Regional-Scale Acacia tortilis Crown Mapping from UAV Remote Sensing Using AI-Assisted Annotation and a Lightweight Hybrid MambaVision-Based Segmentation Framework

Authors : M.B.A. Gibril, R. Al-Ruzouq, A. Shanableh, N. Hammouri, F. Lamghari, S.M. Ahmed, A. Mansour, R. Jena, H. Shanableh, M.A. Almarzouqi, N.S. Alafayfeh, and S.Z. Ghebremeskel

Link : Prioritizing Citrullus colocynthis accessions and... | F1000Research

Context and Motivation

Acacia tortilis (Umbrella Thorn Acacia) is a keystone species of arid and semi-arid ecosystems across Africa and the Middle East. It plays a vital ecological and socio-economic role—improving soil fertility, providing fodder and fuelwood, supporting biodiversity, and contributing to the production of high-value Acacia honey.However, its populations face growing environmental and anthropogenic pressures, including drought, land-use change, and pest infestation. Traditional field-based monitoring is expensive and limited in scale. This research addresses that gap by introducing a UAV- and AI-based mapping framework capable of large-scale, high-resolution monitoring of A. tortilis across diverse landscapes of the UAE.

2. Objectives

The main objective is to design a lightweight, transferable, and accurate deep learning framework for region-wide A. tortilis crown delineation from UAV imagery.Specific aims:

1. Develop an AI-assisted annotation pipeline to overcome manual labeling challenges.

2. Create a hybrid segmentation model integrating CNN, Transformer, and Mamba units.

3. Benchmark the proposed model against existing CNN- and Transformer-based architectures.

4. Assess model generalizability and computational efficiency across heterogeneous landscapes.

3. Study Area and Data

The research covers approximately 250 km² across Fujairah and eastern Sharjah, UAE—an ecologically diverse region encompassing mountains, alluvial plains, and deserts.

• UAV imagery was acquired using a senseFly eBee X fixed-wing drone with an RGB S.O.D.A. camera at 2.5–3 cm GSD.

• Flights were planned with 70% frontal and 40% lateral overlap under clear conditions.

• Data were post-processed using PPK corrections for high positional accuracy and orthomosaic generation in Pix4Dmapper.

• A comprehensive field campaign catalogued 9,100 trees, recording crown dimensions, height, and coordinates using ArcGIS Field Maps.

4. AI-Assisted Annotation Pipeline

Given the massive scale of imagery and the labor-intensive nature of manual labeling, the authors developed an AI-assisted annotation and augmentation pipeline:

• The Segment Anything Model (SAM) was fine-tuned with Low-Rank Adaptation (LoRA), creating SAMLoRA, to delineate crowns automatically.

• SAMLoRA achieved an 88% F-score in crown delineation.

• Predictions were verified through expert validation and Google Street View, producing a reliable dataset of 36,800 Acacia crowns (a 4× increase in training data).

• These labeled crowns formed the foundation for training and validating the deep learning segmentation framework.

5. Hybrid MambaVision Framework

The study introduces U-Net–MambaVision, a U-shaped encoder–decoder architecture that fuses three paradigms:

• CNN modules capture fine-grained local details.

• Transformers provide long-range context and global relationships.

• Mamba units (State Space Models) enable efficient sequence modeling with linear complexity, improving performance and speed.

The encoder comprises four hierarchical stages:

1. Early CNN residual blocks extract high-resolution features.

2. Later MambaVision mixer blocks refine spatial and global dependencies.

3. The decoder progressively upsamples the features using skip connections to restore spatial detail.A lightweight CNN-based decoder was ultimately selected for its balance of accuracy and computational cost.

6. Experimental Design

• Frameworks used: PyTorch, MMSegmentation, and Hugging Face Transformers.

• Training configuration: 100,000 iterations, AdamW optimizer, batch size 2, with random cropping and photometric augmentation.

• Benchmarked models: DeepLabV3+, PSPNet, U-Net (ResNet-50), SegFormer, Segmenter, VMamba, and Mask2Former.

7. Results

AI Annotation Efficiency

• Manual annotation time reduced by more than 70%.

• Generated a diverse and spatially rich dataset, critical for robust model training.

Segmentation Performance

• On the independent test dataset:

  • U-Net–MambaVision variants achieved mIoU = 85.30–85.44% and mF-score = 91.52–91.61%.

• On geographically distinct test regions (generalizability set):

  • The small variant (U-Net–MambaVision-s) reached mIoU = 89.48% and mF-score = 94.17%.

• Outperformed both Transformer and CNN counterparts, combining higher accuracy with shorter training time.

Decoder Design

• U-Net decoder achieved a strong accuracy–efficiency trade-off (11.8 h training vs. 35.9 h for UperNet).

• Demonstrated that complex decoders add computational burden with minimal accuracy gain.

Regional Mapping

• Applied over the full 250 km² UAV dataset.

• Detected approximately 106,000 Acacia crowns, offering the first high-resolution inventory of this species in the UAE.

• Demonstrated robust performance in mountainous, agricultural, and coastal terrains.

8. Discussion

This framework significantly advances ecological monitoring in arid regions by combining AI, UAV imagery, and state-space modeling.Its main advantages include:

• Automation: Reduces manual labor and accelerates annotation.

• Scalability: Performs well over diverse topographies.

• Accuracy: Excels in differentiating Acacia crowns even under occlusion or shadow.

• Efficiency: Lightweight and adaptable to resource-constrained environments.

Limitations included minor misclassifications at image edges and confusion between Acacia and visually similar shrubs. These can be mitigated by augmenting training with background-only tiles and fine-tuning for specific terrains.

9. Conclusions and Future Work

This study marks a regional and methodological first in applying hybrid Mamba-based deep learning to ecological UAV mapping. The proposed U-Net–MambaVision framework delivers:

• Superior segmentation accuracy and computational efficiency;

• Transferability across heterogeneous landscapes;

• Practical deployment potential for environmental authorities and conservation planners.

Future directions include:

1. Integrating satellite imagery to scale mapping from regional to national levels.

2. Combining soil, hydrological, and climatic variables to analyze habitat health.

3. Applying the system to ecosystem restoration, biodiversity assessment, and honey productivity optimization.

10. Significance

Beyond its scientific contribution, this project aligns with the UAE’s sustainability vision and the Fujairah Research Centre’s commitment to developing AI-driven environmental monitoring tools. By providing a scalable model for data-driven conservation, it demonstrates how AI and UAV technologies can bridge the gap between research and actionable ecosystem management—supporting both biodiversity preservation and economic value creation through sustainable natural resource use.

Acknowledgments:Supported by the UAE Ministry of Climate Change and Environment, Fujairah Research Centre (Project No. 133049), and the University of Sharjah.

Keywords: Acacia tortilis, UAV, deep learning, hybrid MambaVision, CNN–Transformer integration, Fujairah, arid ecosystems, AI-assisted annotation.