农业作为经济扩展的重要驱动力,面临着在气候不确定性和资源有限的背景下维持不断增长的全球人口的问题。因此,利用先进的人工智能(AI)支持自主决策的“智能农业”变得越来越受欢迎。本文探讨了物联网(IoT)、人工智能(AI)、机器学习(ML)、遥感和变速率技术(VRT)如何协同作用,推动农业转型。研究的主要应用包括利用复杂算法预测土壤状况、改进农业产量预测、通过传感器数据诊断水分胁迫、通过图像识别识别植物病害和杂草、作物制图以及 AI 指导的作物选择。此外,VRT 精准施用水、农药和肥料,优化资源利用,提高可持续性和效率。为了有效满足全球粮食需求,本研究预测了结合 AI 驱动方法与传统方法的可持续农业未来。
- Crop yield prediction
- Disease and pest detection
- Weed identification and management
- Precision irrigation management
- Nutrient management
- Crop variety selection
- Intelligent harvesting and optimization:| - Crop yield prediction |
| :--- |
| - Disease and pest detection |
| - Weed identification and management |
| - Precision irrigation management |
| - Nutrient management |
| - Crop variety selection |
| - Intelligent harvesting and optimization: |
✓\checkmark
模型失败是由小数据集引起的;多样化的真实世界田间图像至关重要
[51,52]
✓\checkmark
准确性和运行时间受输入的大小和类型影响(例如,背景去除、植被指数)
✓\checkmark
成本高、耗时且专业;需要对无监督技术进行研究
✓\checkmark
准确性和推理时间存在权衡;特定场地的数据集需要重新训练
✓\checkmark
处理开销大且依赖超参数调整
✓\checkmark
轻量级模型和边缘设备是模型的有吸引力的替代方案,而模型通常不切实际
✓\checkmark
需要农民的知识;通过物联网和强化学习可能实现自动化
S. No Technology Applications Limitations and Challenges References
1. Artificial intelligence (AI) "- Predictive crop modeling
- Pest and disease prediction
- Soil health analysis
- Automated machinery" ✓ Limitations in real-time dataset availability [47-50]
✓ Absence of dataset standards
✓ Capturing data at close range
✓ Recognizing small symptoms
✓ Differences in image quality and lighting
✓ Disease development and similarities across classes
✓ Computational difficulties with big datasets
✓ Combining multimodal data
2. Deep learning (DL) and machine learning (ML) models "- Crop yield prediction
- Disease and pest detection
- Weed identification and management
- Precision irrigation management
- Nutrient management
- Crop variety selection
- Intelligent harvesting and optimization:" ✓ Model failures are caused by small datasets; varied, real-world field images are crucial [51,52]
✓ Accuracy and runtime are impacted by the size and type of input (e.g., background removal, vegetation indices)
✓ Costly, time-consuming, and specialized; unsupervised techniques require investigation
✓ Accuracy and inference time are traded off; site-specific datasets require retraining
✓ High processing overhead and reliance on hyperparameter adjustment
✓ Lightweight models and edge devices are attractive alternatives to models, which are frequently impractical
✓ Farmer knowledge is required; automation may be possible with IoT and reinforcement learning | S. No | Technology | Applications | | Limitations and Challenges | References |
| :--- | :--- | :--- | :--- | :--- | :--- |
| 1. | Artificial intelligence (AI) | - Predictive crop modeling <br> - Pest and disease prediction <br> - Soil health analysis <br> - Automated machinery | | $\checkmark$ Limitations in real-time dataset availability | [47-50] |
| | | | $\checkmark$ | Absence of dataset standards | |
| | | | $\checkmark$ | Capturing data at close range | |
| | | | $\checkmark$ | Recognizing small symptoms | |
| | | | $\checkmark$ | Differences in image quality and lighting | |
| | | | $\checkmark$ | Disease development and similarities across classes | |
| | | | $\checkmark$ | Computational difficulties with big datasets | |
| | | | $\checkmark$ | Combining multimodal data | |
