基于改进ResNet18的植物生长阶段识别方法及智慧植物补光的实现

doi:10.16035/j.issn.1001-7283.2025.01.032

摘要/Abstract

摘要：

为了能准确地识别植物的生长阶段从而实现智慧植物补光，设计一套能识别植物种类和生长阶段的智慧植物补光系统，其中识别植物生长阶段以ResNet18模型进行改进，用深度可分离卷积代替传统卷积，并引入SE模块（squeeze and excitation module）来提高模型任务处理的效率和准确性，结合早停法和学习率衰减机制来训练，避免过拟合。以番茄为研究对象进行验证，识别其幼苗期、开花坐果期和果实成熟期。结果表明，改进ResNet18模型的识别准确率达到了96.57%，比原模型提高了4.93个百分点，单张识别时间为0.27 s，比原模型快了0.30 s，模型体积为原模型的14%，同时，改进后的模型在测试集准确率、参数量和Macro F1得分等方面都优于ResNet18、ResNet34、AlexNet和VGG16四种卷积神经网络。最后，将改进ResNet18模型应用于植物补光系统，实际识别番茄生长阶段的准确率达到了96.49%，并能输出预期的光谱。该系统能精准地识别植物种类及其生长阶段，从而调用匹配植物及其生长阶段的光配方，达到智慧补光的目的。

关键词: 智慧补光, 卷积神经网络, 改进ResNet18模型, 植物生长阶段识别

Abstract:

In order to accurately identify the growth stage of plants and achieve intelligent plant light supplement, a set of intelligent plant supplementary light system that can identify plant species and growth stages was designed. The ResNet18 model was used to improve the identification of plant growth stage, and the deep separable convolution was used to replace traditional convolution. The SE module (squeeze and excitation module) was introduced to improve the efficiency and accuracy of model task processing, and early stop method and learning rate attenuation mechanism were combined for training to avoid overfitting. Tomato was used as the research object to verify and identify its seedling, flowering and fruiting, fruit ripening stages. The results showed that the recognition accuracy of the improved ResNet18 model reaches 96.57%, which was 4.93 percentage points higher than the original model. The single recognition time was 0.27 seconds, which was 0.30 seconds faster than the original model. The model volume was 14% of the original model. At the same time, the improved model was superior to the four convolutional neural networks of ResNet18, ResNet34, AlexNet and VGG16 in terms of test set accuracy, parameter quantity and Macro F1 score. Finally, the improved ResNet18 model was applied to the plant light supplement system, and the actual recognition accuracy of tomato growth stage reached 96.49%, and the expected spectrum was output. The system can accurately identify plant species and their growth stage, so as to invoke the light formula matching the plant and its growth stage, to achieve the purpose of intelligent plant light supplement.

Key words: Intelligent light supplement, Convolutional neural network, Improved ResNet18 model, Identification of plant growth stage

王安, 陈湛旭, 孔景徐, 吴思源, 何绍威, 张嘉玲, 万巍. 基于改进ResNet18的植物生长阶段识别方法及智慧植物补光的实现[J]. 作物杂志, 2025 (1): 250–259

Wang An, Chen Zhanxu, Kong Jingxu, Wu Siyuan, He Shaowei, Zhang Jialing, Wan Wei. Identification Method of Plant Growth Stages Based on Improved ResNet18 and Realization of Intelligent Plant Lighting Supplement[J]. Crops, 2025(1): 250–259

图/表 17

图1

图2

图3

表1

ResNet18结构

图层名称 Layer name	输出尺寸 Output size	ResNet18
conv1	60×60	7×7, 64, stride=2
conv2	30×30	3×3 maxpool, stride=2 $3 × 3, ? 64 3 × 3, ? 64 ×2$
conv3	15×15	$3 × 3, ? 128 3 × 3, ? 128 ×2$
conv4	8×8	$3 × 3, ? 256 3 × 3, ? 256 ×2$
conv5	4×4	$3 × 3, ? 512 3 × 3, ? 512 ×2$
	1×1	Average pool, 1000-d fc, softmax
FLOPs		1.8×10⁹

