Remote Sensing Based Crop Type Classification Via Deep Transfer Learning

Machine learning methods using aerial imagery (satellite and unmanned-aerial-vehicles-based imagery) have been extensively used for crop classification. Traditionally, per-pixel-based, object-based, patch-based classifying crops worldwide. Recently, aided by the increased availability of powerful computing architectures such as graphical processing units, deep learning-based systems become popular in other domains natural images. However, building complex neural networks from scratch is a challenging affair, owing to limited labeled data remote sensing domain multitemporal (phenology) geographic variability associated with agricultural data. In this article, we discuss these challenges detail. We then various transfer methodologies that help overcome challenges. Finally, evaluate whether strategy pretrained different helps improve image classification performance on benchmark dataset. Our findings indicate dataset cannot be off-the-shelf feature extractors. network weights initial training or freezing early layers improves compared scratch.