| 2. | Deep learning (DL) and machine learning (ML) models | - Crop yield prediction <br> - Disease and pest detection <br> - Weed identification and management <br> - Precision irrigation management <br> - Nutrient management <br> - Crop variety selection <br> - Intelligent harvesting and optimization: | $\checkmark$ | Model failures are caused by small datasets; varied, real-world field images are crucial | [51,52] |
| | | | $\checkmark$ | Accuracy and runtime are impacted by the size and type of input (e.g., background removal, vegetation indices) | |
| | | | $\checkmark$ | Costly, time-consuming, and specialized; unsupervised techniques require investigation | |
| | | | $\checkmark$ | Accuracy and inference time are traded off; site-specific datasets require retraining | |
| | | | $\checkmark$ | High processing overhead and reliance on hyperparameter adjustment | |
| | | | $\checkmark$ | Lightweight models and edge devices are attractive alternatives to models, which are frequently impractical | |
| | | | $\checkmark$ | Farmer knowledge is required; automation may be possible with IoT and reinforcement learning | |
- Irrigation monitoring and control
- Soil monitoring
- Temperature and humidity monitoring
- Animal monitoring and tracking
- Water monitoring and controlling disease monitoring
- Air monitoring
- Fertilization monitoring| - Irrigation monitoring and control |
| :--- |
| - Soil monitoring |
| - Temperature and humidity monitoring |
| - Animal monitoring and tracking |
| - Water monitoring and controlling disease monitoring |
| - Air monitoring |
| - Fertilization monitoring |
✓ The gathering, storing, processing, and transfer of data pose security threats to IoT-based agriculture
✓ Signal jamming causes inefficiencies and economic losses by interfering with communication, GPS, and remote monitoring
✓ Decision-making, accuracy, and data dependability are all harmed by node capture and outages
✓ Attacks on data transmission lead to surveillance, theft, or poor farming decisions
✓ For prevention, strong procedures like encryption, authentication, monitoring, and system updates are important| $\checkmark$ The gathering, storing, processing, and transfer of data pose security threats to IoT-based agriculture |
| :--- |
| $\checkmark$ Signal jamming causes inefficiencies and economic losses by interfering with communication, GPS, and remote monitoring |
| $\checkmark$ Decision-making, accuracy, and data dependability are all harmed by node capture and outages |
| $\checkmark$ Attacks on data transmission lead to surveillance, theft, or poor farming decisions |
| $\checkmark$ For prevention, strong procedures like encryption, authentication, monitoring, and system updates are important |
S. No Technology Applications Limitations and Challenges References
3. Internet of thing (IoT) "- Irrigation monitoring and control
- Soil monitoring
- Temperature and humidity monitoring
- Animal monitoring and tracking
- Water monitoring and controlling disease monitoring
- Air monitoring
- Fertilization monitoring" "✓ The gathering, storing, processing, and transfer of data pose security threats to IoT-based agriculture
✓ Signal jamming causes inefficiencies and economic losses by interfering with communication, GPS, and remote monitoring
✓ Decision-making, accuracy, and data dependability are all harmed by node capture and outages
✓ Attacks on data transmission lead to surveillance, theft, or poor farming decisions
✓ For prevention, strong procedures like encryption, authentication, monitoring, and system updates are important" [53,54]
4. Variable-rate technology (VRT): map-based and sensor-based techniques "- Variable-rate fertilization