表1

图4

图5

图6

图7

图8

图9

图10

图11

表2

图12

表3

图13

表4

参考文献 34

[1]	邢阿宝, 崔海峰, 俞晓平, 等. 光质及光周期对植物生长发育的影响. 北方园艺, 2018(3):163-172.
[2]	杨其长. LED在农业与生物产业的应用与前景展望. 中国农业科技导报, 2008, 10(6):42-47.
[3]	刘文科. LED植物工厂光质生物学研究与应用现状. 中国农业科技导报, 2018, 20(10):9-14. doi: 10.13304/j.nykjdb.2017.0669
[4]	Stutte G W, Edney S, Skerritt T. Photoregulation of bioprotectant content of red leaf lettuce with light-emitting diodes. HortScience, 2009, 44(1):79-82.
[5]	Chen X L, Xue X Z, Guo W Z, et al. Growth and nutritional properties of lettuce affected by mixed irradiation of white and supplemental light provided by light-emitting diode. Scientia Horticulturae, 2016, 200(8):111-118.
[6]	Giliberto L, Perrotta G, Pallara P, et al. Manipulation of the blue light photoreceptor cryptochrome 2 in tomato affects vegetative development, flowering time, and fruit antioxidant content. Plant Physiology, 2005, 137(1):199-208. doi: 10.1104/pp.104.051987 pmid: 15618424
[7]	汪颖, 杨新琴, 徐沛, 等. 不同红蓝光配比LED光源对蔬菜育苗的影响. 浙江农业科学, 2020, 61(4):649-653. doi: 10.16178/j.issn.0528-9017.20200414
[8]	张现征, 王丹, 董飞, 等. 不同比例红蓝光对番茄幼苗生长发育及光合特性的影响. 江苏农业科学, 2019, 47(14):136-138.
[9]	卢纯, 张亚红, 李青. LED不同光质补光对日光温室冬春茬番茄生长及光合特性的影响. 江苏农业科学, 2020, 48(8):127- 134.
[10]	Hernández R, Kubota C. Physiological responses of cucumber seedlings under different blue and red photon flux ratios using LEDs. Environmental and Experimental Botany, 2016, 121:66-74.
[11]	王芳, 高芳云, 吕顺, 等. 不同比例红蓝LED灯对蔬菜育苗的补光效应. 热带作物学报, 2015, 36(8):1398-1402.
[12]	林魁, 黄枝, 林碧英, 等. 光强和红蓝光配比对瓠瓜幼苗生长及生理生化特性的影响. 西北植物学报, 2017, 37(3):517-525.
[13]	王利, 李小娥, 黄远, 等. 不同红蓝LED组合光源对葫芦和南瓜幼苗生长和生理参数的影响. 长江蔬菜, 2015(8):27-30.
[14]	崔世钢, 陈苗, 张永立, 等. 基于LED光源水培生菜最佳光配方的筛选. 江苏农业科学, 2020, 48(16):152-155.
[15]	李志鑫, 李松霖, 成杰, 等. 白蓝LED补光对温室甜瓜幼苗生长发育的影响. 山东农业科学, 2023, 55(1):84-88.
[16]	高波, 杨振超, 李万青, 等. 3种不同LED光质配比对芹菜生长和品质的影响. 西北农业学报, 2015, 24(12):125-132.
[17]	唐永康, 郭双生, 艾为党, 等. 不同比例红蓝LED光照对油麦菜生长发育的影响. 航天医学与医学工程, 2010, 23(3):206-212.
[18]	胡举伟, 代欣, 孙广玉. 不同红蓝光配比处理对桑树幼苗生长的影响. 森林工程, 2019, 35(4):28-31.
[19]	马旭, 林超辉, 齐龙, 等. 不同光质与光照度对水稻温室立体育秧秧苗素质的影响. 农业工程学报, 2015, 31(11):228-235.
[20]	陈玲, 王月福, 张晓军, 等. 不同红蓝组合光源对花生幼苗根系生长及根系活力的影响. 中国作物学会—2015年学术年会论文摘要集, 2015:219.
