Since independence, the primary contributor to the country’s GDP has been agriculture. Agronomy and their associated industries constituted 59% of the nation’s overall GDP during the fiscal year 1950-1951 [1]. Despite the relatively low agricultural productivity, agriculture remains one of the most prominent economic sectors in India. Precision agriculture is one method we may utilize to boost productivity. Applying precise and suitable amounts of soil, fertilizers, and other elements is what precision farming entails, as the term implies. Globalization has, however, caused a significant shift in the agricultural trend in recent years. Numerous factors have adversely affected India’s agricultural sector. In order to restore crop vitality numerous innovative technologies have been developed. Precision farming is one such method, applied at the right time to the crop to boost yields and production. Precision agriculture refers to farming technology that is site-specific. But in agriculture, it’s important that the guidance provided be precise and correct.
Machine learning, as defined by Arthur Samuel in 1959, is the study of how computers learn without being explicitly programmed. ML algorithms are trained on vast volumes of facts to generate expectations or findings. Recent years have seen a surge in crop prediction research. For example, using loT and machine learning (ML) technology to improve agricultural decision-making [2]. Another research proposes employing neural networks to create a robust, precise, and clear recommendation system [3].
To anticipate the most productive crop, this work proposes a crop suggestion method that makes source of machine learning algorithms to evaluate soil (pH, phosphorus, nitrogen, and potassium) and meteorology (temperature, moisture, and rainfall) data. The F1 score, Recall, and Precision have been utilized to assess the performance of the approach suggested for every class and method. Crop recommendation systems can assist farmers in selecting which crops to plant, increasing yields and reducing resource consumption. Crop suggestion systems can also increase agriculture’s ability to adapt to climate change. The remaining part of this work includes: The literature review and details on the model are provided in Sections 2 and 3. Sections 4 and 5 cover the Experimental Setup.
Including important environmental variables improves the dataset which is used in [4]. The dataset contains Temperature, Humidity, pH, rainfall and label which includes sugarcane, coconut, jute, cotton, papaya, groundnut, maize, graphs, rice, mango, rubber etc. It uses an SVM decision tree (Hybrid approach) which maintains an accuracy rate of 91.8% and Random Forest shows an accuracy of 95%. The Table 1 itself serves as the literature review summary.
Crop Recommendation Techniques ML
| Authors | Methodology | Features | Accuracy | Dataset | Advantages | Limitations |
|---|---|---|---|---|---|---|
| Srilakshmi A., Madhumitha K., Geetha K[4] | SVM decision tree (Hybrid approach), Random Forest | Temperature, Humidity, pH, rainfall, label | SVM decision tree (Hybrid approach)- 91.8%. | sugarcane, coconut, jute, cotton, papaya, groundnut, maize, graphs, rice, mango, rubber etc | Predict crop for any type of field | Small dataset |
| R. Pallavi Reddy, B. Vinitha, K. Rishita, K. Pranavi [2020] [2] | Linear Regression Model | N, P, K, and moisture values | generate recommendations to improve crop production and estimates the price of the yield | Limited in capturing non-linear patterns, Assumes homoscedasticity and independence of errors | ||
| S. Mamatha Jajur, Soumya N. G. [2019] [5] | KNN, Decision trees, SVM, CNN and LSTM, ANNs, K-means clustering | Soil Type, pH value, NPK content of the soil, Water holding, Temperature, Average rainfall, Previously Harvested crop | - | wheat, rice, bajra, maize, jawar, | select the optimum crop while keeping a number of variables in mind to boost the output of agriculture, minimise the deterioration of the soil in fields that are under cultivation and use less fertiliser when growing crops. | Many algorithms are used |
| Mr. Santosh Mahagaonkar, Devdatta A. Bondre [2019] [6] | Random Forest, Support Vector Machine algorithm | crop, crop yield dataset, Location, soil and crop nutrients, fertilizer datasets | soil classification, RF-86.35% crop yield prediction SVM -99.47% | Soybean, Rice, Jowar, Wheat, Sunflower, Cotton, Sugarcane, Tobacco, Onion, Dry Chili, etc. | future prediction of crop yield | Low accuracy in soil classification performance heavily depends on parameter tuning and it is memory intensive, particularly for large datasets |