- Variable-rate seeding
- Variable-rate irrigation:" ✓ ✓quad Younger farmers are more likely than older farmers to use VRTs, and adoption decisions are based on farm economics; farmers place a higher priority on profitability [55]| S. No | Technology | Applications | | Limitations and Challenges | References |
| :--- | :--- | :--- | :--- | :--- | :--- |
| 3. | Internet of thing (IoT) | - Irrigation monitoring and control <br> - Soil monitoring <br> - Temperature and humidity monitoring <br> - Animal monitoring and tracking <br> - Water monitoring and controlling disease monitoring <br> - Air monitoring <br> - Fertilization monitoring | | $\checkmark$ The gathering, storing, processing, and transfer of data pose security threats to IoT-based agriculture <br> $\checkmark$ Signal jamming causes inefficiencies and economic losses by interfering with communication, GPS, and remote monitoring <br> $\checkmark$ Decision-making, accuracy, and data dependability are all harmed by node capture and outages <br> $\checkmark$ Attacks on data transmission lead to surveillance, theft, or poor farming decisions <br> $\checkmark$ For prevention, strong procedures like encryption, authentication, monitoring, and system updates are important | [53,54] |
| 4. | Variable-rate technology (VRT): map-based and sensor-based techniques | - Variable-rate fertilization <br> - Variable-rate seeding <br> - Variable-rate irrigation: | $\checkmark$ | $\checkmark \quad$ Younger farmers are more likely than older farmers to use VRTs, and adoption decisions are based on farm economics; farmers place a higher priority on profitability | [55] |
" MSE "=(1)/(N)sum_(i=1)^(N)(yi- hat(y)i)^(2)
N : Total number of observations in the dataset. i: Index for individual data points. yi: True label. yíi: Predicted value| $\text { MSE }=\frac{1}{N} \sum_{i=1}^{N}(y i-\hat{y} i)^{2}$ |
| :--- |
| N : Total number of observations in the dataset. i: Index for individual data points. yi: True label. yíi: Predicted value |
Model Performance Analysis Parameter Description References
Support Vector Machine (SVM) Accuracy =quad TP + TN where TP = True Positives, TN = True Negatives, FP = False Positives, FN = False Negatives SVM finds the hyperplane that maximizes the margin between classes. It uses hinge loss to discard the incorrect classifications. [56]
K-Nearest Neighbors (KNN) Accuracy =(" Correct Predictions ")/(" Total predictions ") KNN uses its neighbors' approval rating to classify a data point. Performance is strongly impacted by KNN selection. [57]
Decision Tree (DT) Gini Index =1-sum_(i)^(C)p^(2) where p is the proportion of samples belonging to class i To reduce the impurity (such as Gini or entropy), decision trees divide data according to feature thresholds [58]
Random Forest (RF) Accuracy =(1)/(T)sum_(i=1)^(T) Accuracy _(t) where T is the number of trees Random Forest combines predictions using an ensemble of decision trees to increase prediction accuracy and decrease overfitting [59]
Artificial Neural Network (ANN) Accuracy =(" Correct Predictions ")/(" Total predictions ") An ANN uses layers of neurons to map inputs to outputs, optimizing weights by backpropagation with a differentiable loss function, such as cross-entropy [60]
Naïve Bayes Posterior Probability, P(y∣X)= P(X)P(X∣y)P(y) P(yi∣Xi) : Posterior probability of the i-th data point being in class yi Naïve Bayes models eliminate complicated joint probability computations; they simplify calculations and are especially useful when the independence condition is roughly valid [61]