[21]	陈玲, 张晓军, 王月福, 等. 红蓝组合光源对花生幼苗根系生长的影响. 花生学报, 2016, 45(1):15-19.
[22]	王加真, 金星, 冯梅, 等. 不同红蓝光配比对茶树生长及生物化学成分的影响. 江苏农业科学, 2019, 47(10):159-161,172.
[23]	冯志斌, 江华丽. 智能LED植物补光灯控制系统设计. 长春师范大学学报, 2021, 40(12):53-58.
[24]	马州生. 用于植物生长的LED光强自动调节系统. 食品工业, 2019, 40(6):246-249.
[25]	张立萍. 智能型LED植物生长灯控制系统设计. 赤峰学院学报(自然科学版), 2021, 37(1):57-62.
[26]	李东, 郭伟玲, 尹飞, 等. LED植物补光系统控制研究进展. 照明工程学报, 2022, 33(2):64-70.
[27]	Concepcion R S, Alejandrino J D, Lauguico S C, et al. Lettuce growth stage identification based on phytomorphological variations using coupled color superpixels and multifold watershed transformation. International Journal of Advances in Intelligent Informatics, 2020, 6(3):261-277.
[28]	赵康迪, 单玉刚, 袁杰, 等. 基于实例分割的玉米虫害检测研究. 河南农业科学, 2022, 51(12):153-161.
[29]	姜红花, 杨祥海, 丁睿柔, 等. 基于改进ResNet18的苹果叶部病害多分类算法研究. 农业机械学报, 2023, 54(4):295-303.
[30]	牛智有, 于重洋, 吴志陶, 等. 基于改进ResNet18模型的饲料原料种类识别方法. 农业机械学报, 2023, 54(2):378-385,402.
[31]	张垚鑫, 朱荣光, 孟令峰, 等. 改进ResNet18网络模型的羊肉部位分类与移动端应用. 农业工程学报, 2021, 37(18):331-338.
[32]	Shafig M, Gu Z Q. Deep residual learning for image recognition: A Survey. Applied Sciences-Basel, 2022, 12(18):8972.
[33]	Ioffe S, Szegedy C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. International Conference on Machine Learning, 2015, 37:448-456.
[34]	Selvaraju R, Cogswell M, Das A, et al. Grad-CAM: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision, 2020, 128(2):336-359.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

模型 Model	测试准确率 Test-accuracy (%)	参数量 Parameter quantity	Macro F1	识别时间 Identification time (s)
AlexNet	93.56	4.09×10⁶	0.9210	0.13
VGG16	94.56	5.03×10⁷	0.9247	1.57
ResNet18	91.64	1.12×10⁷	0.8990	0.57
ResNet34	95.32	2.13×10⁷	0.9417	0.93
改进ResNet18 Improved ResNet18	96.57	1.53×10⁶	0.9630	0.27

生长阶段 Growth stage	正确次数 Number of correct	错误次数 Number of errors	成功率 Success rate (%)
幼苗期Seedling stage	50	0	100.00
开花坐果期 Flowering and fruit-setting stage	67	3	95.71
果实成熟期Fruit ripening stage	75	5	93.75

生长阶段Growth stage	红光辐照度Er (W/m²)	蓝光辐照度Eb (W/m²)	红蓝比Er-Eb ratio	光通量密度PPFD [μmol/(m²·s)]
幼苗期Seedling stage	10.00	3.31	3.02	68.44
开花坐果期Flowering and fruit-setting stage	8.84	1.24	7.13	54.00
果实成熟期Fruit ripening stage	8.91	1.76	5.07	56.43