| D. Anantha Reddy, Bhagyashri Dadore, Aarti Watekar [2019] [7] | Naïve Bayes, K-NEAREST NEIGHBOUR, RANDOM FOREST, CHAID | Depth, Texture, pH, Soil Colour, Permeability, Drainage, Water holding and Erosion | - | groundnut, pulses, cotton, vegetables, paddy, sugarcane, coriander. | Assist farmers in planting the appropriate seed according to the needs of the soil in order to boost output. | The Naïve Bayes algorithm pretends feature independence, which might not be true when dealing with real-world data., CHAID - Limited to categorical target variables and predictors, making it less versatile for handling continuous data |
| Nidhi H. Kulkarni [8] 2018 | Linear SVM algorithms, Random Forest, Naïve Bayes | Soil type, pH soil, NPK, average rainfall, porosity of soil, sowing season temperature | 99.9 | Cotton, Sugarcane, Rice, Wheat | Crop productivity has improved exponentially for rice, wheat, cotton, and sugarcane. | restricted to a fairly small number of crops |
| Zeel Doshi [3] 2018 | Neural Network Random Forest, Decision Tree, KNN | Temperature rainfall, Location, soil condition | 91% | Jute, sesame, soybean, sugarcane, tobacco, sunflower seeds, ragi, potato, tur, grapeseed, and mustard, bajra, maize wheat, rice gram, barley, cotton, groundnut, and pulses | Neural Networks have the highest accuracy percentage. | predict the crop using the harvest from the previous cycle. Crop supply and demand are not considered |
| Rohit Kumar Rajak [9] 2017 | Random Tree, NB-classifier, ANN, SVM | depth, pH, texture, permeability to store water, color ofthe soil, and drainage from erosion | - | vegetables, rice, sugarcane, sorghum, coriander, bananas, legumes, and groundnuts | boosts agricultural productivity | larger dataset for model training |
| S. Pudumalar [10] 2016 | Random Tree, Naïve Bayes, KNN, CHAID, | Depth, pH, texture, waterholding permeability, Soil color, erosion drainage, | 88% | millet, pulses, groundnut, cotton, banana, vegetables, paddy, sugarcane, sorghum, coriander | Boost productivity | larger dataset for model training |
| Rakesh Kumar [11] 2015 | CSM, Gradient Boosted Decision Tree, and Greedy Forest | soil type, weather, crop type, water density, | ratoi, toria, wheat, potato, sarso, linseed, masoor, khesari, onion, sugarcane, Kanda, mung, til, pumpkin, nenua, ladies’ finger, rice, soybean, sweet potato, toor, vegetable seed, and so on | offers a method to select crops while taking into account the yield forecast rate influenced by various factors. | Adopting a prediction technique that performs well and has greater accuracy is necessary |
The methodology, which integrates machine learning with the IoT, is not well recognized. The authors of [2] proposed it as the Crop Monitoring and Recommendation System utilizing sensors to record certain characteristics of the soil, such as its moisture content and nutrients, and uploading the data to a cloud platform. An Android app receives this data and gives recommendations for crop selection based on soil type among other factors. Furthermore, a price prediction module has been integrated using linear regression. This combined approach is expected to help farmers make good choices and increase farm productivity and profitability aimed at taking into consideration soil health worries.
Authors in [5] have used data including soil type, acidity level, NPK content, permeability, water holding capacity, average rainfall, temperature, as well as previously grown crops. For classification tasks they have supervised learning methods KNN, ensemble learning (EL), and SVM and also unsupervised learning methods (K-means clustering for data analysis).
A technique that farmers throughout India can simply employ is the intelligent crop recommendation system. Three processes are involved in this research: the classification of the soils, crop yield prediction, and fertilizer suggestion utilizing Random Forest and Support Vector Machine, which provide more robust models than traditional logistic regression model as shown in Figure 1. Additionally, the system has apps from third parties that provide weather data. The result of this experiment reveals that soil classification using Random Forests and crop yield prediction with Support Vector Machines are effective. The future is to improve it by including development of a mobile application for farmer as well as implementing crop disease detection through image processing [6].