Long Short-Term Memory (LSTM) "" MSE "=(1)/(N)sum_(i=1)^(N)(yi- hat(y)i)^(2)
N : Total number of observations in the dataset. i: Index for individual data points. yi: True label. yíi: Predicted value" Temporal dependencies in sequential data are captured by LSTM, a recurrent neural network that utilizes gates (forget, input, and output) [62]| Model | Performance Analysis Parameter | Description | References |
| :--- | :--- | :--- | :--- |
| Support Vector Machine (SVM) | Accuracy $=\quad$ TP + TN where TP $=$ True Positives, TN $=$ True Negatives, FP = False Positives, FN = False Negatives | SVM finds the hyperplane that maximizes the margin between classes. It uses hinge loss to discard the incorrect classifications. | [56] |
| K-Nearest Neighbors (KNN) | Accuracy $=\frac{\text { Correct Predictions }}{\text { Total predictions }}$ | KNN uses its neighbors' approval rating to classify a data point. Performance is strongly impacted by KNN selection. | [57] |
| Decision Tree (DT) | Gini Index $=1-\sum_{i}^{C} p^{2}$ where $p$ is the proportion of samples belonging to class i | To reduce the impurity (such as Gini or entropy), decision trees divide data according to feature thresholds | [58] |
| Random Forest (RF) | Accuracy $=\frac{1}{T} \sum_{i=1}^{T}$ Accuracy $_{t}$ where T is the number of trees | Random Forest combines predictions using an ensemble of decision trees to increase prediction accuracy and decrease overfitting | [59] |
| Artificial Neural Network (ANN) | Accuracy $=\frac{\text { Correct Predictions }}{\text { Total predictions }}$ | An ANN uses layers of neurons to map inputs to outputs, optimizing weights by backpropagation with a differentiable loss function, such as cross-entropy | [60] |
| Naïve Bayes | Posterior Probability, $\mathrm{P}(\mathrm{y} \mid \mathrm{X})=$ $\mathrm{P}(\mathrm{X}) \mathrm{P}(\mathrm{X} \mid \mathrm{y}) \mathrm{P}(\mathrm{y})$ $\mathrm{P}(\mathrm{yi} \mid \mathrm{Xi})$ : Posterior probability of the i-th data point being in class yi | Naïve Bayes models eliminate complicated joint probability computations; they simplify calculations and are especially useful when the independence condition is roughly valid | [61] |
| Long Short-Term Memory (LSTM) | $\text { MSE }=\frac{1}{N} \sum_{i=1}^{N}(y i-\hat{y} i)^{2}$ <br> N : Total number of observations in the dataset. i: Index for individual data points. yi: True label. yíi: Predicted value | Temporal dependencies in sequential data are captured by LSTM, a recurrent neural network that utilizes gates (forget, input, and output) | [62] |
土壤水分这一关键方面,Stamenkovic 等人[86]和 Song 等人[87]利用机器学习算法可靠地预测了通过遥感检测的高光谱图像中的水分含量。这些模型,包括支持向量回归和基于深度学习的元胞自动机,表现出高性能,提供了精准灌溉调度的实用解决方案。这些研究共同展示了机器学习和深度学习技术如何革新农业土壤性质的监测与预测,实现更高效和可持续的耕作方式。
农业中的可变速率技术(VRT)利用基于传感器系统的实时数据,精确调整特定作物位置的施肥量。这一技术对于动态管理养分施用至关重要,能够根据土壤和植物状况的实时观察优化使用。Crop Circle 和 Green Seeker 等传感器发挥着关键作用,测量植物冠层在不同光谱波段的反射率,以计算植被指数,如归一化植被指数(NDVI)。NDVI 是最常用的指数,通过比较近红外和红光波段的反射率计算得出,有助于判断植物健康状况,从而确定所需的施肥剂量[136,137]。
Abbreviation Full Form
AI Artificial Intelligence
IoT Internet of Things
PA Precision Agriculture
GPS Global Positioning System
GIS Geographic Information Systems
NDVI Normalized Difference Vegetation Index
SAVI Soil-Adjusted Vegetation Index
VSSI Vegetation Soil Salinity Index
EVI Enhanced Vegetation Index
NLVI Non-Linear Vegetation Index
DVI Differential Vegetation Index
GRVI Green Ratio Vegetation Index
SI Salinity Index
ERSI Enhanced Residues Soil Salinity Index
