| 摘要: |
| 为采用数码相机拍摄的水稻冠层图像来估测作物的氮素含量。以自然环境下获得的水稻冠层图像为研究对象，提出一种基于图像纹理色彩特征(LBPHSV)和ResNet50网络融合算法的氮素含量预测方法。LBPHSV+ResNet50融合算法是通过运用LBP算子和HSV颜色空间矩阵提取图像特征参数，将提取到的融合特征集作为ResNet50模型输入以加强对作物氮素营养的表征，并将预测结果与常用的多元线性回归、随机森林(RF)、支持向量回归模型、多层感知机、卷积神经网络、长短记忆网络(LSTM)及组合模型预测结果进行对比分析。结果显示:相比于浅层机器学习模型，深度学习算法能显著提高预测模型的准确率；LBPHSV+ResNet50融合模型的预测能力和泛化能力达到最优，R2和 RMSE分别为 0.97、0.02。相比于RF、LBP+LSTM、ResNet50，新模型的R2分别提升了16.36%、9.72%、16.55%和1.13%，RMSE 分别下降了 0.35、0.46、0.05和 0.002。因此，LBPHSV+ResNet50融合模型在预测水稻氮素含量时可提供令人满意的性能，能够满足对水稻氮素营养无损精准监测的农业需求。 |
| 关键词: 水稻冠层 氮素 监测 特征选择 ResNet50 |
| DOI：10.11841/j.issn.1007-4333.2023.01.04 |
| 投稿时间：2022-07-12 |
| 基金项目:国家自然科学基金项目(41961048)；江西省重点研发计划项目(20212BBF61013,20212BBF63040,20202BBFL63046,20192BBF60052)；江西省国家级高层次人才创新创业项目；江西省“双千计划”项目联合资助 |
|
| A method for monitoring nitrogen nutrition in rice canopy based on LBPHSV+ResNet50 fusion |
|
YE Chun1,2,LIU Ying1*,LIU Jizhong1,SHU Shifu2,LI Yanda2,WU Luofa2
|
| (1.School of Advanced Manufacturing, Nanchang University, Nanchang 330038, China;2.Institute of Agricultural Engineering/Jiangxi Province Engineering Research Center of Intelligent Agricultural Machinery Equipment/Jiangxi Province Engineering Research Center of Information Technology in Agriculture, Jiangxi Academy of Agricultural Sciences, Nanchang 330200, China) |
| Abstract: |
| 1Abstract: Abstract Nitrogen is an important indicator reflecting the nitrogen nutrition status of crops, and its content is closely related to crop growth and development, photosynthesis capacity, and crop yield. With the increasing maturity of image processing technology, choosing canopy images to estimate the nitrogen content of crops has become an essential technical means. This study takes the canopy images of rice growing under the natural environment as the research object and proposes a new method based on feature extraction and ResNet50 to predict the nitrogen content of rice. LBPHSV+ResNet50 fusion algorithm uses Local Binary Pattern-HSV fusion feature set as input, leaf area concentration(LNC)as output, and ResNet50 as regression prediction algorithm to enhance the characterization of crop nitrogen nutrition. Multiple Linear Regression, Random Forest(RF), Support Vector Regression, Multilayer Perception, Convolutional Neural Network and Long Short-Term Memory(LSTM)are adopted to establish the nitrogen content estimation model, respectively. The results show that: Compared with the machine learning model, the deep learning algorithm can significantly improve the prediction accuracy. The LBPHSV+ResNet50 model proposed in this study has the best prediction ability and generalization ability, and R2 and RMSE are 0. 97 and 0. 02, respectively. Compared with RF, LBP+LSTM, ResNet50 and LBP+ResNet50 fusion models, the R2 of LBPHSV+ResNet50 model increased by 16. 36%, 9. 72%, 16. 55%, and 1. 13%, and the RMSE decreased by 0. 35, 0. 46, 0. 05 and 0. 002, respectively. In conclusion, the LBPHSV+ResNet50 model provides satisfactory performance in predicting rice nitrogen content, and this model can meet the agricultural needs for non-destructive and accurate monitoring of nitrogen nutrition in rice. |
| Key words: rice canopy nitrogen prediction local binary pattern feature selection ResNet50 |
