduce GPP based on the climatological aridity index, representing the degree of stress. Implementing this new approach improved the P model’s performance across all aridity classes, providing a more accurate representation of plant response to water stress.

You can read more about this research in our blog:

Or read the pre-print here: Mengoli, G., Harrison, S. P., and Prentice, I. C. 2023: A global function of climatic aridity accounts for soil moisture stress on carbon assimilation, EGUsphere [preprint],


Biophysical and Biochemical responses of vegetation to soil water stress: an ecosystem-scale analysis across the globe.

Jisu Han and Weiwei Zhan (PhD students at Columbia University supervised by Pierre Gentine)

Jisu’s talk focused on understanding the regulation of vegetation conductance to water at the ecosystem scale, specifically examining the biophysical and biochemical responses to water stress. The aim was to extract downregulation behaviours and explore the implications for land surface models.

There are two major pathways of GPP downregulation in response to water stress: reducing conductance to save water (i.e., stomatal or biophysical regulation) and directly downregulating the maximum photosynthesis capacity (i.e., non-stomatal or biochemical regulation). In order to disentangle these two responses from the observation data, they presented a framework that allowed extraction to ecosystem-scale stomatal and non-stomatal responses using simple formulas. By considering the ecosystem as a single leaf, they upscaled the leaf-scale equations for stomatal conductance and photosynthesis.

The process involved partitioning vegetation conductance from the ecosystem conductance based on soil moisture changes. Using the eddy covariance tower data, they computed the ecosystem conductance by the Penman-Monteith equation and grouped the data by soil moisture levels. By fitting the vegetation conductance using semi-empirical formulas and rescaling the parameters, they extract the slope parameter (G1) for the stomatal response parameter in half-hourly resolution. With this, they further calculated Vcmax25 for the non-stomatal responses parameter by directly inferring intercellular CO2 concentration from the partitioned vegetation conductance.

Figure 1 Global map of site-level downregulation strategies to soil dryness, collected from various EC tower networks across the globe.

Using a framework to estimate ecosystem-scale Vcmax25 using flux tower measurements and soil water content dynamics, they found that biochemical downregulation is universal, while one-third of sites do not exhibit biophysical downregulation. Jisu and Weiwei’s research emphasis the need for accurate representation of downregulation behaviours in land surface models to incorporate the impact of water stress on vegetation.



CASM: a long-term Consistent Artificial-intelligence based Soil Moisture dataset based on machine learning and remote sensing

Olya Skulovich (PhD student at Columbia University supervised by Pierre Gentine).

Olya’s talk introduced the CASM dataset, a significant contribution to the understanding of soil moisture dynamics. The goal of CASM was to create a global, long-term, consistent dataset by training neural networks to extrapolate back-in-time soil moisture (SM) matching Soil Moisture Active Passive (SMAP) SM based on brightness temperature (TB) from different satellites.

For the methodology of the CASM dataset, the soil moisture signal was separated into a seasonal cycle and residuals. The seasonal cycle was fixed to be consistent from year to year per location, while the residuals encompassed other periodic signals, trends, extreme events, and noise. Two neural networks were trained in the approach: the first neural network used SMOS TBs, and the second neural network used AMSR TBs as inputs. Transfer learning was applied to fine-tune the second neural network to adjust for the inconsistencies between AMSR-E and AMSR-2 data. In all cases, the brightness temperature residuals were used as predictors targeting SMAP soil moisture residuals. In addition, an ensemble of neural networks was used to assess the epistemic uncertainty of the resulting SM while introducing noise to the inputs allowed to assess aleatoric uncertainty.

The results demonstrated the success of the CASM approach in creating a temporary and spatially consistent SM dataset. In addition, since the training procedure explicitly aims to account for extreme events and trends decoupled from seasonality, the dataset is suitable for use in studies addressing interannual SM dynamics. The metrics showed a good fit for the residual part of the signal, and when the seasonal cycle was added back, the R2 value increased to 0.95. Comparison with in-situ soil moisture measurements showed similarity to SMAP soil moisture, indicating consistent and high-quality results.

Figure 2. Temporally averaged structural (epistemic) and data (aleatoric) uncertainty averaged globally.

Figure 3. Comparison of CASM soil moisture with in-situ soil moisture measurements. The total median correlation is 0.66 (mean 0.63) when compared with data from 367 stations.

In conclusion, the CASM dataset provided a valuable resource, covering a wide temporal range from 2002 to 2020, with a three-day temporal resolution and a 25x25km grid. This dataset enables researchers to delve into soil moisture dynamics and its implications for ecological processes. The CASM dataset ( ) and the accompanying research paper is available here:

Skulovich, O., Gentine, P. A Long-term Consistent Artificial Intelligence and Remote Sensing-based Soil Moisture Dataset. Sci Data 10, 154 (2023).


An objective estimate of water stress (going beyond PDSI).

Wenli Zhao (Max Planck Institute for biogeochemistry)

Wenli’s talk revolved around a study focused on using the evaporative fraction (EF) as an objective indicator of water stress. Compared to the empirical drought index, e.g., PDSI, EF, could reflect the response of vegetation to water stress conditions more directly. In addition, based on the water bucket model, in the wet season, the water bucket was filled up, while in the dry season, the vegetation consumes stored water. The depth of the vegetation rooting influences the slow or fast EF decay signal during the dry-down periods, with deeper rooting depths showing slower decay compared to shallow depths. The EF dynamics during dry-down periods could further indicate the plant water use strategies under water stress conditions.

Thus, machine learning techniques, specifically long short-term memory (LSTM) models were employed to predict EF dynamics. Driven by the meteorological observations data from FLUXNET2015 Tier 1 eddy-covariance dataset, the model could predict the EF dynamics very well with some routinely available variables, e.g., air temperature, precipitation, incoming shortwave radiation and wind speed, along with static variables including plant functional types and long-term mean temperature, radiation, and precipitation. Importance order are consistent for different memory length, showing Precipitation, incoming shortwave radiation and VPD most affect the EF predictions. Fast and slow EF decay could be further used to indicate the different rooting depth/plant water use strategies during the water stress period.

Figure 4. Importance order are consistent for different memory length, showing Precipitation, incoming shortwave radiation and VPD most affect the EF predictions.

The model exhibited relative insensitivity to variations in plant functional types, indicating its potential for global application. Comparisons with other soil moisture products showed consistent signals of water stress during the dry season. The results have important implications for future water stress estimation, e.g., drought detection, in order to obtain a more direct and accurate estimate of water stress.