CRSI Canopy Response Salinity Index
CI Clay Index
GI Gypsum Index
BI Brightness Index
NMDI Normalized Multi-Band Drought Index
L(lambda) Radiance
rho(lambda) Reflectance
USGS United States Geological Survey
NASA National Aeronautics and Space Administration
ESA European Space Agency
SWIR Shortwave Infrared
NIR Near-Infrared
R Red
B Blue
G Green
pi Pi (Mathematical Constant)
theta Solar Zenith Angle
NRBS Nitrogen-Rich Biosensor Spots
DL Deep Learning
ML Machine Learning
VRT Variable-Rate Technology
SVM Support Vector Machine
KNN K-Nearest Neighbors
DT Decision Tree
RF Random Forest
ANN Artificial Neural Network
MSE Mean Squared Error
LSTM Long Short-Term Memory
TP True Positives| Abbreviation | Full Form |
| :--- | :--- |
| AI | Artificial Intelligence |
| IoT | Internet of Things |
| PA | Precision Agriculture |
| GPS | Global Positioning System |
| GIS | Geographic Information Systems |
| NDVI | Normalized Difference Vegetation Index |
| SAVI | Soil-Adjusted Vegetation Index |
| VSSI | Vegetation Soil Salinity Index |
| EVI | Enhanced Vegetation Index |
| NLVI | Non-Linear Vegetation Index |
| DVI | Differential Vegetation Index |
| GRVI | Green Ratio Vegetation Index |
| SI | Salinity Index |
| ERSI | Enhanced Residues Soil Salinity Index |
| CRSI | Canopy Response Salinity Index |
| CI | Clay Index |
| GI | Gypsum Index |
| BI | Brightness Index |
| NMDI | Normalized Multi-Band Drought Index |
| $\mathrm{L}(\lambda)$ | Radiance |
| $\rho(\lambda)$ | Reflectance |
| USGS | United States Geological Survey |
| NASA | National Aeronautics and Space Administration |
| ESA | European Space Agency |
| SWIR | Shortwave Infrared |
| NIR | Near-Infrared |
| R | Red |
| B | Blue |
| G | Green |
| $\pi$ | Pi (Mathematical Constant) |
| $\theta$ | Solar Zenith Angle |
| NRBS | Nitrogen-Rich Biosensor Spots |
| DL | Deep Learning |
| ML | Machine Learning |
| VRT | Variable-Rate Technology |
| SVM | Support Vector Machine |
| KNN | K-Nearest Neighbors |
| DT | Decision Tree |
| RF | Random Forest |
| ANN | Artificial Neural Network |
| MSE | Mean Squared Error |
| LSTM | Long Short-Term Memory |
| TP | True Positives |
TN
真反例
FP
假正例
FN
假反例
T
随机森林中的树木数量
P(y∣X)\mathrm{P}(\mathrm{y} \mid \mathrm{X})
给定 X 的数据点 y 的后验概率
哨兵-2
欧洲航天局的地球观测卫星任务
Landsat-8
由 NASA 和美国地质调查局管理的遥感卫星
AS7341
农业遥感中常用的光谱传感器
PCM
作物预测模型
HRI
高分辨率影像
EL
集成学习
RS
遥感
ELM
极限学习机
SOC
土壤有机碳
便携式 X 射线荧光光谱仪
便携式 X 射线荧光
DNN
深度神经网络
UAV
无人机
RNN
循环神经网络
MODIS
中分辨率成像光谱仪
SNAP
哨兵应用平台
RMSE
均方根误差
R^(2)\mathrm{R}^{2}
决定系数
LASSO
最小绝对收缩和选择算子
GA
遗传算法
RGB
红、绿、蓝(用于图像的颜色模型)
SIF
太阳诱导叶绿素荧光
pH 值
氢离子浓度(酸碱度)
MLP
多层感知器
YOLO
你只看一次(一个实时目标检测模型家族)
RBF
径向基函数(支持向量机中使用的一种核函数)
PCA
主成分分析
VFH
视点特征直方图(用于物体识别和分类)
mAP
平均精度均值(用于评估物体检测模型)
F1 分数
模型准确度的一个衡量指标,结合了精确率和召回率
Inception-ResNet
结合了 Inception 和 ResNet 模型的混合深度学习架构
SLIC
简单线性迭代聚类(一种图像分割算法)
深度学习模型
深度学习模型
WSN
无线传感器网络
锂离子
锂离子
SPAD
土壤植物分析发展
VRA
变速率施用
NLP
自然语言处理
AI-ML
人工智能与机器学习
USDA
美国农业部
NASS
国家农业统计局
TN True Negatives
FP False Positives
FN False Negatives
T Number of trees in Random Forest
P(y∣X) Posterior probability of data point y given X
Sentinel-2 A satellite mission for Earth observation by the European Space Agency
Landsat-8 A satellite for remote sensing managed by NASA and USGS
AS7341 A spectral sensor commonly used in agricultural remote sensing
PCM Predictive Crop Modeling
HRI High-Resolution Imagery
EL Ensemble Learning
RS Remote Sensing
ELM Extreme Learning Machine
SOC Soil Organic Carbon
pXRF Portable X-ray Fluorescence
DNN Deep Neural Network
UAV Unmanned Aerial Vehicle
RNN Recurrent Neural Network
