Optimized Non-Overlapping Multi-Object Segmentation for Palm Oil Images Using FCN with Squeeze-and-Excitation and Attention Mechanisms

Abstract

Purpose: Palm oil plantation monitoring using UAV imagery presents significant challenges in multi-object segmentation due to homogeneous texture, low resolution, and difficulty in distinguishing disease symptoms. Traditional segmentation methods struggle to accurately separate overlapping and visually similar objects, reducing the effectiveness of automated analysis. This study aims to address these issues by proposing an optimized Fully Convolutional Network (FCN) incorporating Squeeze-and-Excitation (SE-Block) and Attention Mechanisms to enhance segmentation accuracy for multi-object, non-overlapping palm oil images.

Methods: The proposed model utilizes ResNet50 as a backbone, integrating SE-Block to enhance the feature representation of important regions while suppressing less relevant features. Additionally, Attention Mechanisms are incorporated to improve the model's spatial understanding and feature discrimination, which is crucial for segmenting visually similar objects in UAV imagery. A dataset of UAV-captured palm oil images was used to train and evaluate the model, applying deep learning techniques for feature extraction and classification.

Result: Experimental results demonstrate that the proposed method achieves an average Intersection over Union (IoU) of 0.7928, accuracy of 0.9424, precision of 0.9126, recall of 0.8622, F1-score of 0.8693, and mAP of 0.7673. The highest-performing model attained a maximum IoU of 0.8499 and an accuracy of 0.9490, significantly outperforming conventional FCN models. These findings confirm that incorporating SE-Block and Attention Mechanisms enhances segmentation accuracy, making the model more robust in handling UAV imagery complexities.

Novelty: The novelty of this research lies in the integration of SE-Block and Attention Mechanisms within FCN for palm oil segmentation, specifically targeting multi-object, non-overlapping segmentation in challenging UAV imagery conditions. By improving feature extraction and spatial attention, this approach advances deep learning-based agricultural monitoring and can be extended to other remote sensing applications requiring high-precision segmentation.