[Paper Review] A Deep Learning Approach Based on Graphs to Detect Plantation Lines

Diogo Nunes Gonçalves, Mauro dos Santos de Arruda+9 more|arXiv (Cornell University)|Feb 5, 2021

Remote Sensing and LiDAR Applications36 references4 citations

TL;DR

This paper proposes a graph-based deep learning method for detecting plantation lines in UAV-based RGB aerial imagery, leveraging visual features, line probability, and displacement vector alignment to classify edges between plants as part of a line. The approach achieves 98.7% precision, 91.9% recall, and 95.1% F1-score, outperforming state-of-the-art methods in detecting spaced and gap-interrupted plantation patterns.

ABSTRACT

Deep learning-based networks are among the most prominent methods to learn linear patterns and extract this type of information from diverse imagery conditions. Here, we propose a deep learning approach based on graphs to detect plantation lines in UAV-based RGB imagery presenting a challenging scenario containing spaced plants. The first module of our method extracts a feature map throughout the backbone, which consists of the initial layers of the VGG16. This feature map is used as an input to the Knowledge Estimation Module (KEM), organized in three concatenated branches for detecting 1) the plant positions, 2) the plantation lines, and 3) for the displacement vectors between the plants. A graph modeling is applied considering each plant position on the image as vertices, and edges are formed between two vertices (i.e. plants). Finally, the edge is classified as pertaining to a certain plantation line based on three probabilities (higher than 0.5): i) in visual features obtained from the backbone; ii) a chance that the edge pixels belong to a line, from the KEM step; and iii) an alignment of the displacement vectors with the edge, also from KEM. Experiments were conducted in corn plantations with different growth stages and patterns with aerial RGB imagery. A total of 564 patches with 256 x 256 pixels were used and randomly divided into training, validation, and testing sets in a proportion of 60\%, 20\%, and 20\%, respectively. The proposed method was compared against state-of-the-art deep learning methods, and achieved superior performance with a significant margin, returning precision, recall, and F1-score of 98.7\%, 91.9\%, and 95.1\%, respectively. This approach is useful in extracting lines with spaced plantation patterns and could be implemented in scenarios where plantation gaps occur, generating lines with few-to-none interruptions.

Motivation & Objective

• To address the challenge of detecting plantation lines in aerial RGB imagery with spaced and irregular plant patterns.
• To improve robustness in detecting lines amidst gaps, isolated plants, and weeds that disrupt traditional methods.
• To integrate geometric relationships between plants—via displacement vectors—into a deep learning framework for enhanced line detection.
• To develop a method that generalizes across different growth stages and complex plantation geometries.
• To demonstrate the effectiveness of combining graph modeling with deep features for linear object detection in remote sensing.

Proposed method

• A VGG16 backbone extracts initial feature maps from 256×256 RGB image patches of corn plantations.
• A Knowledge Estimation Module (KEM) processes the feature map through three parallel branches: plant position detection, line probability estimation, and displacement vector computation.
• Each plant is modeled as a vertex in a graph, with edges formed between adjacent plants based on spatial proximity.
• Edge classification as part of a plantation line relies on three probabilities: visual feature similarity, likelihood of belonging to a line (from KEM), and alignment of displacement vectors with the edge.
• The final classification uses a threshold of 0.5 for each probability component to determine edge membership in a line.
• The model is trained and evaluated on 564 patches split into 60% training, 20% validation, and 20% testing sets.

Experimental results

Research questions

• RQ1 Can a graph-based deep learning model improve the detection of spaced plantation lines in UAV RGB imagery compared to conventional deep learning methods?
• RQ2 How does incorporating displacement vector alignment enhance the robustness of line detection in the presence of gaps and isolated plants?
• RQ3 To what extent does combining visual features, line probability, and geometric consistency improve detection performance on complex agricultural patterns?
• RQ4 Can the proposed method generalize across different growth stages and irregular row patterns in corn plantations?
• RQ5 Does the integration of graph modeling with deep features lead to superior performance in linear feature extraction for remote sensing applications?

Key findings

• The proposed method achieved a precision of 98.7%, recall of 91.9%, and F1-score of 95.1% on test data, significantly outperforming state-of-the-art methods like Deep Hough Transform and PPGNet.
• The inclusion of displacement vector alignment in the KEM branch reduced false positives from isolated plants and weeds, improving detection robustness.
• The model demonstrated strong generalization across different corn growth stages and irregular row patterns, including those with significant gaps.
• The graph-based edge classification mechanism effectively captured linear continuity even with few-to-none interruptions in plantation lines.
• The ablation study confirmed that the third branch (displacement vector alignment) was critical for performance, as its removal degraded results significantly.
• The method is transferable to other linear feature detection tasks such as road, river, and utility line extraction in remote sensing imagery.