0 Introduction
1 Materials and data
1.1 Study setup
Table 1 Comparison on fertilization methods of different tomato groups |
KNO3/g | NH4H2PO4/g | KH2PO4/g | K2SO4/g | MgSO4·7H2O/g | CaCl2/g | CaNO3·4H2O/g | |
---|---|---|---|---|---|---|---|
CK | 0.0 | 0.0 | 27.2 | 52.3 | 24.6 | 44.4 | 0.0 |
T1 | 13.4 | 7.7 | 0.0 | 34.8 | 24.6 | 44.4 | 0.0 |
T2 | 26.9 | 7.7 | 0.0 | 17.4 | 24.6 | 44.4 | 0.0 |
T3 | 40.4 | 7.7 | 0.0 | 0.0 | 24.6 | 44.4 | 0.0 |
T4 | 40.4 | 7.7 | 0.0 | 0.0 | 24.6 | 29.6 | 11.8 |
T5 | 40.4 | 7.7 | 0.0 | 0.0 | 24.6 | 14.8 | 23.6 |
T6 | 40.4 | 7.7 | 0.0 | 0.0 | 24.6 | 0.0 | 35.4 |
Table 2 Micronutrient application of tomato |
Molecular | Molecular weight standard concentration/(mg/L) |
---|---|
FeSO4·7H2O | 13.206 |
EDTA·H2O | 17.679 |
EDTA·3H2O | 20.000 |
MnSO4·H2O | 1.614 |
MnSO4· 4H2O | 2.130 |
H3BO3 | 2.860 |
ZnSO4·7H2O | 0.220 |
CuSO4·5H2O | 0.080 |
Na2MO4·2H2O | 0.020 |
1.2 Data collection
1.2.1 Automatic data collection
1.2.2 Manual data collection
2 Methods
2.1 Data preprocessing
2.1.1 Handling missing data
2.1.2 Data imputation using interpolation
2.2 The training networks and evaluation metrics
Fig. 1 Architecture of simple CNN |
Fig. 2 Architecture of RNN and LSTM |
2.3 Research framework
Fig. 3 Process flow of tomato growth height prediction method by phenotypic feature extraction using multi-modal data study |
2.4 Experimental setup
Table 3 Parameters for CNN operation of extract image features |
Network Structure | Parameters |
---|---|
Convolutional kernel | 3×3 |
Convolutional stride | 1 |
Convolutional padding | 1 |
Pooling kernel | 2×2 |
Pooling stride | 2 |
Fully connected input | 16×300×300 |
Fully connected output | 16 |
3 Results and analysis
3.1 RNN temporal prediction and model hyperparameter selection
Table 4 Parameters for RNN operation of tomato height prediction |
Network structure | Parameters |
---|---|
Input features | 34 |
RNN hidden size | 30 |
RNN layers | 2 |
Learning rate | 0.001 |
Loss function | MSELoss |
Optimizer | Adam |
|
Table 5 Results of multi-modal prediction using RNN under different hyperparameters in tomato height prediction study |
RNN network layers | Hidden layer size | Prediction scenario | |||||
---|---|---|---|---|---|---|---|
Short-term prediction | Mid-term prediction | Long-term prediction | |||||
MSE | MAPE/% | MSE | MAPE/% | MSE | MAPE/% | ||
2 | 15 | 6.89 | 2.16 | 55.21 | 6.03 | 113.17 | 6.74 |
20 | 1.89 | 1.25 | 68.13 | 6.87 | 288.93 | 10.81 | |
25 | 2.11 | 1.11 | 56.51 | 5.41 | 297.91 | 10.85 | |
30 | 3.09 | 1.87 | 56.78 | 4.83 | 236.51 | 10.12 | |
3 | 15 | 6.03 | 2.01 | 21.09 | 2.82 | 388.68 | 11.26 |
20 | 1.52 | 1.07 | 27.53 | 4.11 | 180.16 | 7.14 | |
25 | 3.07 | 1.74 | 46.15 | 6.36 | 165.09 | 6.51 | |
30 | 2.49 | 1.50 | 60.96 | 6.71 | 203.26 | 8.26 |
Table 6 Results of mono-modal prediction using RNN under different hyperparameters in tomato height prediction study |
RNN network layers | Hidden layer size | Prediction scenario | |||||
---|---|---|---|---|---|---|---|
Short-term prediction | Mid-term prediction | Long-term prediction | |||||
MSE | MAPE/% | MSE | MAPE/% | MSE | MAPE/% | ||
2 | 15 | 10.26 | 2.61 | 25.04 | 3.18 | 386.39 | 10.01 |
20 | 10.82 | 2.62 | 126.04 | 7.70 | 354.20 | 8.61 | |
25 | 10.38 | 2.5 | 30.03 | 3.88 | 546.32 | 11.57 | |
30 | 12.99 | 2.36 | 18.98 | 2.99 | 425.85 | 11.87 | |
3 | 15 | 6.83 | 1.64 | 37.67 | 3.79 | 405.23 | 10.71 |
20 | 8.14 | 2.13 | 1 532.62 | 22.76 | 1 140.67 | 17.18 | |
25 | 31.33 | 5.41 | 144.03 | 8.28 | 432.18 | 10.94 | |
30 | 14.63 | 2.95 | 16.78 | 2.27 | 672.64 | 13.62 |
|
Fig. 4 Variation of loss with training epochs under different prediction horizons in tomato height prediction study |
Table 7 The optimal predictive performance of PFE-RNN in tomato height prediction study |
Prediction scenario | Average accuracy/% | Last day accuracy/% | Average percentage error/% | Last day average percentage error/% | R 2 |
