Hiroaki Okonogi

Key Skills

Software

LabelImg

Other

Top Subject Matter

No subject matter listed

Top Data Types

Image

Geospatial Tiled Imagery

Top Label Types

Bounding Box

Segmentation

Classification

Freelancer Overview

I am a detail-oriented Geospatial AI specialist with extensive experience in satellite image interpretation, data labeling, and AI training dataset preparation. My background includes supervising large-scale image classification and segmentation verification projects, particularly in remote sensing and environmental monitoring domains. I am highly skilled in data annotation, structured dataset validation (10,000+ records), and quality assurance, with a strong focus on accuracy and strict guideline compliance. I am proficient in Python (NumPy, Pandas, Rasterio), QGIS/ArcGIS, and advanced Excel/CSV handling, and have successfully supported AI model development and data validation workflows for both research and operational projects. I thrive in independent remote work and am committed to delivering reliable, high-quality training data for AI applications.

Labeling Experience

Selective Logging Detections

OtherGeospatial Tiled ImageryClassification

Selective logging—the removal of valuable timber species—is often the earliest stage of deforestation in the Amazon, yet it is difficult to detect using conventional methods based solely on satellite imagery. Therefore, labeling of selective logging areas was conducted to enable detection using machine learning approaches. This dataset is also valuable for monitoring sustainably managed forest areas. Using multi-band data from Sentinel-2, image chips containing evidence of selective logging were identified and selected.

Selective logging—the removal of valuable timber species—is often the earliest stage of deforestation in the Amazon, yet it is difficult to detect using conventional methods based solely on satellite imagery. Therefore, labeling of selective logging areas was conducted to enable detection using machine learning approaches. This dataset is also valuable for monitoring sustainably managed forest areas. Using multi-band data from Sentinel-2, image chips containing evidence of selective logging were identified and selected.

2025 - 2025

Mining Detection

OtherGeospatial Tiled ImagerySegmentation

Mining activities in the Amazon cause significant environmental degradation. To enable efficient detection of such activities, a segmentation dataset of mining areas was created using multi-band Sentinel-2 data.

Mining activities in the Amazon cause significant environmental degradation. To enable efficient detection of such activities, a segmentation dataset of mining areas was created using multi-band Sentinel-2 data.

2023 - 2025

Detection of Newly Constructed Roads

OtherGeospatial Tiled ImagerySegmentation

It is known that illegal deforestation tends to occur around newly (illegally) constructed roads. However, because there is limited information on roads built within the Amazon rainforest, a dataset was required to develop a model capable of automatically detecting roads from satellite imagery. Using Planet mosaic imagery, 1,000 image chips (256 × 256 pixels) were manually labeled. Based on this initial dataset, a preliminary model was trained and used to generate additional road data, which was then refined to produce a final dataset of 5,000 labeled image chips.

It is known that illegal deforestation tends to occur around newly (illegally) constructed roads. However, because there is limited information on roads built within the Amazon rainforest, a dataset was required to develop a model capable of automatically detecting roads from satellite imagery. Using Planet mosaic imagery, 1,000 image chips (256 × 256 pixels) were manually labeled. Based on this initial dataset, a preliminary model was trained and used to generate additional road data, which was then refined to produce a final dataset of 5,000 labeled image chips.

2022 - 2025

Cattle Detection

LabelimgImageBounding Box

In efforts to combat illegal deforestation in Brazil, it is necessary to count the number of cattle present in illegally deforested areas in order to seize grazing livestock as part of enforcement actions. To support this objective, an accurate labeled dataset of cattle was created. Annotations were conducted on drone images and video footage acquired during both the rainy and dry seasons in the Brazilian Amazon and the Cerrado regions.

In efforts to combat illegal deforestation in Brazil, it is necessary to count the number of cattle present in illegally deforested areas in order to seize grazing livestock as part of enforcement actions. To support this objective, an accurate labeled dataset of cattle was created. Annotations were conducted on drone images and video footage acquired during both the rainy and dry seasons in the Brazilian Amazon and the Cerrado regions.

2022 - 2025

Education

Work History