Author ORCID Identifier

https://orcid.org/0000-0003-1969-3559

Date of Award

4-2-2025

Document Type

Thesis

School

School of Computing

Programme

Ph.D.-Doctoral of Philosophy

First Advisor

Dr.K.R.Manjula

Keywords

Deep Learning Models, Agricultural Drought Prediction, Time Series Analysis, Machine Leaning Models

Abstract

Accurate and timely prediction of agricultural droughts is fundamental to the management of ecosystems, preservation of biodiversity, and environmental sustainability. It helps to reduce environmental exploitation, promotes eco-friendly irrigation techniques, and tackles climate change efficiently. Its social implications include food safety and financial stability through continuous rural employment. Drought prediction guarantees social progress and sustainability by ensuring zero hunger/poverty and effective water management. A multi-sensor image pre-processing framework that incorporates data from various satellite platforms like Landsat (30 metres), Sentinel-2 (10 metres, 20 metres, 30 metres), and MODIS (500 metres) is used to monitor and assess agricultural drought effectively.

Analysis of satellite imagery brings in many useful improvements. It enhances images using advanced correction techniques and a hybrid classification system that integrates Differential Evolution (DE) optimization with Optimized Learnable Parameter Artificial Neural Network (OLPANN) to produce precise land cover maps. As a complementary study of environmental monitoring, this study develops a framework for drought prediction using Holt Winter-Convolution 2D Long-Short Term Memory (HW-CONV2D-LSTM) models using 22 years of relevant satellite data and reveals useful patterns in temperature, rainfall, soil moisture, and plant growth.

The study’s first objective addresses issues in pre-processing satellite images by integrating radiometric and atmospheric corrections, spatial resolution, and advanced feature extraction techniques. Radiometric correction converts unprocessed satellite data in the form of Digital Numbers (DN) into reliable radiance values to ensure a precise depiction of surface reflectance attributes by considering solar illumination and the characteristics of sensors. The radiometric correction uses noise removal techniques, such as Gaussian, median, and anisotropic filters to eliminate sensor-specific artefacts and enhance the signal-to-noise ratio of multispectral imagery. The atmospheric correction balances the interactions between electromagnetic radiation and various atmosphere constituents to optimise the Aerosol Optical Depth (AOD) by incorporating Dark Object Subtraction (DOS) and considering the maximum pixel value of aerosol particles to achieve precise surface reflectance values during fluctuating atmospheric conditions. The second objective proposes to employ a novel classification system integrating DE optimization with OLPANN architecture. The DE optimization framework, configured with a population size of 42, maximum generations of 100, crossover rate of 0.7, and mutation factor of 0.5, optimizes the OLPANN's hyperparameters, feature weights, and architecture. The OLPANN model examines the potentials of spectral, spatial, texture, and indices information from the multispectral data using Pearson correlation measures. The optimized deep features from the original data set increase the robustness of the artificial neural network model and provide a faster classification result; these optimized features are then trained by the ANN for further processing. Landsat 5 and 8, sentinel-2 data set is used to analyse three different types of features for evaluating seven different land cover classes. McNemar’s test is carried out to evaluate the changes, which endorses the OLPANN and Optimized Extreme Gradient Boosting (OXGB) to make it statistically significant. Friedman’s test demonstrates that the variance of Optimized Random Forest (ORF), Optimized Support Vector Classifier (OSVM), and Optimized Decision Tree (ODT) are significant at 0.01%. The numerical outcomes obtained establish that OLPANN has the potential to achieve the highest accuracy of 94.07%. The system's performance is rigorously evaluated using comprehensive Kappa coefficient metrics. Based on the test results, the OLPANN classifier is recommended as the best candidate that produces ideal measures for the various characteristics considered for the study. This analysis empowers the government to identify urban extension, delineate any damages in natural land cover, ascertain legal boundaries for property assessment, and detect roads, bridges, and water and other land surface interfaces.

Agricultural droughts significantly affect rain-fed crops and thereby decrease employment possibilities and per capita income. Agriculture drought affects all nations significantly and cannot be avoided owing to the changing climate conditions. However, its impact on the environment could be reduced by predicting its occurrence in a timely and accurate manner. Presently, the detection of droughts largely relies on ground-based monitoring stations, however, satellites can scan vast land masses from above and provide accurate and reliable monitoring solutions. Multispectral imagery and spatio-temporal data from satellites offer quick responses, which enables the decision-makers to successfully manage agricultural resources and crop quality.

The third objective establishes a comprehensive time series analysis framework for drought prediction by implementing and comparing ARIMA, SARIMA, and Holt-Winters models based on 22 years (from 2000 to 2022) of multispectral data obtained from Thanjavur Station. The methodology incorporates sophisticated seasonal decomposition techniques, automated model selection protocols, and multi-scale temporal analysis evaluated with rigorous statistical metrics including Bayesian Information Criteria (BIC) and Akaike Information Criteria (AIC). The framework demonstrates superior performance by implementing advanced models, pattern recognition algorithms and adaptive parameter optimization techniques while maintaining robust validation protocols across diverse temporal scales. The proposed method employs the Holt Winter Conventional 2D-Long Short-Term Memory for meteorological and agricultural droughts' prediction based on the precipitation index data sets from Climate Hazards Group Infrared Precipitation with Station, MODIS 11A1 temperature index, and MODIS 13Q1 vegetation index. The time series data for trends and seasonality are extracted from the satellite images based on Holt Winter alpha, beta, and gamma parameters. Finally, an efficient procedure for predicting drought is developed based on the Conv2D-LSTM for computing the spatio-temporal correlation among drought indices. The HW-CONV2D-LSTM presents an improved value of R2 equal to 0.97 and holds promise to act as an effective computer-assisted strategy for predicting drought and maintaining agricultural productivity, which is important to feed the ever-increasing human population.

Download

Share

COinS