Burundi's economy is predominantly based on agriculture which is not only a source of food for most households but also a source of revenue both in food sales and in the creation of labour opportunities. The agricultural sector of Burundi employs 90 percent of the workforce through small-scale, mainly subsistence-orinted famly farming units. The sector contributes about 95 percent to the food supply and represents more than 50 percent of GDP. It is the first earner of foreign exchanges accounting for 80 percent of the country's exports, mainly coffee and tea. The country is recognized as potentially self -sufficient in food production thanks to its bimodal rainfall pattern allowing three cropping seasons.

However, the performance of this sector is very poor and 70 percent of the population lives under the national poverty line. The high prevalence of poverty and hunger are mainly attributed to the rapid population growth which leads to land scarcity and degradation. The increase in agricutural production remains generally lower compared to the rate of population growth. The main objective of this study was therefore to find out what really matters among small scale farms and suggest a model for agricultural production in response to the problem.

As an attempt to answer the obove problem'a research survey was carried out in Ngozi, one of the northern province of Burundi with a heavy population density (462 inhabitants persquare kilometre). Surveys were held in August 2010. The sampling frame was based on a list of 360 households selected at random on 90 "collines" or hills (institutional demarcation) of Ngozi in 2007 by the project UA-FACAGRO) (Université d'Anvers-Faculté d'Agronomie de l'Université du Burundi) A sample of 60 farms was purposely selected using a cluster analysis based on resource endowments and livelihood strategies. This stepwise approach was used in ordre to avoid sampling errors and hence to ensure a more reliable and representative sample.

Results of statistical analysis showed that smallholdes farmers are disadvanteged in both asset endowments and labour remunerations. The limited access to land leads to livelihood diversification through both on-farm and non-farm activites. Poor farmes have to diversity in order to get the amount of money necessary to buy food and to cover additional famly expenses. However, the limited non-farm job opportunities constrain them to engage in gricultural labour where the levels of payment are very are very low. The low wages provided to agriculture labourers are mainly attributed to the low agricultural income which results from risk attitude of farmers and market failure. Despite the inverse relationship between farm size and labour availability in the region, the level of complementarities between large and small farms is very poor. The rate of own consumption among farmers (both rich and poor) is very high which adversely affects productivity and hence rural wages.

Nevertheless, despite the rampant land scarcity, the model developed in this Master's Dissertation shows that here are possibilities for households of Ngozi to optimally produce crops that meet their food securty requirements while generating a certain level of income. By growing a limited number of crops, farmers can benefit from an optimum land use which further contributes in improving farm output and inceasing agricultural wages offered to landless households. However, the model highlighted in infrasrtucture especially strorage system a necessary condition for the success of the agricultural specialization in fighting rural poverty.

Finally, despite the great effort made in producing this Master's Dissertation, no claim cas be made to have fully covered the issue of agricultural production in Burundi,. Amongst other limitations encountered, the complexity of agricultural system of Burundi and the use of Linear programming model which relies on average farms drive up the author to formulate suggestions towards more research. Further researches could assess the scope of exchanges between farms both in input and outputs using Mathematical programming based mult-Agents systems in order to capture the real word heterogeneity and simulate sustainable resource use.