Md Sayedul Islam
Md Sayedul Islam
DUET CSE Student
Ongoing Undergraduate ResearchComputer Vision • Agriculture

Rice Panicle Detection and Growth-Stage Recognition

Agricultural Computer Vision & Lightweight Deep Learning for Resource-Constrained Environments

Academic Supervisor

Assistant Professor Umme Fawzia Rahim

Department of Computer Science & Engineering (CSE), DUET

Host Institution

Dhaka University of Engineering & Technology (DUET)

Gazipur-1700, Bangladesh

Research Overview:An ongoing academic research initiative focused on detecting rice panicles and classifying rice growth stages under natural field conditions. The project investigates computer vision and deep learning techniques with a particular emphasis on exploring lightweight, efficient models suitable for resource-constrained hardware.
Context & Background

Research Motivation

Rice is a primary staple crop globally and the foundational agricultural product in Bangladesh. Accurate identification of rice growth stages and panicle emergence is essential for timely agricultural interventions, such as fertilizer scheduling, water management, disease monitoring, and yield estimation.

Field-Level Challenges

  • •Manual in-field visual inspection is labor-intensive, subjective, and difficult to scale across broad acreage.
  • •Dense plant canopies, overlapping panicles, variable natural illumination, and visual clutter create complex detection conditions in field environments.
  • •Standard state-of-the-art vision models often require heavy computational resources that are impractical for field-level computing devices.

Intended Research Value

  • •Automated growth-stage tracking to support data-informed crop management decisions.
  • •Objective visual assessments to help reduce observation discrepancies.
  • •Investigation into practical edge intelligence that can eventually function in rural environments with limited computing infrastructure.
Research Scope

Core Objectives

1

Collect and curate a structured aerial-view photographic dataset spanning three physiological growth stages of rice.

2

Perform precise region/bounding-box annotation of rice panicles utilizing the Computer Vision Annotation Tool (CVAT).

3

Explore machine learning, deep learning, and computer vision methods for robust panicle detection and growth-stage recognition.

4

Investigate lightweight vision architectures capable of efficient inference on resource-constrained devices.

5

Explore the feasibility of future workflow integration into an accessible Android mobile application.

Data Foundation

Research Dataset & Distribution

Structured balance across critical physiological milestones of rice cultivation.

Total Acquired Images

4,500 Field Photographs

3 Growth Stages • 1,500 Images Each
Stage Distribution100% Balanced (Equal Splits)
Booting: 1,500 (33.3%)Heading: 1,500 (33.3%)Maturity: 1,500 (33.3%)

Booting

1,500 img

Early reproductive stage where panicle development occurs enclosed within the flag leaf sheath.

Dataset Proportion33.33%

Heading

1,500 img

Panicle emergence stage as floral structures emerge from the leaf sheath into the open canopy.

Dataset Proportion33.33%

Maturity

1,500 img

Ripening and grain filling stage characterized by color changes and mature grain development.

Dataset Proportion33.33%
Methodology

Data Collection & Annotation

Image Acquisition

Images were collected directly in rural agricultural fields using a smartphone camera captured from a drone-style aerial perspective to simulate top-down observation.

Perspective: Drone-style aerial viewpoints to simulate automated field surveying.

CVAT Annotation

In Progress

Using the open-source Computer Vision Annotation Tool (CVAT) to generate high-quality bounding-box and stage annotations across all collected field imagery.

Goal: Consistent multi-class ground-truth labels for panicle instances and stage recognition.
Pipeline

Current Research Workflow

End-to-end pathway from field data collection to eventual mobile-edge exploration.

Phase 1Completed

Field Image Collection

Acquisition of 4,500 field photographs from an aerial viewpoint across local rice paddies.

Phase 2Completed

Dataset Organization & Curation

Categorization and verification of images into three distinct physiological stages: Booting, Heading, and Maturity (1,500 each).

Phase 3In Progress

CVAT Annotation

Detailed bounding-box labeling of panicles and growth-stage labeling using CVAT.

Phase 4Planned

Efficient Model Exploration

Investigating lightweight object detection architectures suitable for constrained compute budgets.

Phase 5Planned

Empirical Evaluation

Evaluating detection fidelity, speed, model size, and generalization across varied lighting and canopy densities.

Phase 6Planned

Potential Android Integration

Exploring mobile integration to examine on-device edge inferencing for practical field usage.

Edge & Lightweight AI

Efficiency & Deployment Direction

Focus on Resource-Constrained Environments

High-end multi-GPU workstations are unavailable in typical agricultural field settings. This research emphasizes model efficiency—investigating compact model footprints, low memory bandwidth, and reduced parameter counts so that detection algorithms can operate reliably on edge devices.

Target Hardware Profile:

Resource-constrained embedded systems, edge AI boards, and modern mobile processors.

Potential Mobile Direction:

A planned direction is prototyping an Android application that could eventually bring computer vision capabilities directly to agricultural workers and researchers in the field.

Note: Android application integration and edge optimization are project goals and research directions, not claims of completed deployment or benchmarked performance.
Status Breakdown

Current Progress & Next Steps

Completed Work

  • ✓Field photography campaign collecting 4,500 aerial-view rice images.
  • ✓Dataset structuring and distribution balancing into Booting, Heading, and Maturity stages (1,500 images each).
  • ✓Setup and configuration of CVAT annotation workspace and labeling taxonomy.

In Progress

  • •Image-by-image bounding box annotation of panicles using CVAT.
  • •Data quality verification and consistency auditing across stage groupings.
  • •Preliminary literature exploration of lightweight vision backbones.

Planned Future Steps

  • →Training baseline and lightweight object detection models once annotation is complete.
  • →Rigorous comparative evaluation across stage recognition accuracy and latency.
  • →Exploratory prototype development for Android-based edge execution.
Academic Context

Supervision & Department Affiliation

Assistant Professor Umme Fawzia Rahim

Assistant Professor, Department of Computer Science & Engineering (CSE)

Faculty Supervisor
University: Dhaka University of Engineering & Technology (DUET)
Location: Gazipur-1700, Bangladesh
Student Researcher: Md Sayedul Islam (B.Sc. in CSE, 3rd Year)
Academic Focus: Computer Vision & Deep Learning