The study of links between bird species richness and forest fragmentation contributes to a better understanding of landscape biodiversity. Difficulties arise from the necessity to deal with multiple non-linear relationships between the involved variables. Neural network models provide an interesting solution thanks to their internal set of non-linear neuron-like components. Their ability is well established for prediction, but their complex structure limits the understanding of underlying processes. To open the ‘black box’ and get a more transparent ‘glass box’ model, we selected a simple neural network (2 inputs, 1 hidden layer with 3 neurons and 1 output neuron), that improves the prediction of birds species richness (lower root mean square error) compared to linear, log-linear and logistic models, and simple enough to analyze its internal components and identify patterns in the data. The first hidden neuron provided a sigmoid relationship related to the forest area, the second was like a Boolean operator separating two groups according to the distance to the nearest source forest larger than 100 ha, and the third acted on the smallest isolated woodlots. We revealed a group of isolated woodlots with a higher species richness than less isolated woodlots for a given forest area. This result, unexpected according to the literature, was not obvious in the raw data, and could be explained by a regional differentiation in fragmentation history. Our neural network showed its ability to improve prediction accuracy in respect to other models, to remain ecologically understandable and to give new insights into data exploration.