AlphaEarth embeddings evaluated for crop classification across US and India
Original title: Evaluating AlphaEarth Embeddings for Crop Classification: Separability, Temporal Dynamics, and Cross-Regional Transferability
64-dimensional embeddings separate similar crops better than NDVI or harmonics, but transfer worse from the US to India.
A representation-agnostic study benchmarks a 64-dimensional AlphaEarth embedding space against Sentinel-2 reflectance, NDVI phenology, and harmonic features for field-scale crop classification. Data cover the conterminous United States from 2017 to 2023 and northern India in 2021. Embeddings separate spectrally similar crops and multi-year rotations better, while harmonic and NDVI features transfer better from the US to India. A combined feature space is reported to balance both.
- Corn versus soybeans reaches distances of 3.9 for embeddings against 0.10.
- Rotation separability is 1.784 versus at most 1.003 for engineered features.
- US-to-India macro F1 is 0.493 for harmonics and NDVI versus 0.326.
- Mean residual spatial autocorrelation of errors is about Moran's I 0.17.
- Products
- