Intelligent Decision Support Systems for Smart City Applications

10Water Content Prediction in Smart Agriculture of Rural India Using CNN and Transfer Learning Approach

Rohit Prasan Mandal¹*, Deepanshu Dutta², Saranya Bhattacharjee² and Subhalaxmi Chakraborty^2†

¹Department of Computer Science and Technology, University of Engineering & Management, Kolkata, West Bengal, India

²Department of Computer Science and Engineering, University of Engineering & Management, Kolkata, West Bengal, India

Abstract

Soil moisture prediction is one of the emerging fields in the domain of smart agriculture. Ample moisture levels are of high importance to yields; thus, plants will not grow and develop with inadequate soil moisture, which influences soil temperature and heat capacity and also prevents soil from weathering. Accurate soil moisture detection can lead to proper growth of all kinds of trees and crops. In this chapter, we have implemented a deep learning and few transfer learning algorithms, such as convolutional neural network (CNN), VGGNet (VGG-16, VGG-19), Inception v3, ResNet50, Xception for prediction of soil moisture based on a very small custom dataset of 626 field images of rural India over a period of 10 days, both with and without the effect of data augmentation. We have also demonstrated how these pretrained algorithms save both time and resources compared to traditional deep learning models. Here, multiclass classification is considered based on three different classes, viz. “dry,” “wet,” “extremely wet.” The algorithms CNN, VGG-16, VGG-19, ...