ellaampy/SpatioTemporalYield

On exploring spatio-temporal encoding strategies for county-level yield prediction

Crop yield information plays a pivotal role in ensuring food security. Advances in Earth Observation technology and the availability of historical yield records have promoted the use of machine learning for yield prediction. Significant research efforts have been made in this direction, encompassing varying choices of yield determinants and particularly how spatial and temporal information are encoded. However, these efforts are often conducted under diverse experimental setups, complicating their inter-comparisons.

The dataset SpatioTemporalYield is the data used in our comparative studies.

Data coverage

• The United States of America (USA) is the world’s largest producer of corn, accounting for approximately one-third of global production.
• SpatioTemporalYield covers the USA’s top five corn-producing states: Iowa, Illinois, Indiana, Nebraska, and Minnesota.
• Altogether, they accounted for over one-half of the USA’s corn(grain) production in 2021.

Structure of the data

The folder contains numpy arrays (over 8000) in the form YYYY_GEOID.npy and a single json file labels.json corresponding to their labels.

• YYYY - is the year of acquisition (from 2003 to 2021)
• GEOID - a five-character code representing a stateid (first two characters) and a county id (next three characters)
• Each .npy file has the structure T (time) x C (channel) x S (number of pixels)
• There are 46 sequences (observed from January to December) and 12 channels in each array.

The channels/features are in the following order

Order	Band	Source
Index 0	red	MOD9A1.061
Index 1	nir	MOD9A1.061
Index 2	blue	MOD9A1.061
Index 3	green	MOD9A1.061
Index 4	nir2	MOD9A1.061
Index 5	swir1	MOD9A1.061
Index 6	swir2	MOD9A1.061
Index 7	tmin	Daymet
Index 8	tmax	Daymet
Index 9	prcp	Daymet
Index 10	ndvi	MOD9A1.061 (derived)
Index 11	ndwi	MOD9A1.061 (derived)

Citation

If you use this data, please cite our work as:

TBD

Notes

The SpatioTemporalYield is a first version of our on-going initiative to create a multi-task and multi-sensory benchmark dataset for agricultural monitoring in the USA Please check back for updates.