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

Automated summary

Machine learning methods using aerial imagery have been widely used for crop classification. Traditional per-pixel-based, object-based, and patch-based methods have been used, but deep learning-based systems are becoming popular. However, building complex deep neural networks for aerial imagery is challenging due to limited labeled data and the variability associated with agricultural data. This article discusses these challenges and evaluates transfer learning methodologies for improving remote sensing image classification performance.