Remote sensing can provide useful explanatory variables for tree species distribution modeling, but only a few studies have explored this potential in Amazonia at local scales. Particularly for tropical forest management it would be useful to be able to predict the potential distribution of important tree taxa in areas where field data is as yet missing. Forest concessions produce valuable census data that cover large areas with high sampling effort and can be used as occurrence data in species distribution models (SDM). Nevertheless, these tree records are often spatially clumped and possibly only provide accurate predictions over areas close to where the training occurrence records are located. Here, we aim at investigating to what degree SDM performance and spatial predictions differ between models that have different spatial configurations of the occurrence data. For this, we divided the available occurrence data from a forest concession census in Peruvian Amazonia into different spatial configurations (narrow, elongated and compact), each of which contained approximately 20% of the full dataset. We then modelled the distributions of five tree taxa using Landsat data and elevation. More elongated configurations of the training data were more representative of the available environmental space, and also produced more robust SDMs. Average model performance (expressed as AUC) was 5% higher and variation in model performance 50% lower when elongated rather than compact configurations of training area were used. This confirms that covering only a small fraction of the environmental variability in the area of interest may lead to misleading SDM predictions, which needs to be taken into account when forest management decisions are based on SDMs.