基于迁移学习的无人机影像耕地信息提取方法

随着精准农业技术的发展,对农作物用地信息快速、准确提取的需求越来越高。同时,无人机技术以其方便、高效、具有低空云下飞行能力等优势被广泛应用于自然资源的调查中。但无人机影像普遍光谱信息较为匮乏,因此很难准确、快速地提取出耕地信息。基于此,提出了一种利用迁移学习机制的耕地提取方法(TLCLE)。首先,利用深度卷积神经网络(DCNN)剔除线状地物(道路、田埂等),然后,通过引入迁移学习机制将DCNN特征训练过程中得到的特征提取方法迁移到耕地提取中,最后,将所提方法与利用易康(e Cognition)软件进行耕地提取(ECLE)结果进行对比。研究结果表明:对于实验影像1、2,TLCLE方法耕地提取总体精度分别为91.9%、88.1%,ECLE方法总体精度分别为90.3%、88.3%,2种方法提取精度相当,在保证耕地地块完整、连续性上TLCLE方法优于ECLE方法。

Cultivated Land Information Extraction from High Resolution UAV Images Based on Transfer Learning

The development of precision agriculture demands high accuracy and efficiency of cultivated land information extraction. Due to the low spatial resolution of satellite remote sensing images, it is difficult to identify cultivated land of small areal extent in critical regions, which requires image data of high spatial resolution for specific or general cases. Simultaneously, unmanned aerial vehicle (UAV) has been increasingly used for natural resource applications in recent years as a result of their great availabilities, the miniaturization of sensors, and the ability to deploy UAV relatively quickly and repeatedly at low altitudes. But most UAV images lack spectral information and cultivated land information extraction which usually leads to an unsatisfactory result. Based on this, a novel cultivated land information extraction method based on transfer learning (TLCLE) was proposed. Firstly, linear features (roads and ridges, etc.) were rejected based on deep convolutional neural network (DCNN). Secondly, feature extraction method learned from DCNN was used for extracting cultivated land information by introducing transfer learning mechanism. Finally, cultivated land information extraction results were completed by the TLCLE method and eCognition software for cultivated land information extraction (ECLE). The experimental results show that TLCLE can obtain equivalent accuracy to ECLE, and it outperforms ECLE in terms of guaranteeing the integrity and continuity of cultivated land information.