A ML-AI ENABLED ENSEMBLE MODEL FOR PREDICTING AGRICULTURAL YIELD

Figures & data

Table 1. Mostly used machine learning algorithms

Most used machine learning algorithms	No. of times used
Neural Networks	27
Linear Regression	14
Random Forest	12
Support Vector Machine (SVM)	10
Gradient Boosting Tree	4

Source: Klompenburga et al. (2020).

Figure 1. The nomenclature for the proposed “ensemble model”.

Figure 2. Procedure followed for data analysis.

Table 2. Independent variables used in the model

Variables/ dimensions	Different elements of the variables/ dimensions used in the model	Variables which evolved as statistically significant
Weather	Water level (gallons of water per square feet), humidity (average humidity in percentage), rainfall (average rainfall in mm), temperature (maximum temperature, minimum temperature, and average temperature-degree Celsius). The average maximum wind speed at the district (Max_WS_Avg).	The average maximum wind speed (Max_WS_Avg), Average rainfall and average temperature
Soil	Alluvial soil, black soil, loamy soil, red soil, acidic, alkaline, Amount of Boron content in the field (B_NA), Amount of chlorine content in the field (LT_CL), Whether the iodine content was more than the critical level (I Above Critical Level), Level of electrical conductivity in the seeds (EC_ Critical for germination), No of fields having the prevalence of red soil as soil characteristics in the district (ST_Red Soil)	Amount of Boron content in the field (B_NA), Amount of chlorine content in the field (LT_CL), Whether the iodine content was more than the critical level (I_Above Critical Level), Level of electrical conductivity in the seeds (EC_ Critical for germination), No of fields having the prevalence of red soil as soil characteristics in the district (ST_Red Soil)
Farmer	Number of farm labourers, number of farm owners, number of tenants, number of marginal farmers, number of small farmers., No of farms who cultivate a specific crop in the district (Farms_Cropwise), No of firms in the district having the prevalence of modern irrigation facilities including air pumps, etc. (IR_ Lift Irrigation)	No of farms who cultivate a specific crop in the district (Farms_Cropwise), No of firms in the district having the prevalence of modern irrigation facilities including air pumps, etc. (IR_ Lift Irrigation)
Irrigation	Type of irrigation for the field—Dry, Rainfed, Borewell, Canal, Tank, Open well, etc (IR_Irrigated)	Type of irrigation for the field—Dry, Rainfed, Borewell, Canal, Tank, Open well, etc (IR_Irrigated)
Fertiliser	Normal fertiliser, fertiliser injurious to crop, high-organic fertiliser, low-organic fertiliser, high potassium fertiliser, high nitrogen fertiliser. Amount of Propyne oxide contaminated in the field categorized into Low, Medium, and High based on the distribution (PO)	Amount of Propyne oxide contaminated in the field categorized into Low, Medium, and High based on the distribution (PO)

Table 3. Level 1 validation: model performance statistics for all the different models on the validation data

Category	Models	Measurement parameters
		ROC or AUC	Accuracy	Precision	Recall	F Score	Champion Model
Equation based models	Full fit regression	0.483	0.503	0.523	0.576	0.574	Forward regression
	Backward regression	0.492	0.517	0.516	0.504	0.502
	Forward regression	0.497	0.527	0.575	0.598	0.597
	Stepwise regression	0.467	0.502	0.517	0.557	0.562
Algorithm/ Tree Based Models	Decision Tree	0.914	0.907	0.907	0.907	0.907	Random Forest
	Random Forest	0.938	0.900	0.902	0.900	0.900
	XG Boost	0.860	0.860	0.863	0.860	0.860
	Stochastic Gradient Descent	0.933	0.933	0.938	0.933	0.933
	Gradient Boosting	0.500	0.923	0.852	0.923	0.886
Self-Learning / Machine Learning Based Models	Neural Network	0.487	0.517	0.565	0.588	0.587	Neural Network from Champion Tree
	Neural Network from Champion Regression	0.683	0.506	0.505	0.506	0.497
	Neural Network from Champion Tree	0.960	0.940	0.940	0.940	0.940
	Support Vector Machines (SVM)	0.955	0.900	0.900	0.900	0.900
Models based on other modern approaches	CN2 Rule Inducer (Learning Algorithm for Rule Induction)	0.923	0.887	0.887	0.887	0.887	CN2 Rule Inducer (Learning Algorithm for Rule Induction)
	K Nearest Neighbour (KNN)	0.727	0.673	0.675	0.673	0.673
	Naïve Bayes	0.850	0.908	0.909	0.908	0.908

