💻 computer science

ResAF-Net: An Anchor-Free Attention-Based Network for Tree Detection and Agricultural Mapping in Palestine

This paper introduces ResAF-Net, an anchor-free attention-based deep learning framework that leverages satellite imagery and integrates with local cadastral data to enable scalable, accurate tree detection and agricultural monitoring in the resource-constrained and fragmented landscapes of Palestine.

Original authors: Rabee Al-Qasem

Published 2026-04-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rabee Al-Qasem

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to count every single tree in a country where the land is chopped up into tiny, irregular pieces, and where you aren't allowed to fly drones over the fields to take pictures. That is the challenge faced by farmers and officials in Palestine. To solve this, a developer named Rabee Al-Qasem built a special computer program called ResAF-Net.

Here is how this paper explains the solution, using simple analogies:

The Problem: A Broken Puzzle

Think of Palestinian agriculture like a giant, broken mosaic. The land is fragmented, meaning farms are small and scattered. Traditional ways of counting trees (like sending people out to walk the fields) are too slow and difficult. Also, because of political restrictions, they can't use drones to get a bird's-eye view. They only have satellite photos, which are like looking at a giant puzzle from very far away. In these photos, trees often look small, blurry, or crowded together, making them hard to spot.

The Solution: A Smart "Tree Finder"

The author built an AI system (ResAF-Net) that acts like a highly trained detective looking at satellite photos. Instead of guessing where trees might be, it scans the image pixel by pixel to find them.

Here is how the "detective" works, broken down into its parts:

The Eyes (ResNet-50): This is the part of the AI that looks at the picture. It's like a pair of eyes that has been trained on millions of other images. It knows what a tree usually looks like, even if the photo is a bit blurry or the lighting is weird.
The Magnifying Glass (ASPP): Sometimes trees are tiny, and sometimes they are big. This module acts like a set of different-sized magnifying glasses. It zooms in and out simultaneously to make sure it doesn't miss a small sapling or a huge olive tree.
The Focus Filter (Self-Attention): Satellite photos are messy; they have rocks, roads, and shadows. This part of the AI acts like a noise-canceling headphone. It ignores the background "noise" (rocks and dirt) and focuses its energy only on the parts of the image that look like trees.
The Flexible Net (Anchor-Free Head): Old detection systems used "anchor boxes"—imagine trying to catch fish with a net that has holes of only one specific size. If the fish is too big or too small, it slips through. This new system is "anchor-free." It's like using a flexible, stretchy net that can mold around a tree of any shape or size, whether it's a tiny shrub or a massive canopy.

The Training: Learning from a Global Library

To teach this AI, the author didn't just use local photos (which were scarce). They used a massive global library of tree data called MillionTrees. It's like teaching a student to recognize apples by showing them photos of apples from all over the world, not just the ones in their own garden.

They fed the AI over 10,000 images and 700,000 labeled trees.
They taught it to handle different image qualities, sizes, and angles.

The Results: A High-Recall Net

When they tested the AI, it performed very well at one specific thing: finding trees.

The Catch Rate: It found 82% of the trees that were actually there. This is a huge improvement compared to other models, which only found about 53% or less.
The Trade-off: Because the AI is so eager to find trees, it sometimes gets a little "over-enthusiastic" and marks a rock or a shadow as a tree (false positives). However, the paper emphasizes that for agricultural planning, it is better to find almost every tree and double-check the extras later, than to miss the trees entirely.
Accuracy: Even with the extra noise, the AI was very good at pinpointing exactly where the center of the tree was.

Real-World Use: The Digital Map

The author didn't just stop at the computer code. They built a web-based map application that anyone can use.

Connecting the Dots: They linked this AI to official government land records (called GeoMolg).
Three Levels of Viewing:
1. Scene Level: You can look at any random spot on the map, and the AI instantly counts the trees there.
2. Parcel Level: You can click on a specific farmer's registered land plot, and the AI will count only the trees inside that specific legal boundary.
3. Community Level: You can look at an entire village, and the AI will process the whole area to give a total tree count for the community.

Why This Matters

The paper concludes that this tool proves it is possible to do large-scale agricultural monitoring in places with limited resources and strict restrictions. It turns satellite photos into a practical tool for farmers and policymakers to understand their land without needing to fly drones or hire armies of surveyors.

In short: The paper presents a new, flexible AI that uses satellite photos to count trees in Palestine with high accuracy, overcoming the challenges of small, fragmented farms and restricted airspace, and puts this data directly into a map tool for local use.

