tanitdata
FREN

Policy brief

Southern Tunisia's Date Oases: Monitor the Aquifer, Not the Yields

Twenty-two years of satellite data: the aquifer is draining, harvests keep growing, but climate is starting to show through in good and bad years again.

By Tarek Gasmi Sajeda Albarghati

6 min read
  • agriculture
  • water
  • climate
  • sustainability
  • governance
  • policymakers
  • tunisia
  • north-africa

Finding

Southern Tunisia's aquifer has lost about 1 cm/yr since 2007. Yields keep rising, but climate is starting to show through in Kébili and Tozeur harvests.

Recommendation

Stop asking satellites to forecast the harvest. Ask them to monitor the aquifer — and anchor the 2025 Finance Law to that measurement system before the next season.

Key findings

  • NASA's GRACE satellites measure a cumulative 16.6 cm loss of groundwater over 22 years in southern Tunisia. A formal statistical test places the tipping point at 2007: before, the aquifer was stable; after, it loses about 1 cm every year.
  • At most 5 cm of this loss is attributable to climate. The rest — at least 12 cm — is anthropogenic. In Kébili, the number of registered wells nearly tripled over the same period: 3,733 wells in 2002, 7,872 in 2012, 10,632 in 2024.
  • Yields keep rising. In Kébili, productivity per hectare went from 3.7 t/ha in 2002 to 8.4 t/ha in 2024. But over the recent period, good climate years produce above the trend and bad ones below it: irrigation no longer absorbs everything.
  • Three families of machine-learning models trained on 2002–2020 and tested on 2021–2024 fail to forecast yields from satellite data. None beats a simple per-governorate trend. Not a technical defect — the signature of an irrigated system that decouples yields from climate.
  • The climate–yield signal is clear in Tozeur and Kébili. It needs confirmation in Gafsa (sensitive to the choice of statistical method). It does not yet appear in Gabès.

Recommendations

For policymakers

Stand up a public monthly aquifer dashboard for the four southern governorates, hosted under DGRE or ONAGRI, with a refresh latency below 60 days.

Cap and audit new boreholes in Kébili. Require volumetric metering reported to the CRDA for wells above a flow-rate threshold to be defined.

Anchor the 2025 Finance Law's pricing tools on agricultural water explicitly to the dashboard above, before the 2026–2027 season.

Methodology

Per-governorate compilation of 22 years (2002–2024) of satellite series (NASA GRACE and GRACE-FO for groundwater, MODIS and ERA5-Land for surface climate) calibrated against CRDA extraction totals, joined to date production reported by ONAGRI. Analyses are run on yield per hectare after correction for effective palm-grove area. The aquifer's slope break is identified by a formal statistical test, not by visual reading.

Questions this brief answers

  • Can satellites forecast date palm yields in southern Tunisia?
  • How fast is the deep aquifer draining, and since when?
  • Does irrigation still shield harvests from climate?
  • What should be done with the water provisions of the 2025 Finance Law?

In short

Southern Tunisia’s date crop grows over a fossil aquifer that NASA’s GRACE satellites have observed since 2002. Over 22 years, 16.6 cm of water have been lost; a formal statistical test places the slope break in 2007. Over the same period, production has surged — in Kébili, yields per hectare went from 3.7 t/ha to 8.4 t/ha — but irrigation can no longer fully mask the climate: in Tozeur and Kébili, good climate years produce above trend and bad ones below it. Three families of machine-learning models fail to forecast yields from satellites. The operational conclusion is sharp: the satellite instrument is not for predicting harvests, it is for measuring how much water is left under the ground. The 2025 Finance Law introduced pricing tools on agricultural water. Without a credible measurement system, those tools float in a void.

1. The ground

Tozeur, Kébili, Gafsa, Gabès. Four governorates, one shared resource underground: deep aquifers — the Continental Intercalaire and Complexe Terminal — that rainfall can no longer refill at any human time scale. It is this water, pumped to the surface, that makes the date palm grow. For two decades, irrigation acted as a shield between climate and harvest.

2. The buffer worked

Over 2002–2010, climate moves sharply — dry years, wet years — and harvests barely show it. This is the expected behaviour of an irrigated system: as long as water is available, yields don’t track climate. It is also why satellite-based agricultural forecasting models, designed for rainfed agriculture, were never built for this context.

3. The aquifer is draining — and the human cause is clear

The central number: −16.6 cm of groundwater over 22 years. A formal statistical test places the break at 2007 (95% confidence interval 2006–2008): before, the aquifer was stable; after, it loses about 1 cm per year. A first-order water balance attributes at most 5 cm of this loss to climate. The rest — at least 12 cm — is anthropogenic.