MODIS Moderate-Resolution Imaging Spectroradiometer
SNAP Sentinel Application Platform
RMSE Root Mean Square Error
R^(2) Coefficient of Determination
LASSO Least Absolute Shrinkage and Selection Operator
GA Genetic Algorithm
RGB Red, Green, Blue (color model used for images)
SIF Solar-Induced Chlorophyll Fluorescence
pH Potential of Hydrogen
MLP Multi-Layer Perceptron
YOLO You Only Look Once (a family of real-time object detection models)
RBF Radial Basis Function (a kernel function used in SVM)
PCA Principal Component Analysis
VFH Viewpoint Feature Histogram (used in object recognition and classification)
mAP Mean Average Precision (used to evaluate object detection models)
F1 Score A measure of a model's accuracy, combining precision and recall
Inception-ResNet A hybrid deep learning architecture combining Inception and ResNet models
SLIC Simple Linear Iterative Clustering (an algorithm for image segmentation)
DL Models Deep Learning Models
WSN Wireless Sensor Networks
Li-ion Lithium Ion
SPAD Soil Plant Analysis Development
VRA Variable Rate Application
NLP Natural Language Processing
AI-ML Artificial Intelligence and Machine Learning
USDA United States Department of Agriculture
NASS National Agricultural Statistics Service| TN | True Negatives |
| :--- | :--- |
| FP | False Positives |
| FN | False Negatives |
| T | Number of trees in Random Forest |
| $\mathrm{P}(\mathrm{y} \mid \mathrm{X})$ | Posterior probability of data point y given X |
| Sentinel-2 | A satellite mission for Earth observation by the European Space Agency |
| Landsat-8 | A satellite for remote sensing managed by NASA and USGS |
| AS7341 | A spectral sensor commonly used in agricultural remote sensing |
| PCM | Predictive Crop Modeling |
| HRI | High-Resolution Imagery |
| EL | Ensemble Learning |
| RS | Remote Sensing |
| ELM | Extreme Learning Machine |
| SOC | Soil Organic Carbon |
| pXRF | Portable X-ray Fluorescence |
| DNN | Deep Neural Network |
| UAV | Unmanned Aerial Vehicle |
| RNN | Recurrent Neural Network |
| MODIS | Moderate-Resolution Imaging Spectroradiometer |
| SNAP | Sentinel Application Platform |
| RMSE | Root Mean Square Error |
| $\mathrm{R}^{2}$ | Coefficient of Determination |
| LASSO | Least Absolute Shrinkage and Selection Operator |
| GA | Genetic Algorithm |
| RGB | Red, Green, Blue (color model used for images) |
| SIF | Solar-Induced Chlorophyll Fluorescence |
| pH | Potential of Hydrogen |
| MLP | Multi-Layer Perceptron |
| YOLO | You Only Look Once (a family of real-time object detection models) |
| RBF | Radial Basis Function (a kernel function used in SVM) |
| PCA | Principal Component Analysis |
| VFH | Viewpoint Feature Histogram (used in object recognition and classification) |
| mAP | Mean Average Precision (used to evaluate object detection models) |
| F1 Score | A measure of a model's accuracy, combining precision and recall |
| Inception-ResNet | A hybrid deep learning architecture combining Inception and ResNet models |
| SLIC | Simple Linear Iterative Clustering (an algorithm for image segmentation) |
| DL Models | Deep Learning Models |
| WSN | Wireless Sensor Networks |
| Li-ion | Lithium Ion |
| SPAD | Soil Plant Analysis Development |
| VRA | Variable Rate Application |
| NLP | Natural Language Processing |
| AI-ML | Artificial Intelligence and Machine Learning |
| USDA | United States Department of Agriculture |
| NASS | National Agricultural Statistics Service |
参考文献
Lal, R. 在 5 亿公顷谷物作物面积上养活 110 亿人。食品能源安全。2016,5,239-251。