Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Accueil du site → Doctorat → Soudan → Developing Neural Network Models for Forecasting Sorghum Production in the Rain-fed Sector of Gedarif State - Sudan

University of Khartoum (2017)

Developing Neural Network Models for Forecasting Sorghum Production in the Rain-fed Sector of Gedarif State - Sudan

Ahmed, Isam Ahmed Basheir

Titre : Developing Neural Network Models for Forecasting Sorghum Production in the Rain-fed Sector of Gedarif State - Sudan

Auteur : Ahmed, Isam Ahmed Basheir

Université de soutenance : University of Khartoum

Grade : Doctor of Philosophy (PhD) 2017

Résumé
Sorghum is the main cereal crop in the Sudanese food basket. Most of Sorghum in Sudan is produced under the rain-fed agricultural sector. In the absence of reliable forecasting models, the country will continue to suffer from serious gaps in Sorghum production for food security. There is a lack of research of implementing neural networks in Sudan for forecasting. The study aims to fill in some of the research gaps in forecasting of Sorghum production in Sudan using neural networks. In this study, Artificial Neural Network (ANN) architectures are developed, investigated, and tested for forecasting Sorghum yield for the selected study area in Gedarif state. Rainfall data is provided by Meteorological Authority, Satellite images are collected from (Famine Early Warning System (FEWS) website, and Sorghum yield is provided by Ministry of Agriculture – Gedarif state. The study has developed and analyzed 32 models of neural networks to predict the production of Sorghum yield (as output data) by using the seasonal rainfall data and satellite images vegetation indices (as input data). The accuracy of each model was investigated for different input patterns. Multi-Layer Perceptron (MLP) and Radial Basis Function (RBF) networks have been adopted to develop Sorghum yield forecasting models. MLP has been implemented for developing 18 models using rainfall data as input data with different monthly, dekadal and cumulative rainfall patterns, respectively. Also 12 models have been developed using satellite images vegetation indices as input data to forecast Sorghum yield (as output data) (6 models using iv MLP and 6 models using 6 RBF). Both MLP and RBF are used to develop an early Sorghum yield forecasting models using cumulative monthly rainfall data and dekadal rainfall data. In the early Sorghum yield forecasting model there is two-phase training method using RBF in the first phase and MLP in the second phase. In the first phase, RBF has been used for training to forecast September and October cumulative monthly rainfalls using June, July and August cumulative monthly rainfall data. In the second phase, the forecasted cumulative monthly rainfall of September and October together with cumulative monthly rainfall data of June, July and August (five months) have been used as input data for the neural network model to forecast sorghum yield, using MLP for developing the model. The above-mentioned procedure has also been implemented using dekadal rainfall data. The correlation coefficient R-square has been used to evaluate the performance of the developed Artificial Neural Network (ANN) models, by comparing actual Sorghum yield data with forecasted Sorghum yield. The study concluded that the best model with highest value of R-square occurs when using Satellite images vegetation indices for months (July, August and September) as input data implementing MLP. If Satellite images vegetation indices data is not available the best two models that can be adopted are those models with either input data of monthly rainfall of (June, July, August and September) or cumulative rainfall of (June, July, August, September and October). Also, an early forecasting model, using two-phase training procedure, has been developed to enhance Sorghum yield forecasting using dekadal rainfall of June, July and August as input data, which is the best model compared to that using cumulative monthly rainfall of June, July and August. In addition the study has developed an algorithm to convert RGB satellite images into vegetation indices for the study area

Présentation

Page publiée le 15 mars 2019