Deep Convolutional Neural Networks for Weeds and Crops Discrimination From UAS Imagery
Deep Convolutional Neural Networks for Weeds and Crops Discrimination From UAS Imagery
复制标题
DOI：
10.3389/frsen.2022.755939
复制
发表时间：
2022-02-11
期刊：
影响因子：
--
通讯作者：
Dorbu, Freda
中科院分区：
文献类型：
--
作者：
Hashemi-Beni, Leila;Gebrehiwot, Asmamaw;Dorbu, Freda
杂草是危害作物产量的重要因素之一，它通过入侵作物和窒息牧场，并显着降低收获作物的质量。除草剂在农业中被广泛用于控制杂草，然而，在农业中过量使用除草剂会导致环境污染以及产量降低。准确绘制作物/杂草地图对于确定杂草的位置和对这些区域进行局部处理至关重要。对灵活、准确和低成本的精准农业技术的需求日益增加，导致了基于UAS的遥感数据收集和方法的进步。深度学习方法已成功用于不同领域的UAS数据处理和映射任务。这项研究调查，比较和评估了深度学习方法在两个开源和已发布的基准数据集上的作物/杂草识别性能，这些数据集由不同的UAS（田间机器人和无人机）捕获并由专家标记。我们具体研究了以下架构：1）U-Net Model 2）SegNet 3）FCN（FCN-32 s，FCN-16 s，FCN-8 s）4）DepLabV 3+。对深度学习模型进行了微调，将UAS数据集分为三类（背景、作物和杂草）。U-Net实现的分类准确率为77.9%，高于SegNet的62.6%，FCN-32 s的68.4%，FCN-16 s的77.2%，略低于FCN-8 s的81.1%和DepLab v3+的84.3%。实验结果表明，与其他分类器相比，基于ResNet-18的分割模型（如DepLab v3+）可以准确地提取杂草。
Weeds are among the significant factors that could harm crop yield by invading crops and smother pastures, and significantly decrease the quality of the harvested crops. Herbicides are widely used in agriculture to control weeds; however, excessive use of herbicides in agriculture can lead to environmental pollution as well as yield reduction. Accurate mapping of crops/weeds is essential to determine weeds' location and locally treat those areas. Increasing demand for flexible, accurate and lower cost precision agriculture technology has resulted in advancements in UAS-based remote sensing data collection and methods. Deep learning methods have been successfully employed for UAS data processing and mapping tasks in different domains. This research investigate, compares and evaluates the performance of deep learning methods for crop/weed discrimination on two open-source and published benchmark datasets captured by different UASs (field robot and UAV) and labeled by experts. We specifically investigate the following architectures: 1) U-Net Model 2) SegNet 3) FCN (FCN-32s, FCN-16s, FCN-8s) 4) DepLabV3+. The deep learning models were fine-tuned to classify the UAS datasets into three classes (background, crops, and weeds). The classification accuracy achieved by U-Net is 77.9% higher than 62.6% of SegNet, 68.4% of FCN-32s, 77.2% of FCN-16s, and slightly lower than 81.1% of FCN-8s, and 84.3% of DepLab v3+. Experimental results showed that the ResNet-18 based segmentation model such as DepLab v3+ could precisely extract weeds compared to other classifiers.