In Réunion, sugarcane represents more than half of the useful agricultural area and is the main export product of the Réunion economy. The humid tropical climatic conditions promote the development of a high diversity of weeds in sugarcane fields (from 6 to 48 species at the field level but more than 200 species for the whole sugarcane area), which can cause yield losses from 30 to 80% if weed control is not carried out in time. For farmers, optimizing the efficiency of weeding requires a good knowledge of the weed flora that will develop on their fields. However, knowing this weed flora in advance is particularly difficult. The presence and abundance of species depend strongly on environmental factors, geographic location, seedbank and season. In our study, we formalize the weed-prediction problem as a multi-label supervised learning problem. We propose a comparative study of different multi-label learning algorithms to predict presence/absence and abundance of weeds in a sugarcane field. The experimental dataset used for this study is based on 20 years of historical data from phyto-ecological surveys conducted on sugarcane fields in Réunion. The results show that the ML-ARAM and ML-kNN algorithms obtain the best performances. However, the correspondence between the predictions and the ecological profiles and the confrontation of the results to experts' opinions is not always excellent. New perspectives are proposed to improve the quality of the predictions.