Diagnosis of Potassium Nutrition in Rice Based on CA_MobileViT Model

doi:10.13304/j.nykjdb.2024.0146

Abstract

Abstract:

In order to rapidly and accuratelely diagnose and recognize potassium stress degree of rice， The Huanghuazhan variety of late rice was used as material， and 4 potassium fertilization levels were set respectively for base fertilizer and topdressing at jointing stage. Taking scanned images of 3 spread leaves of main stem at tiller stage and jointing stage as data set， and MobileViT as the skeleton， the CA_MobileViT model was constructed after introducing Coordinate Attention to the BN Layer of each 3×3 convolution in the layer of MobileViT and introduce transfer learning. The results were verified by comparing EfficientNet-V2， ConvNeXt， MobileViT and CA_MobileViT models. The results showed that the accuracy rates of the 4 models were 98.4%， 98.5%， 94.2% and 95.3%， respectively. The accuracy of CA_MobileViT model was 95.3%， 1.1 percent point higher than that of MobileViT model， but 3.1 percent point lower than that of EfficientNet-V2 model and 3.2 percent point lower than that of ConvNeXt model. However， the number of parameters of CA_MobileViT models was about 1/4 of the EfficientNet-V2 and ConvNeXt large models， and the training time was reduced by about 1/3. The improved CA_MobileViT model has a high accuracy for the diagnosis of potassium stress degree of rice， and could effectively guide the scientific potassium topdressing management of rice， and also provide a universal and feasible method for the rapid and accurate diagnosis of nutrition of other crops.

Key words: rice, potassium nutrition diagnosis, lightweight hybrid model, transfer learning, Coordinate Attention

摘要：

为快速、准确地诊断水稻钾素的胁迫程度，以晚稻品种黄华占为材料进行大田试验，对基肥和拔节期追肥分别设置4个钾素施肥水平处理。以分蘖期和拔节期水稻主茎顶部3片展开叶的扫描图像作为数据集，以MobileViT为骨架，将Coordinate Attention引入到MobileViT的Layer中的每个3 $×$ 3卷积的BN层之后，构建CA_MobileViT模型并引入迁移学习。通过对比EfficientNet-V2、ConvNeXt、MobileViT和CA_MobileViT模型进行验证，结果表明：4个模型的准确率分别为98.4%、98.5%、94.2%、95.3%。其中CA_MobileViT模型的准确率为95.3%，比MobileViT模型提高了1.1百分点，却比EfficientNet-V2模型准确率低3.1百分点，比ConvNeXt模型准确率低3.2百分点，但CA_MobileViT模型参数量约EfficientNet-V2和ConvNeXt大模型的1/4，训练时间缩短了约1/3。改进的CA_MobileViT模型对于水稻钾素胁迫程度诊断具有较高的准确率，能有效地指导水稻钾素科学追肥管理，也为其他农作物的营养快速、精确诊断提供了一种普适、可行的方法。

关键词: 水稻, 钾素营养诊断, 轻量混合模型, 迁移学习, Coordinate Attention

CLC Number:

S126

Zheng WU, Hongyun YANG, Aizhen SUN, Jie KONG, Shumei HUANG. Diagnosis of Potassium Nutrition in Rice Based on CA_MobileViT Model[J]. Journal of Agricultural Science and Technology, 2025, 27(8): 80-88.

吴正, 杨红云, 孙爱珍, 孔杰, 黄淑梅. 基于CA_MobileViT模型的水稻钾素营养诊断研究[J]. 中国农业科技导报, 2025, 27(8): 80-88.