Over the same window, the number of registered wells in Kébili went from 3,733 in 2002 to 10,632 in 2024. Production grew 4.1×, of which roughly 1.9× came from area expansion and 2.2× from per-hectare gains. This is not a coincidence: pumping tripled where the aquifer broke.

4. The buffer is wearing thin

Yields keep rising — that has to be said upfront, to avoid misreading what comes next. In Kébili, per-hectare productivity went from 3.7 t/ha in 2002 to 8.4 t/ha in 2024, more than doubling.

What has changed is something else: over the recent period, the gap between good and bad climate years starts showing in the harvest again. A dry year produces below trend; a wet year above. The effect is clear in Tozeur and Kébili. It needs confirmation in Gafsa (the result depends on the choice of statistical method). It does not yet appear in Gabès, where irrigation still absorbs climate variability.

This is what a thinning water buffer looks like: protection does not collapse at once, it lets through more and more signal.

5. Satellites do not forecast the harvest

Three families of machine-learning models — regularized linear regression, random forests, gradient boosting — trained on 19 years of satellite data (2002–2020) and tested on the next 4 (2021–2024). None beats a simple per-governorate trend. Test R² is consistently negative.

Not a modelling flaw. The very signature of a water buffer: as long as irrigation decouples yields from climate, the variables satellites can see — moisture, temperature, vapour pressure — carry no predictive power. Asking satellites to forecast harvests in an irrigated oasis is using the wrong instrument.

The useful instrument is GRACE — which measures not the harvest, but the water that’s left.

6. Three levers, calibrated on the 2025 Finance Law

Public monthly aquifer dashboard, per governorate, hosted under a Tunisian authority (DGRE or ONAGRI), with a refresh latency below 60 days and open methodological documentation. Alert thresholds calibrated on the post-2007 slope.

Cap and audit on Kébili wells. The tripling observed over 2002–2024 coincides with the GRACE break. Require volumetric metering reported to the CRDA for wells above a flow-rate threshold to be defined. Condition any new authorisation on a hydrogeological opinion grounded in the calibrated series.

Anchor the 2025 Finance Law. The pricing and fiscal tools introduced presuppose a credible measurement system. Explicitly designate the dashboard above as the reference instrument before the 2026–2027 agricultural season.

Methodology in brief

Per-governorate compilation of 22 years (2002–2024) of satellite series: NASA GRACE and GRACE-FO for groundwater, MODIS and ERA5-Land for surface climate, joined to date production reported by ONAGRI and calibrated against CRDA extraction totals. Analyses are run on yield per hectare, corrected for effective palm-grove area. The aquifer’s slope break is identified by a formal statistical test, not by visual reading. Technical details — choice of detrending method, autocorrelation correction, sensitivity to specification variants — are documented in the manuscript linked below.

Limitations

The sample is short: 23 years, 4 governorates. No individual finding survives a strict multiple-testing correction across the full grid of specifications tested; signal robustness comes from multi-source convergence (the GRACE break, well proliferation, signal persistence under several detrending methods), not from a single corrected test. The Gafsa signal collapses under one specific methodological variant and is reported tentatively throughout the brief. Finally, the GRACE calibration against CRDA totals assumes those totals are reliable; an independent audit of extraction metering is a separate operational priority.

Sources

  • Research paper (EarthArXiv preprint): Gasmi, T., Guesmi, R., Abdelbari, S., Boulaares, S., Albarghathi, S. (2026). Emerging climate–yield re-coupling in overexploited date palm oases: satellite evidence from a 22-year temporal decoupling index in southern Tunisia. DOI: 10.31223/X5648P
  • Code and data (GitHub): github.com/tanitdata/DatePalm
  • Data archive (Zenodo): DOI 10.5281/zenodo.20172850
  • Primary sources: NASA PO.DAAC JPL Mascon GRACE/GRACE-FO RL06.3 v4; MODIS MOD13A2 and MOD11A2; ERA5-Land (Copernicus C3S); CHIRPS (UCSB CHG); ONAGRI (per-governorate date production 2002–2024); CRDA Tozeur, Kébili, Gafsa, Gabès (extraction totals and well inventory).
  • AI-native access to ONAGRI agricultural data: agridata MCP — the studio’s Model Context Protocol server that exposes the agridata.tn portal directly to LLM clients (source code on GitHub).