Aerial Surveillance Land Classification using Global Sum Pooling in CNNs: A Study on EuroSAT Dataset

Global sum pooling (GSP) is a recent pooling approach in neural networks that provides an alternative method for feature aggregation. In this work, we explore the applicability of GSP to various deep-learning tasks, paying special attention to satellite-image classification on the EuroSAT dataset. GSP sums all the feature map values while maintaining full intensities of activations, whereas conventional pooling techniques, such as max-pooling and Global Average Pooling (GAP), pay greater emphasis to extreme values and averaged summaries. This property is uniquely valuable in applications, for example, land use classification for satellite photography, where the total strength of features matters. This paper evaluates the effectiveness of GSP by conducting heavy experimentation on the EuroSAT dataset, which showed a great potential in improving model accuracy robustness, and capabilities of generalization. The authors show that GSP outperforms traditional pooling methods, especially in scenarios where the difference in features makes a huge difference. Therefore, it emphasizes the integration of GSP in Convolutional Neural Networks (CNNs) and how it performs its functions in a flexible and beneficial manner compared to other modern deep learning architectures associated with remote sensing and satellite imagery.