Logistic function [12]
SVM [5]
The authors in [7] suggested a method so that they can help farmers to make decisions about crops depending on their soil types. It applies soil-specific characteristics such as depth, texture, pH, and waterholding capacity to recommend appropriate crops. It makes use of a group method that incorporates Random Tree, CHAID, Naïve Bayes and KNN machine learning algorithms as demostrated in Figure 3 and 5. This study demonstrates the use of Random Forests for soil categorization and the use of Vector Machines for crop production prediction.
Decision Tree [3]
Random Forest [3]
Four crops have been taken into consideration in a crop recommendation system: wheat, cotton, sugarcane, and rice in [8]. Accurate crop selection is provided by crop recommendation systems, which take soil, surface temperature, and rainfall into account. The proposed model uses a high degree of efficiency and accuracy of ensembling approach to predict the crop that will boost yield.
The authors of this work [3] have introduced AgroConsultant, an intelligent system that Indian farmers may use to make well-informed decisions on which crops to produce by utilizing data on soil properties, geographic location, and climatic parameters like rainfall and temperature. This is accomplished by the utilization of various ML methods, like neural networks, KNNs, Random Forests, and decision trees.
The authors [9] have concluded that crop yield prediction is essential for the country’s planned guiding principles made in the field of agriculture development, to provide greater agricultural output and effective use of water resources while assisting farmers in minimising the use of pesticides in crop production and preventing soil deterioration.
Based on the needs of the soil, data mining tools would assist farmers in choosing the best seeds to plant, guaranteeing higher yield to make a profit. To precisely and effectively recommend a crop based on site-specific data, a collaborative recommendation model is built using techniques including the algorithms Random Tree, CHAID, KNN, and Naïve Bayes [10].
To grow the crop’s net yield rate, the authors of [11] recommend the Crop Selection Method (CSM), which recommends the series of crops that will be sown throughout the season depending on the crop yield forecast. A solution for crop selection based on factors such as crop type, weather, soil type, and water density is offered by the proposed method. This approach suggests a crop sequence’s daily production is maximal for a given season, considering the crop, sowing time, number of days in planting, and season-wise yield rate as inputs.
Machine learning is one of several methods to predict crop production in agriculture. Machine learning techniques such as Random Forests and Support Vector Regression are used, which provide more robust models than traditional linear regression models [12].
One technique used for binary classification is called Logistic Regression. In this technique we predict the probability of an input example falling into one of two classes. Therefore, the output should be in discrete value i.e., either Yes or No, true or false, etc. It is classified into three basic categories: Binomial, Multinomial and Ordinal.
x = input value
y = predicted output
a0 = bias or intercept term
a1 = coefficient for input (x)
One of the most fundamental and successful probabilistic classification methods, Naïve Bayes, is based on the Bayes’ theorem. It generates a likelihood table by finding probabilities of the features. Gaussian Naïve Bayes is specifically applied when continuous features follow a Gaussian distribution. It’s efficient, simple, and performs well in limited training data.
Eq.(2) is Bayes’ theorem in which:
P (A) = The probability of A occurring
P (B) = The probability of B occurring
P (A|B) = The probability of A given B
P (B|A) = The probability of B given A
Support Vector Machine (SVM) is a supervised machine learning technique that may be applied to both regression and classification applications. The objective of this method is to determine the hyperplane that effectively divides the two classes with the widest possible margin. The SVM algorithm utilizes two crucial elements to choose the most suitable hyperplane for classifying labels: two measurements: the support vectors, or the data points closest to the hyperplane, and the margin, or the distance between the hyperplane and the closest data points. Figure 2 highlights the SVM plan and support vectors.
Eq.(3) equation of hyperplane in which:
w = a vector normal to hyperplane
b = an offset
KNN, K-Nearest Neighbours is a supervised ML method, is usually used for division but also for regression. This method starts with selecting the nearest Neighbors. Then, on can try different values for K to find the optimal one. The Euclidean distance between K Neighbors is also calculated. We count the number of data points in each category between these K Neighbors, and we assign a new data point to the category with the highest number of Neighbors. Expanding the training data collection could enhance this technique.
Eq. (4) is Euclidean distance formula in which:
The coordinates of one point are (x1, y1)
The coordinates of the other point are (x2, y2)
Distance between (x1, y1) and (x2, y2) is d.