Table 4. Performance of ensemble model (average aggregation of individual probabilities)

Category	Aggregation function of the probabilities used for Ensembling	Measurement parameters
		ROC or AUC	Accuracy	Precision	Recall	F Score	Champion model
Ensemble Model	Averaging	0.985	0.953	0.955	0.953	0.953	Ensemble model with the average aggregation

Table 5. Performance of different ensemble models on the validation data

Category	Aggregation function of the probabilities used for Ensembling	Measurement parameters
		ROC or AUC	Accuracy	Precision	Recall	F Score	Champion Model
Ensemble Model	Averaging	0.985	0.953	0.955	0.953	0.953	Ensemble Model with Average Aggregation
	Maximum Probability	0.959	0.869	0.866	0.869	0.867
	Voting Option (Averaging the posterior probabilities of the most popular class)	0.932	0.852	0.855	0.857	0.861
	Voting Option (Recalculate the posterior probability of each class by using the proportion of posterior probabilities that predict that class)	0.912	0.849	0.852	0.849	0.852

Table 6. Optimized cut-off for the ensemble model

Cut Off	Sensitivity	Specificity	Sum
0.01	69.70	100	169.70
0.10	90.60	93.20	183.80
0.20	92.30	93.20	185.50
0.30	93.10	93.60	186.70
0.40	94.10	94.30	188.40
0.50	95.30	95.50	190.80
0.60	95.30	76.50	171.80
0.70	92.30	76.50	168.80
0.80	87.80	75.30	163.10
0.90	100	33.30	133.30
0.99	100	0	100.00

Figure 3. Snapshot of results obtained.

Figure 4. Lift curve analysis of the ensemble model (average probability aggregation).

Table 7. Lift analysis of ensemble model (average probability aggregation): 2nd level of model validation

Decile	Lift	Interpretation
10	2	The Ensemble model is able to classify twice the events (High Yield, Medium Yield and Low Yield) compared to a no model or a baseline model.
20	1.90	The Ensemble model is able to classify 1.9 times the events (High Yield, Medium Yield and Low Yield) compared to a no model or a baseline model.
40	1.86	The Ensemble model is able to classify 1.86 times the events (High Yield, Medium Yield and Low Yield) compared to a no model or a baseline model.
60	1.64	The Ensemble model is able to classify 1.64 times the events (High Yield, Medium Yield and Low Yield) compared to a no model or a baseline model.
80	1.36	The Ensemble model is able to classify 1.36 times the events (High Yield, Medium Yield and Low Yield) compared to a no model or a baseline model.

Table 8. Cumulative gain analysis of ensemble model: 2nd level of model validation

Decile	Cumulative gain	Interpretation
20	0.40	By targeting top 20% of the observations in the validation data based on the predicted probability (Category—High Yield) we will be able to focus on 40% of the high yield cases.
40	0.70	By targeting top 40% of the observations in the validation data based on the predicted probability (Category—High Yield) we will be able to focus on 70% of the high yield cases.
60	0.95	By targeting top 60% of the observations in the validation data based on the predicted probability (Category—High Yield) we will be able to focus on 95 % of the high yield cases.

Figure 5. Lift curve analysis of the ensemble model (average probability aggregation) compared with other ML models.

Figure 6. Cumulative gain analysis of the ensemble model.

Figure 7. Cumulative gain analysis of the ensemble model compared with other ML models: 2nd level of model validation.

Figure 8. 10-fold model cross validation: 3^rd level of model validation.

Figure 9. Cumulative Gain of Ensemble Model

Thomas van Klompenburg, Ayalew Kassahun, Cagatay Catal. et al (2020). Crop yield prediction using machine learning: A systematic literature review. Computer and Electronics in Agriculture, 177, 105709. https://doi.org/10.1016/j.compag.2020.105709 https://www.sciencedirect.com/science/article/pii/S0168169920302301. ISSN 0168-1699

 Google Scholar