Figures/Tables 8

FIG. 1 MobileViT flowA：Overall architecture of MobileViT；B：MobileViTBlock structure；P represents the pixels in the patch； N is the number of patches； d represents the processed feature dimension； H and W represent the height and width of the feature map or input image， respectively

Fig. 2 CA_MobileViT modelA： CoordATT module structure； B： Improved network backbone； C： Model process structure

Fig. 3 Performance curve of the jointing stage modelA：Model accuracy curve；B：Model loss curve

Table 1 Hyperparameter Settings of the model and the number of model parameters

模型 Model	优化器 Optimizer	学习率 Learning rate	总参数 Total parameters	轮次时间 Epoch time/s
EfficientNet-V2	SGD	0.010 0~0.000 1	20 182 612	276
ConvNeXt	AdamW	0.010 0~0.000 2	27 816 580	289
MobileViT	AdamW	0.010 0~0.000 2	5 578 632	198
CA_MobileViT	AdamW	0.010 0~0.000 2	5 643 988	203

Fig. 4 Performance curve of the tillering stage modelA：Model accuracy curve；B：Model loss curve

Table 2 Comparison of model evaluation indicators

模型 Model	准确率 Accuracy/%	精度 Precision/%	召回率 Recall/%	特异性 $S p e c i f i c i t y$ /%	F₁分数 F₁-score
EfficientNet-V2	98.4	98.5	98.5	99.5	0.985
ConvNeXt	98.5	98.5	98.5	99.5	0.985
MobileViT	94.2	94.4	94.2	98.0	0.942
CA_MobileViT	95.3	95.3	95.3	98.4	0.952

Table 2 Comparison of model evaluation indicators

模型 Model	准确率 Accuracy/%	精度 Precision/%	召回率 Recall/%	特异性 $S p e c i f i c i t y$ /%	F₁分数 F₁-score
EfficientNet-V2	98.4	98.5	98.5	99.5	0.985
ConvNeXt	98.5	98.5	98.5	99.5	0.985
MobileViT	94.2	94.4	94.2	98.0	0.942
CA_MobileViT	95.3	95.3	95.3	98.4	0.952

