Crop-Recommendation System

This repository contains a Machine Learning project, implemented in python, named 'Crop Recommendation System using Optimized Extreme Learning Machine'.

Dataset

A dataset named ‘Crop Recommendation Dataset’ was opted to make appropriate recommendations with the backing of Machine Learning techniques. The indicated dataset is available in a website named ‘Kaggle’. The Crop Recommendation Dataset consists of 2200 rows and 8 columns. These 8 columns correspond to 8 different attributes or features of the dataset. The attributes and their meaning are as follows:-
● N - Proportion of Nitrogen in soil content
● P - Proportion of Phosphorus in soil content
● K - Proportion of Potassium in soil content
● Temperature - Temperature in degree Celsius unit
● Humidity - Percentage of relative humidity
● ph - pH level of soil
● Rainfall - Amount of rainfall received in mm
● Label - Names of crops
Here, we choose 7 attributes, namely N, P, K, Temperature, Humidity, ph and Rainfall collectively as the input and the label as the output. This label contains 22 unique values which specify the name of different crops.

Exploratory Data Analysis

The basic statistical measures are calculated from the dataset to get an overall notion of the dataset. The number of instances corresponding to each unique crop is 100. From the heat map drawn, we could observe that there exists no null value in the dataset. The distance plot drawn from the attributes Temperature and ph conveys that both these attributes are normally distributed. A bar plot is drawn for comparing the N-P-K values which converts the data capable of easier analysis through observation alone. Thus after subjecting the original data for some statistical and visual analysis, we were able to obtain a clear image about the dataset and gather a comprehensible idea about the structure of the dataset. Therefore we could reach the conclusion that the 'crop recommendation dataset' is a befitting choice for building a crop recommendation system.

Comparitive Study

After the exploratory data analysis of the Crop Recommendation Dataset, it is submitted to the comparative study of classical supervised learning algorithms. These commonly used algorithms of machine learning are chosen one by one and they are trained using the available predefined models and their accuracies corresponding to its default parameters are noted down. Then each of these algorithms is provided with normalized data which is developed using different Normalization techniques. Afterwards, the data is submitted to hyperparameter tuning which enables us to choose the best parameters to yield an optimized accuracy for the model. GridSearchCV is the function we use for hyperparameter tuning. All the supervised algorithms we have chosen for analysis and normalization techniques are available in the scikit-learn library in Python.
The algorithms used for the comparative study are K Nearest Neighbor Algorithm, Decision Tree, Random Forest, Naive Bayes Algorithm, Logistic Regression, Multi-Layer Perceptron and Support Vector Machine.
We succeeded in optimizing the accuracies of these algorithms which they had from training the model using default parameters. Now, we move on to another advanced neural network algorithm to build the proposed model. Extreme Learning Machine algorithm is used in our project to build the crop recommendation system with an optimized accuracy and enhanced performance.

Implementation

Extreme Learning Machine or ELM is an advanced Neural Network algorithm for a single hidden layer feed-forward network. This is a three-step learning model algorithm. Random parameters are assigned for hidden nodes. Then, a matrix is calculated using the hidden layer output. The output weight is computed with the assistance of the Moore-Penrose generalized inverse of this matrix. All these steps are carried out using basic mathematics computation. Thus it needs much less training time and effort.
In our system, we create an Extreme Learning Machine model according to five different activation functions. These activation functions are hyperbolic tangent function, sine square function, triangular basis function, hard limit transfer function and radial basis function. Including these five activation functions, an Extreme Learning Machine classifier is defined. While defining this classifier the number of hidden neurons is given as 500. This number was chosen after examining the accuracy of the model while considering the number of hidden neurons within an interval of 50 to 650. After tuning, the model with hyperbolic tangent function as activation function gained the maximum accuracy and efficient performance. Thus, our proposed model developed a system with ease of implementation and minimum human intervention.

Conclusion

The Crop Recommendation System nurtured through the optimized Extreme Learning Machine model has several advantages over the conventional classification algorithms. When compared with the existing Crop Recommendation Systems, our model acknowledges to result in the maximum accurate model. In our project, a comparative study of the existing classification algorithms was carried out. But, the Extreme Learning Machine model managed to generate the maximum prediction accuracy of 99.77% by using the hyperbolic tangent function as the activation function. This was the highest accuracy value when compared with all the other existing conventional algorithms.