基于深度学习的无人机遥感生态灌区杂草分类

为了更好地推进生态灌区建设,对灌区内杂草进行检测、控制,提出一种基于卷积神经网络的杂草分类和密度测算方法.通过无人机低空拍摄采集3种杂草(藜草、葎草、苍耳)和3种作物(小麦、花生、玉米)作为数据集,经过裁剪、灰度化等前期处理,并通过旋转方式扩充数据集,最后收集17 115张训练样本和750张测试样本,然后将训练集输送给卷积神经网络,采用softmax回归,实现6类植物的分类.为降低网络参数,文中试验了100×100和300×300不同分辨率图像对识别精度的影响,分类结果表明300×300分辨率时识别率最高可达到95.6%.另外为实现针对特定杂草的防控,提出了一种检测单一杂草密度的方法,可实现对灌区内各种杂草的精确监控,为后期杂草防控的精准施药提供依据,对实现高效、绿色、安全的现代农业具有重要理论意义和实用价值.

Classification of weeds in ecological irrigation areas based on deep learning and remote sensing images taken by drone at low-altitude

In order to better promote the construction of ecological irrigation areas, it is required to detect and control weeds in the areas. Thus, in this paper a method was proposed based on convolutional neural network(CNN)for weed classification and density estimation in ecological irrigation areas.The images were taken by a drone at low-altitude for three kinds of weeds namely chenopodium album, humulusscandens and xanthium sibiricum as well as 3 sorts of crops such as wheat, peanut seeding and maize, and then 17 115 training samples and 750 test samples were harvested through trimming, gray scale and rotation. Finally, the training sets were input into the CNN, and the classification of 6 types of plants was conducted by means of softmax regression.In order to reduce the network parameters, the effect of 100×100 and 300×300 resolution images on recognition accuracy was also clarified.The results showed that the highest recognition rate of 300×300 resolution could reach as high as 95.6%accuracy.In order to prevent and control specific weeds, a method of detecting single weeds density was presented, too.Through this method, accurate monitoring of various weeds in irrigation areas can be achieved. This method can provide a basis for precise applying pesticide, and has important significance and theoretical and practical values for realizing efficient, green, and safe modern agriculture. 