1. Problem Statement

The paper addresses the critical need for reliable, large-scale agricultural data in Palestine to support food security, land-use planning, and economic resilience. Several unique challenges hinder traditional data collection in this region:

Political and Regulatory Constraints: Strict restrictions on Unmanned Aerial Vehicle (UAV/drone) usage and limited field access make aerial monitoring difficult.
Environmental Complexity: The agricultural landscape is highly fragmented, with small, irregular tree crowns often obscured by shadows or dense vegetation.
Data Scarcity: There is a lack of locally annotated datasets suitable for Mediterranean agricultural conditions.
Limitations of Existing Models: Traditional object detection methods relying on "anchor boxes" (predefined bounding box templates) struggle with the dense, overlapping, and varying scales of tree crowns found in these fragmented landscapes.

2. Methodology: ResAF-Net Architecture

The authors propose ResAF-Net, a custom deep learning framework designed specifically for satellite-based tree detection in resource-constrained environments. The architecture integrates four key modules:

Encoder (ResNet-50 + ASPP):
- Utilizes a pre-trained ResNet-50 backbone for hierarchical feature extraction, chosen for its balance of performance and computational efficiency compared to heavier Transformer-based models.
- Integrates an Atrous Spatial Pyramid Pooling (ASPP) module at the final stage (C5) to aggregate multi-scale contextual information, addressing the challenge of detecting trees at varying sizes and improving feature richness.
Feature Fusion (FPN):
- Employs a Feature Pyramid Network (FPN) to unify feature maps of different channel dimensions into a single unified dimension.
- Includes smoothing operations to enhance the detection of small and medium-sized trees.
Attention Refinement (Multi-Head Self-Attention):
- A novel Multi-Head Self-Attention module is inserted before the detection head.
- It dynamically reweights features to amplify salient tree signals while suppressing background noise.
- Incorporates learnable 2D positional encodings (row and column embeddings) rather than fixed sinusoidal ones, providing greater flexibility for the varying spatial dimensions of the feature pyramid.
Anchor-Free Detection Head (FCOS):
- Replaces traditional anchor-based detection with a Fully Convolutional One-Stage (FCOS) head.
- Treats detection as a per-pixel prediction task, predicting:
  1. Classification Scores: Probability of a tree center.
  2. Regression Offsets: Distances to bounding box boundaries ( $l, t, r, b$ ).
  3. Centerness: A score indicating how close the pixel is to the box center, used to suppress low-quality predictions.
- Loss Function: Optimized using a weighted sum of Focal Loss (classification), Generalized IoU (regression), and Binary Cross-Entropy (centerness). The regression loss is weighted higher ( $\lambda_{reg}=2.0$ ) to prioritize precise localization of small, fragmented crowns.

3. Key Contributions

First Anchor-Free Application for Tree Crowns: This is the first work to test anchor-free object detection methods specifically for tree crown identification, demonstrating superior flexibility in handling dense and irregular canopies compared to anchor-based methods.
Integration with Official Cadastral Data: The study presents the first system to integrate deep learning-based monitoring with GeoMolg (Palestinian Ministry of Local Government's official cadastral data), enabling analysis at the parcel and community levels.
Resource-Constrained Optimization: The model is designed to run efficiently on limited hardware (e.g., single consumer-grade GPU) while maintaining high sensitivity, making it feasible for deployment in developing regions.
Practical Deployment: The authors moved beyond theoretical modeling to build a web-based GIS application that visualizes tree counts across scenes, individual land parcels, and entire communities.

4. Experimental Results

The model was trained on the MillionTrees benchmark (utilizing 10 high-quality sources, totaling ~10,000 images and ~700,000 annotations) and evaluated on a held-out validation set.

Performance Metrics:
- Recall: 82% (Significantly outperforming the DeepForest baseline at 53% and SAM3 at 20.6%).
- mAP@0.50: 63.03%.
- mAP@0.50:0.95: 35.47%.
- Precision: 13.89% (Lower precision is attributed to false positives in complex terrain, but the high recall ensures comprehensive coverage).
Training Efficiency: The model converged stably over 18 epochs on an NVIDIA GeForce RTX 3060 (12GB VRAM).
Real-World Validation:
- Scene Level: Detected 6,184 trees in a single view, correctly distinguishing cultivated cropland from trees.
- Parcel Level: Successfully detected 384 trees in a specific parcel known to contain ~410 trees, despite low image quality and stone chains.
- Community Level: Mapped 6,641 trees across the village of Al Zawiya, providing high-level indicators for community-level planning.

5. Significance and Impact

Strategic Policy Tool: The system provides a scalable, cost-effective alternative to traditional agricultural censuses, offering data-driven insights for policymakers in regions where aerial surveillance is restricted.
Food Security & Resilience: By enabling accurate monitoring of agricultural assets (olive, almond, palm trees), the tool supports better land-use planning and helps mitigate the impacts of land fragmentation and climate change.
Scalability: The web-based GIS integration allows stakeholders to operate at different scales (from individual farmers to national planners), facilitating the transition from manual data collection to automated, AI-assisted inventorying.
Future Potential: The framework lays the groundwork for future tasks such as species-specific identification, tree height estimation, and the creation of a custom, human-verified dataset for the Palestinian landscape.