Replicating Human Perception of Proximity

Applying Neural Networks in the Monitoring of Suspicious Air Activities in Amazon

Abstract

This study investigates whether a neural network can approximate human perception of proximity in the context of air traffic surveillance involving suspicious aircraft near regular and irregular airstrips. Using a DenseNet model trained on a synthetic dataset of graphical representations, the study evaluates the network's ability to classify visual proximity relationships without explicit distance computation. While the model achieved moderate performance (F1-Score of 78.4%), results were limited by overfitting and the low variability of the data. Larger datasets did not improve performance, suggesting the importance of visual diversity over quantity. These findings validate the feasibility of modeling human-like spatial reasoning through neural networks in controlled environments. The research establishes an experimental baseline for future studies involving more complex data and architectures, such as EfficientNet or Transformers, to further improve model generalization and practical applicability.