Fig. 5 Confusion matrix at jointing stage

Fig. 6 Grad-CAM thermal maps

References 23

[1]	赵鹏姝,杨顺瑛,郝东利,等.水稻钾素营养的基因型特征及分子机制初探[J].土壤,2021,53(1):37-46.
	ZHAO P S, YANG S Y, HAO D L, et al.. Genotypic characteristics and molecular mechanism of potassium nutrition in rice [J]. Soils, 2021,53(1):37-46.
[2]	朱向东,王宏庭.土壤钾素管理研究进展[J].山西农业科学,2013,41(11):1274-1281.
	ZHU X D, WANG H T. Research progress of soil potassium management [J]. Shanxi Agric. Sci., 2013,41(11): 1274-1281.
[3]	王绍华,曹卫星,王强盛,等.水稻叶色分布特点与氮素营养诊断[J].中国农业科学,2002,35(12):1461-1466.
	WANG S H, CAO W X, WANG Q S, et al.. Positional distribution of leaf color and diagnosis of nitrogen nutrition in rice plant [J]. Sci. Agric. Sin., 2002,35(12):1461-1466.
[4]	罗建军.基于计算机视觉技术与高光谱技术的水稻氮素营养诊断研究[D].南昌:江西农业大学,2021.
	LUO J J. Nitrogen nutrition diagnosis of rice basedon Computer vision and Hyperspectral technology [D]. Nanchang: Jiangxi Agricultural University,2021.
[5]	FANG S D, WANG Y F, ZHOU G X, et al.. Multi-channel feature fusion networks with hard coordinate attention mechanism for maize disease identification under complex backgrounds [J/OL]. Comput. Electron. Agric., 2022,203:107486 [2024-11-12]. .
[6]	ESPEJO-GARCIA B, MALOUNAS I, MYLONAS N, et al.. Using efficientNet and transfer learning for image-based diagnosis of nutrient deficiencies [J/OL].Comput.Electron.Agric., 2022,196:106868 [2024-11-12]. .
[7]	BEVERS N, SIKORA E J, HARDY N B. Soybean disease identification using original field images and transfer learning with convolutional neural networks [J/OL]. Comput. Electron.Agric., 2022,203:107449 [2024-11-12]. .
[8]	AZGOMI H, HAREDASHT F R, SAFARI MOTLAGH M R. Diagnosis of some apple fruit diseases by using image processing and artificial neural network [J/OL]. Food Contr., 2023,145:109484 [2024-11-12]. .
[9]	周琼,杨红云,杨珺,等.基于BP神经网络和概率神经网络的水稻图像氮素营养诊断[J].植物营养与肥料学报,2019,25(1):134-141.
	ZHOU Q, YANG H Y, YANG J, et al.. Feasibility study of BP neural network and probabilistic neural network for nitrogen nutrition diagnosis of rice images [J]. J. Plant Nutr. Fertil., 2019,25(1):134-141.
[10]	杨红云,郭紫微,郭高飞,等.基于Stacking集成卷积神经网络的水稻氮素营养诊断[J].植物营养与肥料学报,2023,29(3):573-581.
	YANG H Y, GUO Z W, GUO G F, et al.. Rice nitrogen nutrition diagnosis based on stacking integrated convolutional neural network [J]. J. Plant Nutr. Fertil., 2023,29(3):573-581.
[11]	YU S, XIE L, HUANG Q L. Inception convolutional vision transformers for plant disease identification [J/OL]. Internet Things,2023,21:100650 [2024-11-12]. .
[12]	THAI H T, LE K H, NGUYEN N L T. FormerLeaf: an efficient vision transformer for Cassava Leaf Disease detection [J/OL]. Comput. Electron. Agric., 2023,204: 107518 [2024-11-12]. .
[13]	DHAR P. The carbon impact of artificial intelligence [J]. Nat. Mach. intell., 2020, 2:423-425.
[14]	LI S, SONG W F, QIN H, et al.. Deep variance network:an iterative,improved CNN framework for unbalanced training datasets [J/OL]. Pattern Recognit., 2018,81:294-308.
[15]	朱劲松,李欢,王世芳.基于卷积神经网络和迁移学习的钢桥病害识别[J].长安大学学报(自然科学版),2021,41(3):52-63.
	ZHU J S, LI H, WANG S F. Defect recognition for steel bridge based on convolutional neural network and transfer learning [J].J. Chang’an Univ. (Nat.Sci.), 2021,41(3):52-63.
[16]	CHOLLET F. Xception: Deep learning with depthwise separable convolutions [EB/OL]. (2017-04-04) [2024-10-16]. .
[17]	LIU Z, MAO H Z, WU C Y, et al.. A convnet for the 2020s [C]// Proceedings of 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).IEEE, 2022:11966-11976.
[18]	LIU X N, QI J T, ZHANG W R, et al.. Recognition method of maize crop rows at the seedling stage based on MS-ERFNet model [J/OL]. Comput. Electron. Agric., 2023, 211:107964 [2024-11-12]. .
[19]	SINGH A, GANAPATHYSUBRAMANIAN B, SINGH A K, et al.. Machine learning for high-throughput stress phenotyping in plants [J]. Trends Plant Sci., 2016,21(2):110-124.
[20]	YE Y B. An improved efficientnetv2 model based on visual attention mechanism: Application to identification of cassava disease [J/OL]. Comput. Intell. Neurosci., 2022, 1:1569911 [2024-11-12]. .
[21]	CHEN S, ZOU X J, ZHOU X Z, et al.. Study on fusion clustering and improved YOLOv5 algorithm based on multiple occlusion of Camellia oleifera fruit [J/OL]. Comput. Electron. Agric., 2023,206:107706. .
[22]	YANG Y, ZHANG L, REN L, et al.. MMViT-Seg:a lightweight transformer and CNN fusion network for COVID-19 segmentation [J/OL]. Comput. Meth. Programs Biomed., 2023, 230:107348 [2024-11-12]. .
[23]	徐明岗,李冬初,李菊梅,等.化肥有机肥配施对水稻养分吸收和产量的影响[J].中国农业科学, 2008,(10):3133-3139.
	XU M G, LI D C, LI J M, et al.. Effects of organic manure application combined with chemical fertilizers on nutrients absorption and yield of rice in Hunan of China [J]. Scientia Agric. Sin., 2008,(10):3133-3139.