Zing Forum

Reading

ELDOR: A Large-Scale Drone Dataset and Benchmark for Monitoring Illegal Gold Mining in the Amazon Rainforest

ELDOR is a drone aerial dataset covering over 2500 hectares, specifically designed for monitoring illegal gold mining activities in the Amazon Rainforest. This dataset provides pixel-level semantic annotations, supports four benchmark tasks, and reveals the limitations of existing methods in identifying small-scale mining structures and fine-grained restoration categories.

ELDOR亚马逊雨林非法金矿开采无人机遥感语义分割环境监测计算机视觉基准数据集
Published 2026-05-15 04:30Recent activity 2026-05-18 16:17Estimated read 4 min
ELDOR: A Large-Scale Drone Dataset and Benchmark for Monitoring Illegal Gold Mining in the Amazon Rainforest
1

Section 01

ELDOR: Introduction to the Large-Scale Drone Dataset and Benchmark for Monitoring Illegal Gold Mining in the Amazon

ELDOR is a drone aerial dataset covering over 2500 hectares, designed specifically for monitoring illegal gold mining in the Amazon Rainforest. It provides pixel-level semantic annotations, supports four benchmark tasks, reveals the limitations of existing methods in identifying small-scale mining structures and fine-grained restoration categories, and fills a gap in the field of environmental remote sensing.

2

Section 02

Background and Challenges: Environmental Threats of Illegal Mining in the Amazon and Monitoring Dilemmas

Illegal gold mining causes deforestation, water mercury pollution, and ecological damage in the Amazon Rainforest. Traditional satellite remote sensing has limitations such as insufficient resolution, cloud cover interference, and difficulty detecting subtle changes, making it hard to accurately monitor over 2000 illegal mining sites.

3

Section 03

Core Features of the ELDOR Dataset: Scale and Annotation System

The ELDOR dataset covers a continuous area of over 2500 hectares, including high-resolution orthophotos. Pixel-level semantic segmentation annotations distinguish mining-affected areas from surrounding ecosystems, enabling analysis of the penetrating impact of mining on ecosystems.

4

Section 04

Four Benchmark Tasks: Comprehensive Evaluation of Computer Vision Model Performance

Four tasks based on ELDOR: 1. Semantic segmentation (pixel classification); 2. Segmentation-derived recognition (high-level semantic concepts); 3. Direct multi-label classification (end-to-end scene recognition); 4. Vision-language category presence recognition (natural language feature detection).

5

Section 05

Experimental Findings: Limitations of Existing Models and Improvement Directions

Existing models have issues such as missed detection of small targets (small-scale mining structures), insufficient fine-grained classification (ecological restoration stages), and limited context understanding. There is a need to develop multi-modal models that integrate multi-source data (topography/hydrology/temporal sequence).

6

Section 06

Interactive Tools: Facilitating Dataset Application and Model Development

The research team developed an interactive platform that supports aerial data visualization navigation, model prediction comparison, expert annotation correction, and custom monitoring model prototype development, promoting the practical application of the dataset.

7

Section 07

Research Significance and Application Prospects: Value and Future Directions of ELDOR

ELDOR provides a standardized benchmark for illegal mining monitoring. Its application directions include real-time monitoring systems, policy decision support, ecological impact assessment, and multi-modal model research, driving technological progress in Amazon Rainforest protection.