StemNet: A Dataset, Benchmark and Method for Scene Recognition in Remote Sensing

This study addresses the challenges in remote sensing scene recognition, traditionally treated as an image classification problem, leading to issues with false positives and false negatives, especially in complex images. We propose a paradigm shift by framing scene recognition as an advanced object detection task and introduce a specialized dataset to assess models in realistic scenarios. Our approach includes StemNet, an innovative fusion technique integrating hyperspectral and RGB imagery, surpassing traditional methods in accuracy, precision, and robustness. Through extensive experimentation, StemNet consistently outperforms conventional techniques, offering a groundbreaking perspective and setting a benchmark for future methodologies in remote sensing scene recognition. The introduced dataset and StemNet contribute significantly to advancing research and practice in this field.