Executive summary 2 sites have been selected for the study: the lake Aoatra area, considered as CA successful and Vakinakratra/Highlands considered as CA failure. Modeling at plot and cropping systems level The comparison of soil loss results obtained by direct and indirect estimations showed that generally soil loss calculated from direct measures were lower than those obtained by RUSLE. Therefore, to improve and calibrate RUSLE model at Lake Alaotra, a correction factor was proposed, mainly the reduction of the P factor by making the amount of soil loss by simulation closer to the direct measurement on field. So, extrapolation or modeling of soil loss depending on the soil management mode appeared valid for rice and maize crops, by using for CA systems different types of mulch the most used in the region at Lake Alaotra such as mulch of rice, maize+dolichos and stylosanthes. The DSSAT experiment was realized on two experimental fields that had been installed by the URP SCRiD at Andranomanelatra. From these studies, it was impossible to use the cropping model to assess the weight of water factor in the variability of inter-annual yield in these regions. It was demonstrated that in the Vakinankaratra, the part of yield variability due to stress water in conditions of study was not important, but at Lake Alaotra, it is responsible for a significant inter-annual variability of yields. The study also highlighted the big differences between achievable yields to the yields observed in field, even under controlled conditions, demonstrating the impact of other limiting factors for both regions. These studies allowed us in isolating the weight of climate and water factor on the variability of rainfed rice yields. Thus, it appears that plant health problems including weeds are similarly responsible of some differences between observed and achievable yields and the difference between treatments in the test, but there is probably also an important problem of nutrition mineral crops, contributing strongly to the differences between observed and achievable yields, and to the differences between treatments. The models tested can be used for the prediction of the potential and achievable yields with the available water in the soil conditions and the climates in which they were calibrated. Some potential advantages of CA for cropping rainfed rice (improvement of water balance, and control on water erosion) have been efficiently tested with the models f But for taking into account other potential effects of CA (weeds control, P availability, improvement of Soil organic matter stocks…) the crop models need still to be improved.