Although decision trees are among the most powerful tools available, they are typically employed for classification jobs. They can also be utilized for regression assignments. As shown in Figure 4, the Decision Node and Leaf Node are the two nodes that make up this system. Leaf Nodes represent the decision’s output, and Decision Nodes are used to make decisions. Decision trees are easy to understand, treatment of both numerical and categorical data though their failure to prune could lead them into over-fit situations especially when noisy datasets are considered.
Among the widely used ensemble models based on decision trees is Random Forest. Every observation is fed into one of the many decision trees that are generated in Random Forest. Random Forest can deal with high-dimensional data, maintain some level of interpretability, and has less chance of being affected by overfitting than individual decision trees themselves [13].
We aim to develop a CR System by using ML techniques to help farmers select which crop should yield based on soil and climate parameters. The dataset used contains nitrogen (N), phosphorus (P), potassium (K) levels, temperature, humidity, pH, and amount of rainfall, along with the label of crops suitable for the environment and to get a better understanding of the dataset we have visualized the correlation between label and other features.
To ensure data integrity, exploratory data analysis identifies the dataset’s dimensions, contents, missing values, and duplicates. Furthermore for smooth flow of data, crop labels were mapped into numerical identifiers. Each crop is associated with a unique numerical value, ranging from 1 to 22. A new column ‘crop_num’ is created corresponding to existing column ‘label’ containing crop names.
The dataset is divided into features (X) containing N, P, K, temperature, humidity, pH and rainfall and labels (y) containing newly created column ‘crop_num’. To standardize and normalize the data for model compatibility, we have used Min-Max Scaler. So that it scales and translates each feature individually using the function MinMaxScaler().
The Figure 7 above demonstrates Various Machine learning algorithms were used for the development of the model. Each model’s accuracy in crop prediction was evaluate using training dataset (X_train, y_train) after that, the suitable model was trained using a training dataset (X_train, y_train). Naïve Bayes has shown to have the highest accuracy among these models.
(a). Relationship between Nitrogen Levels and Crop Yield
(b). Relationship between Potassium Levels and Crop Yield
(c). Relationship between Phosphorus Levels and Crop Yield
(d). Relationship between Temperature and Crop Yield
(e). Relationship between Humidity and Crop Yield
(f). Relationship between pH and Crop Yield
(g). Relationship between Rainfall and Crop Yield
For the most accurate crop recommendation as shown in the Figure 8, a Gaussian Naïve Bayes model is trained on the entire dataset. A recommendation function is developed, which will allow users to input soil and climatic parameters such as nitrogen, phosphorus, potassium levels, temperature, humidity, pH, and rainfall to predict the most suitable crop for cultivation then a trained model is used, and the function simply predicts while mapping the predicted crop number against a predefined dictionary storing its corresponding crop name [14] as displayed in Table 2.
Block Diagram of Crop Recommendation System
Accuracy Comparison
Crop Label and Corresponding Numerical Representation
| label | Crop_num |
|---|---|
| Rice | 1 |
| Maize | 2 |
| Jute | 3 |
| Cotton | 4 |
| Coconut | 5 |
| Papaya | 6 |
| Orange | 7 |
| Apple | 8 |
| Muskmelon | 9 |
| Watermelon | 10 |
| Grapes | 11 |
| Mango | 12 |
| Banana | 13 |
| Pomegranate | 14 |
| Lentil | 15 |
| Blackgram | 16 |
| Mungbean | 17 |
| Mothbeans | 18 |
| Pigeonpeas | 19 |
| Kidneybeans | 20 |
| Chickpea | 21 |
| Coffee | 22 |
For better understanding of the performance of all models used and to make right decisions we have also generated Confusion Matrix demonstrated in Figures 9 & 9 and Classification report containing Precision, Recall and F1 score of all models (Logistic Regression, Naïve Bayes, SVM, KNN, Decision Tree, and Random Forest) [15].
(a). Confusion Matrix - Logistic Regression
(b). Classification Report - Logistic Regression
The research presented here compares machine learning approaches used in crop recommendation systems to recommend high-yielding crops. To get the better insight of the dataset we have plotted graphs of correlation between each feature (Nitrogen, Phosphorus, Potassium, Temperature, pH and Humidity) and the label. Furthermore, each model is evaluated for accuracy using testing data. Table 3 and 4 displays the standardize and normalize the data for model compatibility, using Min-Max Scaling.
