Aerial Images and Convolutional Neural Network for Cotton Bloom Detection

Area of Science:

• Agricultural Science
• Computer Vision
• Remote Sensing

Background:

• Accurate monitoring of crop flowering is crucial for production management, yield estimation, and genotype selection.
• Traditional methods for counting cotton blooms are labor-intensive and time-consuming.
• Developing automated, high-throughput methods for monitoring flowering progress is essential for modern agriculture.

Purpose of the Study:

• To develop and validate a methodology for detecting and counting cotton flowers (blooms) using unmanned aerial system (UAS) imagery.
• To create a system capable of determining the 3D location of individual blooms.
• To enable automated, continuous monitoring of cotton flowering progress.

Main Methods:

• Acquisition of aerial color images from cotton test fields using a UAS.
• Design and training of a convolutional neural network (CNN) for bloom detection in raw images.
• Utilizing structure from motion (SfM) to generate a dense point cloud and calculate 3D bloom locations.
• Development of a constrained clustering algorithm for registering blooms across multiple images based on 3D coordinates.
• Analysis of dense point cloud quality and its impact on 3D location accuracy.

Main Results:

• The CNN effectively detected cotton blooms in aerial images.
• The 3D location accuracy of blooms was influenced by dense point cloud quality.
• The constrained clustering algorithm demonstrated good efficiency and accuracy in bloom registration.
• Bloom counts using the proposed method showed comparability with manual counts, with minor errors in single-plant plots.
• Underestimation of bloom counts occurred in dense, multi-plant plots due to occluded blooms.

Conclusions:

• The developed methodology offers a high-throughput and automated approach for monitoring cotton flowering progress.
• The system provides valuable data for crop production management and yield estimation.
• Further refinement may be needed to address challenges in dense planting scenarios and occluded blooms.