A COMPARATIVE ANALYSIS OF THE SEMI-AUTOMATIC AND AUTOMATIC APPROACHES TO THE IDENTIFICATION OF CROPS AND AGRICULTURAL AREAS IN THE DEPARTMENT OF LOUETSI-WANO (LEBAMBA, GABON)

Mapping of food-crop farming systems increasingly relies on artificial intelligence algorithms. In our previous work [13-14], we used a series of automatic and semi - automatic algorithms to map agricultural plantations and to discriminate between cassava and plantain crops. This paper presents a systematic comparative analysis of these two approaches. Starting from the methodology, we examine the statistical robustness of the models (automatic and semi-automatic) and their sensitivity to sample size. The results show that tree-based ensemble methods (RF, XGBoost) consistently outperform neural architectures and the parametric classifier when the training sample is small, raising the Kappa index for cassava and plantain discrimination from 0.63 to 0.89 within the same geographic context and the same three-level hierarchical protocol. The analysis concludes that, for crops with small training samples, the choice of algorithm family matters more than the data source alone, and it suggests ways to strengthen the methodology of future agricultural mapping studies in tropical forest areas.