Before MinMax Scaling
| N | P | K | temperature | humidity | pH | rainfall | |
|---|---|---|---|---|---|---|---|
| 1656 | 17 | 16 | 14 | 16.396243 | 92.181519 | 6.625539 | 102.944161 |
| 752 | 37 | 79 | 19 | 27.543848 | 69.347863 | 7.143943 | 69.408782 |
| 892 | 7 | 73 | 25 | 27.521856 | 63.132153 | 7.288057 | 45.208411 |
| 1041 | 101 | 70 | 48 | 25.360592 | 75.031933 | 6.012697 | 116.553145 |
| 1179 | 0 | 17 | 30 | 35.474783 | 47.972305 | 6.279134 | 97.790725 |
After MinMax Scaling
| 0.12142857 | 0.07857143 | 0.045 | 0.21723408 | 0.9089898 | 0.48532225 | 0.29685161 |
| 0.26428571 | 0.52857143 | 0.07 | 0.53710965 | 0.64257946 | 0.56594073 | 0.17630752 |
| 0.05 | 0.48571429 | 0.1 | 0.53647858 | 0.57005802 | 0.58835229 | 0.08931844 |
| 0.72142857 | 0.46428571 | 0.215 | 0.47446209 | 0.708898 | 0.39001747 | 0.34576958 |
| 0. | 0.08571429 | 0.125 | 0.76468429 | 0.39318139 | 0.43145185 | 0.2783274 |
The Figure 8 shown above demonstrates that Random Forest and Naïve Bayes algorithms exhibit the utmost level of accuracy, while Logistic Regression and K-nearest Neighbors methods display the lowest level of accuracy as seen in the Table 5. Metrics such as the confusion matrix, precision, recall, and F1 score offer a more insightful analysis of the prediction results than accuracy alone. Confusion matrices display the true positive, true negative, false positive, and false negative predictions given by each model shown in the Figures 14(a) & 14(b), enabling us to evaluate the precision and efficacy of crop categorization. Additionally, the categorization reports provide a thorough evaluation of performance parameters such as precision, recall, and F1 score for each crop category. Precision assesses the model’s accuracy in identifying specific crops by determining the ratio of correctly identified instances to all instances predicted for that crop, for example, correctly identifying wheat crops out of all predicted wheat instances. Recall provides crucial insights into the model’s capability to capture and include all relevant events, such as ensuring all instances of a specific crop like rice are correctly identified and included in the predictions. The term refers to the proportion of true positive predictions relative to the combined number of false negative predictions and true positive predictions for each category. Taking the weighted harmonic mean of precision and accuracy, one can calculate the F1 score. The number of actual occurrences of the class in the provided dataset is referred to as support. Consequently, we have generated a classification report and visualization for each model. To obtain the most accurate forecast, it is essential to utilize above mentioned metrics provided in the analysis [16].
Accuracy of ML Models
| Models | Accuracy |
|---|---|
| Logistic Regression | 0.9636 |
| Naïve Bayes | 0.9954 |
| SupportVector Machine | 0.9681 |
| K-Nearest Neighbors | 0.9590 |
| Decision Tree | 0.9818 |
| Random Forest | 0.9931 |
(a). Confusion Matrix - Naïve Bayes
(b). Classification Report - Naïve Bayes
(a). Confusion Matrix - Support Vector Machine
(b). Classification Report - Support Vector Machine
(a). Confusion Matrix - K-Nearest Neighbors
(b). Classification Report - K-Nearest Neighbors
(a). Confusion Matrix - Decision Tree
(b). Classification Report - Decision Tree
(a). Confusion Matrix - Random Forest
(b). Classification Report - Random Forest
The proposed study will assist farmers in increasing agricultural output, reducing soil degradation in cultivated fields, and using less fertiliser during crop production by suggesting the optimal crop amount to plant based on a wide variety of criteria. The suggested activity helps farmers sustainability by helping them choose the right crops to cultivate. We have identified the advantages and disadvantages of each model through comparative analysis, providing important information for decision-making in the agricultural sector. We can improve the system later by adding more features to the dataset. Furthermore, with the support of remote sensing technologies and loT devices, real-time monitoring can be made possible which would allow the system to recommend crops along with climate change. To make the project beneficial to farmers in every corner of our nation, we may also incorporate all regional languages.