本文为德国耶拿·弗里德里希·席勒大学（作者：Christoph Rieke）的硕士论文，共69页。

本论文的目的是通过深度学习实例分割，从卫星图像中描绘分割出农田地块。虽然手动描绘是准确的，但非常耗时，许多采用传统图像分割技术的自动化方法难以捕捉可能出现的各种场景。深度学习已经被证明在各种计算机视觉任务中都是成功的，并且可能是一个很好的候选者，能够对农业领域的应用进行准确、有效、概括的描述。

本文对来自丹麦的Sentinel-2图像数据和相应的农田多边形进行了完全卷积的实例分割架构设计（算法改进来自Li等人，2016年）。与许多其他方法不同，该模型在未经处理的RGB图像上运行，而没有进行显著的前处理和后处理。经过训练，该模型成功地预测了外置图像芯片的区域边界，将研究成果在不同的区域大小、形状和其他性质上进行了推广，但在某些情况下会表现出特征问题。在第二个实验中，该模型被训练为同时预测田地实例的作物类型，在这种情况下的性能明显较差。虽然许多田地都被正确地划定了，但预测的作物种类是错误的。总体而言，该方向的研究结果是有希望的，证明了深度学习方法的有效性，此外，该方法为将来的进一步改进提供了许多方向。

This thesis aims to delineate agriculturalfield parcels from satellite images via deep learni ng instance segmentation.Manual delineation is accurate but time consuming, and many automatedapproaches with traditional image segmentation techniques struggle to capturethe variety of possible field appearances. Deep learning has proven to besuccessful in various computer vision tasks, and might be a good candidate toenable accurate, performant and generalizable delineation of agricultural fields.Here, a fully convolutional instance segmentation architecture (adapted from Liet al., 2016), was trained on Sentinel-2 image data and correspondingagricultural field polygons fromDenmark. In contrast to many otherapproaches, the model operates on raw RGB images without significant pre- andpost-processing. After training, the model proved successful in predictingfield boundaries on held-out image chips. The results generalize acrossdifferent field sizes, shapes and other properties, but show characteristic problemsin some cases. In a second experiment, the model was trained to simultaneouslypredict the crop type of the field instance. Performance in this setting wassignificantly worse. Many fields were correctly delineated, but the wrong cropclass was predicted. Overall, the results are promising and prove the validityof the deep learning approach. Also, the methodology offers many directions forfuture improvement.

1 引言

2 基于深度学习的实例分割

3 研究方法

4 研究结果与评估

5 未来工作展望

6 结论

更多精彩文章请关注公众号：