---|---|---|---|---|---|
Short-term prediction | 91.67 | 91.67 | 0.81 | 0.83 | 0.999 678 |
Mid-term prediction | 88.54 | 87.50 | 2.66 | 3.11 | 0.991 119 |
Long-term prediction | 70.37 | 66.67 | 14.05 | 16.63 | -0.041 393 |
Fig. 5 Distribution of predicted values versus actual values for PFE-RNN prediction in tomato height values |
3.2 LSTM temporal prediction and model hyperparameter selection
Table 8 Parameters for LSTM operation in tomato height prediction |
Network structure | Parameters |
---|---|
Input features | 34 |
LSTM hidden size | 30 |
LSTM layers | 2 |
Learning rate | 0.001 |
Loss function | MSELoss |
Optimizer | Adam |
Fig. 6 Variation of loss with training epochs under different prediction horizons in tomato height prediction study |
Table 9 Results of multi-modal prediction using LSTM under different hyperparameters in tomato height prediction study |
LSTM network layers | Hidden layer size | Prediction scenario | |||||
---|---|---|---|---|---|---|---|
Short-term prediction | Mmid-term prediction | Long-term prediction | |||||
MSE | MAPE/% | MSE | MAPE/% | MSE | MAPE/% | ||
2 | 15 | 1.48 | 0.96 | 121.90 | 8.44 | 153.98 | 7.61 |
20 | 3.27 | 1.33 | 104.02 | 7.99 | 204.30 | 8.82 | |
25 | 0.57 | 0.88 | 88.01 | 6.63 | 254.83 | 9.67 | |
30 | 3.85 | 2.24 | 119.26 | 8.42 | 188.33 | 8.56 | |
3 | 15 | 0.21 | 0.54 | 81.72 | 7.60 | 547.96 | 15.03 |
20 | 1.56 | 1.44 | 105.02 | 8.89 | 286.87 | 10.56 | |
25 | 0.91 | 0.99 | 87.00 | 7.09 | 286.13 | 8.62 | |
30 | 0.34 | 0.73 | 98.91 | 7.16 | 172.19 | 6.54 |
Table 10 Results of mono-modal prediction using LSTM under different hyperparameters in tomato height prediction study |
LSTM network layers | Hidden layer size | Prediction scenario | |||||
---|---|---|---|---|---|---|---|
Short-term prediction | Mid-term prediction | Long-term prediction | |||||
MSE | MAPE/% | MSE | MAPE/% | MSE | MAPE/% | ||
2 | 15 | 6.68 | 1.83 | 31.11 | 3.39 | 802.94 | 15.17 |
20 | 2.99 | 1.40 | 18.01 | 2.56 | 378.78 | 9.44 | |
25 | 6.59 | 2.22 | 39.56 | 4.25 | 486.95 | 11.55 | |
30 | 5.15 | 1.52 | 30.60 | 3.43 | 358.90 | 9.49 | |
3 | 15 | 3.79 | 1.37 | 37.67 | 3.79 | 870.38 | 15.67 |
20 | 3.89 | 1.39 | 21.53 | 2.84 | 507.24 | 10.95 | |
25 | 3.41 | 1.40 | 30.07 | 3.58 | 485.04 | 11.79 | |
30 | 7.01 | 1.74 | 153.36 | 7.81 | 385.27 | 8.66 |
Table 11 Optimal predictive performance of PFE-LSTM model in tomato height prediction study |
Prediction scenario | Average accuracy/% | Last day accuracy/% | Average percentage error/% | Last day average percentage error/% | R 2 |
---|---|---|---|---|---|
Short-term prediction | 97.92 | 97.92 | 0.40 | 0.38 | 0.999 901 |
Mid-term prediction | 53.13 | 41.67 | 6.36 | 7.72 | 0.965 263 |
Long-term prediction | 50.93 | 33.33 | 14.49 | 19.48 | -0.277 455 |
Fig. 7 Distribution of predicted values versus actual values for PFE-LSTM model |
3.3 Comparative analysis results with other methods
3.3.1 Performance evaluation of LLM
Fig. 8 Distribution of predicted values versus actual values for LLM prediction in tomato height |
3.3.2 Performance evaluation of Transformer-based model
Fig. 9 Distribution of predicted values versus actual values for transformer prediction of tomato height |
Table 12 Comparison of results from different methods of tomato height prediction study |
Model | Prediction scenario | MAPE/% | MSE | Last day MAPE/% |
---|---|---|---|---|
PFE-RNN | Short-term prediction | 0.81 | 0.56 | 0.83 |
Mid-term prediction | 2.66 | 17.04 | 3.11 | |
Long-term prediction | 14.05 | 651.35 | 16.63 | |
PFE-LSTM | Short-term prediction | 0.40 | 0.12 | 0.38 |
Mid-term prediction | 6.36 | 82.54 | 7.72 | |
Long-term prediction | 14.49 | 696.22 | 19.48 | |
Transformer | Short-term prediction | 6.72 | 74.71 | 6.33 |
Mid-term prediction | 10.29 | 144.62 | 9.65 | |
LLM | Short-term prediction | 8.00 | 588.26 | 7.90 |
Mid-term prediction | 27.43 | 4 067.70 | 28.93 |