Case study

We know that agriculture is an important part of gross domestic product. This project cussed in the advantage of insurance companies so they have efficient insurance coverage. In this project, we have taken two test datasets. One is 2CSV files and another is image dataset.

CSV file has lots of features like temperature, humanity, pressure and precipitation type, such as snow, rain and all other weather conditions and also wind speed visibility. Here we are predicting a crop field condition. The second dataset is the image dataset which consists of 10 images of crop fields. With drone images, we have manually counted the number of crops in each image. Here we have used Convolutional Neural Network (CNN) regression to predict the contradicting crops (Figures 5.10 and 5.11).

Next, we discuss the preprocessing of the dataset and different machine learning algorithms used in this case study. So, first in the preprocessing part, we have to input all the set variable data using one odd in coding, then we have to normalize the data using max scalar. Then, we have to fill nine values by the mean values of that respective columns. Afterward we split whole dataset in the two parts, the first part is the training set, the second is testing set. Training set consists of 70% of the data and the second part consists 30%. Now we come to the different algorithms used. First, we used random forest, after that we applied support vector regression then, deep neural networks for regression.

To explain in detail the case study we studied the crops in the field that can be counted using the images. We used ten images taken by a drone and counted them manually with several crops per image. It took about 2 hours to count this. We then took eight of the images for training a model and two images for testing a model to see error rates. We used CNN model and modified it for predictive purposes by including a dancing layer with a linear activation function in the output layer. We normalized the images to 400 × 400 pixels with the length. So, feeding images into the model became easier. A model has about 14 hidden layers with a 25% dropout function. Since we do not have a large dataset, we cannot expect very large accuracy but we can predict a moderate level of accuracy based on the dataset (Figures 5.12–5.14).