Author ORCID Identifier

0000-0001-7639-6863

Date of Award

17-8-2025

Document Type

Thesis

School

School of Computing

Programme

Ph.D.-Doctoral of Philosophy

First Advisor

Dr.K.Geetha

Keywords

Machine Learning, Image Processing, Deep Learning, Precision Agriculture

Abstract

Precision agriculture also referred as precision farming or smart farming, is an innovative approach to agricultural management that leverages technology and data to optimize various aspects of the farming process. This approach aims to make farming more effective, sustainable, and profitable by affording farmers with the application tools and information they need to make more informed decisions.

Precision agriculture combines elements of agriculture, technology, and data science to enhance crop production, and resource utilization. Precision agriculture techniques can be highly effective in leaf disease detection within crop fields. Machine learning has been developed incredibly across multiple domains and shown it excellence in precise predictions, detailed analysis of data, recommendation tasks etc.

Deep learning which is a subdivision of machine learning performs outstanding tasks by utilizing artificial neural networks that mimics the human brain’s learning process thereby automating the feature extraction tasks. It also uses a good optimization algorithm to adjust its weights thereby reducing the loss and impacting more on accuracy.

Many challenges lie behind deep learning because of its black box architecture and can be addressed by giving optimal solutions. Overall agricultural process can be automated by employing the neural network and poses significant challenges at each farming tasks. Early and accurate detection of leaf diseases is crucial for preventing yield loss and optimizing the use of pesticides. Precision agriculture for leaf disease detection not only enhances crop management but also contributes to sustainable and environmentally friendly farming practices by reducing the excessive use of chemicals. It enables farmers to protect their crops more effectively while optimizing resource utilization and increasing overall productivity.

The significant contribution of the research is highlighted as follows:

1. At the initial stage every agricultural field needs the best crop to be planted according to the environmental and soil conditions. In the current situation there is a tough weather and many floods raised and damaged the field and yielding a heavy loss to farmers. So, it is important to analyze the meteorological factors and soil conditions in order to choose a right crop. Hence a crop recommendation model based on meteorological and soil parameters is implemented using machine learning technique. An ensembled with lightGBM outperforms the other machine learning approaches and yielded best accuracy by classifying a maximum of 70 different classes.

2. If the recommended crop is soybean for the field it has to be monitored to track if it is affected by any disease or healthy. Two major diseases such as angular leaf spot, bean rust is compared with healthy leaves. Since the input data are images, a neural network-based classification will be appropriate. An optimized densenet model based on Bayesian hyperparameter tuning is proposed to accurately classify the disease with reduced model parameters.

3. The overlapping of bean leaves or dense leaves cause inappropriate classification or it will make the model to learn unwanted features from the neighbor leaves also instead of focusing on the exact diseased portions. Overlapping is removed from the focused image and based on the segmented region classification is done. A DIM-UNet model is proposed to perform segmentation of the overlapped leaves and a sparse regularized auto encoded feature extractor (SR-AE) captures best features from the segmented leaves. SR-AE features are compared with CNN model-based features for classification where SR-AE outperforms the other approaches.

4. Soybean is prone to many diseases and among that some diseases such as mossaicvirus, crestamento, septoria and ferrugen has very less samples which is not sufficient for classification. Such images can be generated synthetically using GAN approach and can have adequate data. An optimized DCGAN model is suggested which synthetically generates images as well as overcome mode collapse problem.

Download

Share

COinS