DEVELOPMENT OF A VISION BASED CROP YIELD ESTIMATION SYSTEM

CHIBUEZE, GIDEON UWAOMA (2026) DEVELOPMENT OF A VISION BASED CROP YIELD ESTIMATION SYSTEM. Other thesis, GODFREY OKOYE UNIVERSITY, ENUGU.

Text COMPLETE WORK CHIBUEZE GIDEON UWOMA pdf.pdf Download (2MB)

Abstract

Food security and agriculture planning at the global level necessitate accurate prediction of crop yield estimates. The methods involved in predicting crop yield estimates usually involve manual farm surveys and simple statistical models which lack regional flexibility, do not capture the ecological diversity, and neglect the visual cues from the crops. With an aim to bridge these gaps, the current research seeks to propose a vision-based multi-modal system for estimating crop yield estimates in the Nigerian setting, emphasizing two of the most prevalent cash crops in the region, viz. cassava and maize. This research uses an OOD & Experimental Deep Learning-based approach..The new CNN architectures used in the model were invented on the basis of the EfficientNet-B0 architecture for analysis and identification of features from the visual images of the crops’ leaves to categorize their health condition and diseases. In contrast to the XGBoost model, which receives as input the automatically acquired climate data from NASA (temperature, precipitation, and humidity), along with historical agricultural data from FAOSTAT, to calculate the amount of the output in metric tons per hectare, the innovative characteristic of the methodology presented here is that it does not explicitly measure the plant counts but rather infers the density of the crops by its CNNs. The architecture has been experimented and deployed using FastAPI – a fast and simple to use framework along with web app interface to integrate two explainable AI (XAI) pipelines: Grad-CAM and SHAP for explainable image classification and explainable feature attribution of the climate data correspondingly. The results of these experiments show that the test accuracies of maize diseases classifier is $81.17\%$ and cassava diseases classifier is $70.7\%$. From the two baselines of Linear Regression and Random Forest, the suggested yield regression head based on XGBoost algorithm became the most efficient one, with MAE=$0.00988$, RMSE=$0.02482$ and $R^2$ score=$0.99994$. Production in tons, along with other factors, such as crop growth periods and total area, turned out to be the most impactful. It can be seen that an application of computer vision and weather data may significantly contribute to the development of a very efficient, non-invasive and reliable system of precision agriculture with many possible applications. Keywords: Precision Agriculture, Computer Vision, Deep Learning, CNN, XGBoost, Explainable AI (XAI), Yield Prediction

Item Type:	Thesis (Other)
Subjects:	H Social Sciences > H Social Sciences (General)
Divisions:	Faculty of Computing And Information Technology (FACIT)
Depositing User:	CHIBUEZE EZE
Date Deposited:	24 Jul 2026 13:12
Last Modified:	24 Jul 2026 13:12
URI:	http://eprints.gouni.edu.ng/id/eprint/5980

Actions (login required)

View Item