欢迎访问《草业学报》官方网站,今天是 分享到:

草业学报 ›› 2023, Vol. 32 ›› Issue (12): 90-103.DOI: 10.11686/cyxb2023046

• 研究论文 • 上一篇    下一篇

栽培苜蓿草地智能感知系统关键生物物理指标实时监测及分析算法研究

苗春丽1(), 李仲贤2, 赵志成3, 伏帅1, 高金龙1, 刘洁1, 冯琦胜1, 梁天刚1()   

  1. 1.兰州大学草地农业科技学院,草地农业生态系统国家重点实验室,兰州大学农业农村部牧草创新重点实验室,兰州大学草地农业教育工程研究中心,甘肃 兰州 730020
    2.兰州大学网络安全与信息化办公室,甘肃 兰州 730000
    3.宁夏彭阳县畜牧技术推广服务中心,宁夏 彭阳 756500
  • 收稿日期:2023-02-13 修回日期:2023-05-04 出版日期:2023-12-20 发布日期:2023-10-18
  • 通讯作者: 梁天刚
  • 作者简介:Corresponding author. E-mail: tgliang@lzu.edu.cn
    苗春丽(1995-),女,山东日照人,在读博士。E-mail: 120220900380@lzu.edu.cn
  • 基金资助:
    中国工程院战略研究与咨询项目(2022-HZ-09);财政部和农业农村部:国家现代农业产业技术体系(CARS-34);甘肃省林业和草原局科技创新项目(kjcx2022010);兰州大学中央高校基本科研业务费专项资金(lzujbky-2022-sp13)

Real-time monitoring and analysis algorithm for key biophysical indicators of cultivated alfalfa in a grassland intelligent perception system

Chun-li MIAO1(), Zhong-xian LI2, Zhi-cheng ZHAO3, Shuai FU1, Jin-long GAO1, Jie LIU1, Qi-sheng FENG1, Tian-gang LIANG1()   

  1. 1.College of Pastoral Agriculture Science and Technology,Lanzhou University,State Key Laboratory of Grassland Agro-ecosystem,Key Laboratory of Grassland Livestock Innovation,Ministry of Agriculture and Rural Affairs,Engineering Research Center of Grassland Industry,Ministry of Education,Lanzhou 730020,China
    2.Network Security and Informatization Office,Lanzhou University,Lanzhou 730000,China
    3.Ningxia Pengyang County Animal Husbandry Technology Extension Service Center,Pengyang 756500,China
  • Received:2023-02-13 Revised:2023-05-04 Online:2023-12-20 Published:2023-10-18
  • Contact: Tian-gang LIANG

摘要:

苜蓿作为重要的优质牧草,其产量和品质的监测对草牧业发展具有十分重要的作用。传统大范围栽培苜蓿盖度和产量的地面调查以及卫星遥感反演易受天气、人力和财力等因素的影响,在时空动态监测方面具有一定局限性。近年来随着传感器和人工智能(AI)的快速发展及其在作物监测和分析方面的普遍应用,为栽培苜蓿盖度及产量的精准实时估测提供了新的契机。本研究以新疆、内蒙古、甘肃、宁夏等北方四省区栽培苜蓿为研究对象,结合地面实测资料,利用深度学习算法(DL)、多元线性回归(MLR)和随机森林(RF)方法建立了栽培苜蓿盖度和产量估测模型;并对模型精度进行了评价。研究结果表明:1)总体而言,我国新疆、甘肃河西等地区的栽培苜蓿以灌溉为主,地块集中连片、地势平坦,一年刈割3~4次,苜蓿草地在盛草期的平均产量和盖度达5362.81 kg·hm-2、96.29%;以旱作生产方式为主的甘肃陇东、宁夏南部等地区的栽培苜蓿草地大多种植在山区水平梯田,一年刈割2~3次,其盛草期的平均产量和盖度达3987.57 kg·hm-2、91.55%;2)基于无人机可见光遥感数据的苜蓿草地盖度深度学习模型的R2达0.99,均方根误差(RMSE)为1.44%,模型准确度为92%,对栽培苜蓿草层盖度的动态监测效果较好;3)利用经度、纬度及海拔和苜蓿关键生物物理指标草高、盖度及二者乘积构建的苜蓿产量RF模型相较于MLR模型可以提升产量的估测精度,最优估测模型测试集的R2为0.69,RMSE为1151.24 kg·hm-2。研究结果可以为栽培苜蓿智能感知系统的关键生物物理指标快速评估提供算法依据,对多点位高时频的网络化、自动化和智能化栽培苜蓿生长数据采集与动态分析系统应用具有重要的技术支撑作用。

关键词: 智能感知系统, 苜蓿盖度, 苜蓿产量, 机器学习, 深度学习, 无人机

Abstract:

Alfalfa, as an important high-quality forage, plays a vital role in the development of grassland animal husbandry. Traditional ground surveys and satellite remote sensing to evaluate large-scale alfalfa cover and yield are easily affected by weather, manpower, and financial factors, and have certain limitations when used in spatio-temporal dynamic monitoring. In recent years, with the rapid development of sensors and artificial intelligence (AI) and their widespread application in crop monitoring and analysis, new methods have been developed for accurate and real-time estimation of alfalfa cover and yield. In this study, alfalfa cover and yield estimation models were established using deep learning (DL) and multiple linear regression (MLR) and random forest (RF) methods combined with ground-measured data for alfalfa cultivated in four northern provinces of Xinjiang, Inner Mongolia, Gansu, and Ningxia. The accuracy of all the models was evaluated. It was found that: 1) In general, cultivated alfalfa in northern Xinjiang, Hexi in Gansu, and other regions is mainly irrigated, with concentrated and contiguous plots and flat terrain, and is harvested three to four times a year, with an average yield and vegetation cover of 5362.81 kg·ha-1 and 96.29%, respectively, at the peak growth period. Alfalfa in Eastern Gansu, southern Ningxia, and other regions is mainly cultivated using a dryland production system, is mostly grown on horizontal terraces in mountainous areas, and is harvested two to three times a year, with an average yield and cover of 3987.57 kg·ha-1 and 91.55%, respectively, at the peak growth period. 2) For the alfalfa cover deep learning model using unmanned aerial vehicle visible light remote sensing data as inputs, the R2 was 0.99, the root mean squared error was 1.44%, and the model accuracy was 92%, indicative of good ability to dynamically monitor cultivated alfalfa cover. 3) Compared with the MLR model, the RF model based on longitude, latitude, altitude, and key biophysical indicators of alfalfa height, cover, and their product provided more accurate estimates of alfalfa yield. The R2 of the optimal estimation model test set was 0.69, and the RMSE was 1151.24 kg·ha-1. In summary, we have established an intelligent perception system based on an algorithm for the rapid evaluation of key biophysical indicators of cultivated alfalfa. These findings provide technical support for the application of networked, automated, and intelligent alfalfa growth data acquisition and dynamic analysis systems in multiple locations and with a high frequency.

Key words: intelligent sensing system, alfalfa coverage, alfalfa yield, machine learning, deep learning, UAV