Crop yield prediction using multi-parametric deep neural networks

E Kalaiarasi

doi:10.17485/ijst/v14i2.2115

Abstract

Objective: To propose Multi-parametric Deep Neural Network (MDNN) for modeling the impact of climate changes, multiple parameters related to the weather and soil for accurate crop yield prediction. Methods: In MDNN, a measure called Growing-Degree Day (GDD) is introduced for measuring the overall effect of weather conditions related to the crop yield. One of the key elements in MDNN is the neuron’s layer-wise activation function. In order to enhance the crop yield predictive performance, a leaky rectified linear unit is used in the activation units of MDNN. For the analysis of performance of DNN and MDNN, data about weather, crop and soil are collected from http://www.ccafs-climate.org/climatewizard/, https://data.world/thatzprem/agriculture-india and https://data.gov.in/search/site?query=soil respectively. From the collected data, 60000 records are used for training and 40,000 records are used for testing. Findings: By considering multiple parameters of climate and the effect of weather on crop yield, the accuracy of MDNN is improved for predicting the crop yield. The effectiveness of MDNN is tested and compared with DNN for different types of crops. The MDNN achieves 91.84% of mean accuracy for five different crops compared to the DNN classification. Novelty: This proposed work tries to predict the crop yield more accurately by analyzing the climate, weather and soil parameters. The MDNN considerably improves statistical efficiency over typical DNN by using previous knowledge about important phenomena and functional forms relating them to the crop yield. Keywords: Crop yield prediction; machine learning; DNN; climatic changes; soil parameters; growing degree-day

Highlights

One third of the land on Earth is used for farming, with the majority of it being used for agriculture
The efficiency of the [7]Artificial Neural Network (ANN), [7]Multiple Linear Regression (MLR), [6]DNN and Multi-parametric Deep Neural Network (MDNN) method for crop yield prediction is tested based on different metrics used for evaluating the classifier performance
The leaky ReLU (lReLU) permits non-zero gradient during idle state of node which enhance the precision of MDNN for crop yield prediction

Summary

Introduction

One third of the land on Earth is used for farming, with the majority of it being used for agriculture. Agricultural demand is increasing due to the growing world population. The control of farming activities is faced by unique problems including the seasonal cycles of crop phenology and crop production depends on weather, climate and soil parameters. A large number of farmers don’t get the planned profit because of these challenges. Farmers require timely guidance on the potential profitability of crops, Kalaiarasi & Anbarasi / Indian Journal of Science and Technology 2021;14(2):131140 so an analysis must be conducted to assist farmers to increase their profit. One of the major issues in agriculture field is yield prediction [1,2]

Methods

Findings

Discussion

Conclusion