In the last decade, image analysis using machine learning algorithms proved its potential for the detection and counting of plant organs. Numerous studies provided fruit tree yield estimates based on machine learning with high levels of efficiency. However, most of these studies were conducted under homogeneous conditions of fruit aspect. The aim of this study was to develop an efficient machine learning method for ripe mango fruit detection from RGB images and to test it under heterogeneous field conditions for tree yield estimation in Senegal. The algorithm consisted in a k-nearest neighbours classification based on colour and texture features followed by a post-treatment based on shape indices. The F1 score, which accounts for both precision and recall performances, reached 0.73 for a set of 42 images of 'Kent' trees in homogeneous conditions. When performed on a second set of 128 images representing the actual heterogeneity in tree structure (height, canopy volume) and cultivars ('Kent', 'Keitt' and 'Boucodiekhal') found in the Niayes region of Senegal, the F1 score fell to 0.48. This decrease resulted from the heterogeneous field conditions in terms of fruit size, fruit colour and light exposure caused by different tree structures among cultivars. Despite the algorithm was less efficient under these conditions, significant linear relationships were evidenced between the number of detected fruits and the actual number of fruits per tree for each cultivar ('Kent': R2=0.92, 'Keitt': R2=0.93, and 'Boucodiekhal': R2=0.90). These models were cross-validated and reached a relative RMSE of 14%. Those results offer new opportunities to accurately and rapidly estimate mango yield and to further identify the parameters that drive its variability at the tree and orchard scales.