The scenarios of the frequency of landslide reactivation f generally follow the mean precipitation scenarios. Distinct trends in f were simulated in none of the GCMs. Apparently, decreased summer precipitation and increased evapotranspiration are cancelled out by increased winter and spring precipitation.

A distinct downward trend is simulated in ECHAM4. In HCGG a weaker downward trend is simulated. The stronger trend in ECHAM4 is not surprising, given the strong precipitation decreases in spring and autumn. In HCGS, the slight increases in precipitation are cancelled out by increased evapotranspiration.

The presentation of scenarios for landslide reactivation, based on low-quality analog-downscaled precipitation scenarios, seems senseless. However, it is done merely to show that the underestimation of precipitation is amplified in f (ECHAM4, 1960 to 1989: 17.5% precipitation to 81% f on average). This shows the importance of consistent, plausible precipitation scenarios in order to obtain consistent impact scenarios.

A distinct change in the frequency of landslide reactivation was simulated with only 1 specific combination of GCM and downscaling approach (multiple regression, ECHAM4). This indicates that the consideration of not only different GCMs but also different downscaling methods is justified and recommended in order to better quantify the overall uncertainty in climate change impact studies.

The influence of interdecadal variability of precipitation is considerable [in the scenario of landslide reactivation for multiple regression and ECHAM4 for all the time intervals]. This influence seems greater than the influence of increasing temperature, which would have resulted in a smoother decline. Temperature is important, but becomes decisive when precipitation changes are slight or become cancelled out between the seasons.The major anomalous situations occur in the target periods 2030 to 2059 and 2040 to 2069 have also been studied. These were not considered in the other 2 downscaling approaches. This illustrates the importance of considering the entire scenario period instead of only 2 or 3 selected target periods in order to capture the most radical impact changes.

Finally, the scenarios of different f quantiles [in the anomalous situations] are further apart than the scenarios [of landslide reactivation for multiple regression and ECHAM4 for all the time intervals]. This shows that climate variability not captured by the multiple regression, or in general by the applied methods, adds a considerable amount of uncertainty to the simulated impacts. This variation is not revealed when using the other downscaling methods, but should always be considered as an important part of local climate.

Precipitation and temperature

The Boisivre landslide is situated on the eastern slope of the Riou Bourdoux valley, in the basin of Barcelonnette in the French Alps. The nearest weather station, Barcelonnette ‘Le Verger’ is about 4 km to the southeast.

The threshold is 270 mm. This means that whenever the amount of net precipitation over 3 mo exceeds 270 mm, a landslide reactivation should occur, according to the model. The threshold only allows events and ‘nonevents’ to be discriminated; the model is too coarse for event magnitudes to be assessed. The match is not perfect, because in the dry years 1974-1975 a reactivation occurred. Furthermore, the duration of landslide reactivation is systematically overestimated by the model.

Tree ring data from 1956 to 1980 were used to reconstruct the landslide reactivation record. The associated temporal resolution is yearly. A time series of net precipitation covering the same period was calculated with a simple water balance model for the till. The model requires monthly precipitation and temperature data, and soil moisture retention curves of the till. To meet the desired climatic time scale outlined earlier, the net precipitation time series was aggregated to 3 mo running sums. For details of the model see Buma (1998).

The frequency of landslide reactivation (f) was subsequently calculated. The annual maxima of the climatic time series (during the period 1960 to 1989) were identified and ranked in ascending order (30 values). The return interval of each annual maximum is related to this ranking according to the theory of statistics of extreme values formulated by Gumbel (1958). A linear regression relating these 2 variables provided a significant fit with an r2 of about 0.95. Substituting the threshold value (270 mm) in the regression provided a f of 0.27 yr–1 (landslide reactivation about once every 4 yr).

Several factors complicate the relation. First, the tree ring record may be contaminated by false rings or the absence of rings (as a result of abnormal weather conditions during the growing season) or counting errors. Second, tree slanting may have other causes such as wind action or disease. Third, the model does not take into account the time it takes the water to percolate through the weathered marl. This time lag may also be variable in time; more water may be needed to raise the groundwater table, as a result of drainage in a preceding period of relative drought. Finally, landslide reactivation may be more likely to occur after long periods of slope stability in which the build-up of shear stress is allowed. Conversely, reactivation may cease as a result of feedback mechanisms once the shear stress is released. This may explain why the duration of reactivation is in general overestimated.

Utilisation of the GCMs ECHAM4-OPYC3 , HadCM2-GG, HadCM2-GS. The performance of the 2 statistical downscaling methods in reproducing the observed Barcelonnette precipitation was tested by downscaling from the observed large-scale climate. Temperature was not downscaled with these methods; the temperature scescenarios presented in Section 5 were derived by direct GCM interpolation (average of several grid boxes). This was motivated by the fact that in general the correlation between local and large-scale values is better for temperature than for precipitation.

Observed temperature (1956 to 1994) and precipitation time series (1928 to 1994) of the weather station Barcelonnette ‘Le Verger’ were obtained from Météofrance. A monthly time scale for downscaling is considered sufficient to capture the time scale of the climatic landslide triggering mechanism, as outlined in Section 2.

The direct precipitation and temperature scenarios were used as input for the landslide model without further processing. Five target periods were selected: 1870–1899, 1910–1939, 2020–2049, 2050–2079 and 2070–2099. Again, the simulation of the ‘observed’ frequency of landslide reactivation f in the reference period is trivial.

(4) - Remarques générales

It may be concluded that the presented model for landslide reactivation as a function of climatic parameters still carries too many uncertainties for a successful application to climate change impact assessment. However, the focus of this paper is to determine the influences of different GCMs and downscaling methods on simulated climate change impacts on landsliding. For this purpose even a theoretical model relating climate to landslide activity would have been suitable. Therefore, the landslide model is considered good enough for the scenario study.

Uncertainty in scenarios of climate change impacts is an important issue in this paper. It pertains to all stages of the climate change impact modelling approach: from errors and biases in the observations on the landslide and generalisations in the landslide model to numerous problems associated with the construction of climate scenarios. This study has shown that the use of different GCMs and downscaling methods results in a broad range of impact scenarios. There seems little point in calibrating impact models meticulously in order to improve the impact scenario with such great uncertainty remaining in the climate scenarios (obviously, this does not mean that there is no point at all in calibrating these models). Our recommendation for climate change impact model studies does not concern the use of one specific downscaling method or GCM in combination with a fully validated impact model; instead we suggest that as many uncertainties as possible be taken into account by considering different parameter settings in impact models and climate scenarios. In the meantime, while impact models are improved, GCMs and downscaling methods are improved as well.