Drone Imagery for Crop Type Classification in Rwanda

RTI International generated 2,611 labeled point locations across Rwanda's agroecological zones. This dataset provides labels for three crop types (Banana, Maize, Legume) and other land cover, with corresponding 4.5m x 4.5m image chips from three time periods in late 2018 and early 2019. It was used in research documented by Chew et al. in 2020.

Use Cases

• Classify land cover type from latitude, longitude, and corresponding image chips using convolutional neural networks.
• Analyze temporal changes in crop patterns using image chips labeled P1 (Dec 2018), P2 (Jan 2019), and P3 (Feb 2019).
• Assess model confidence by incorporating the labeler's confidence degree attribute during training or evaluation.
• Train models for specific crop identification (Banana, Maize, Legume) versus non-crop types (Forest, Structure, Other).

Strengths

• 2,611 labeled point locations provide a substantial sample for model training.
• Covers 19 original land cover types, reduced to 6 distinct classes for focused analysis.
• Data is clustered across 5 distinct agroecological zones within Rwanda, offering environmental variety.

Limitations

• The dataset is limited to a single country (Rwanda) and a narrow three-month time range, limiting geographic and temporal generalizability.
• Without column details, the exact structure and available metadata for each point are unclear.
• Sample size of 2,611 points may be limited for training very large deep learning models without augmentation.

Provenance

Source

RTI International, via MLHUB and NASA Earthdata.

Collection Method

Labeled point locations generated from drone (UAV) imagery, with labeler confidence recorded.

Time Range

December 2018 to February 2019.

Freshness

Last updated in 2023, but the core imagery and labels are from 2018-2019.

Geography

Rwanda, across 5 agroecological zones.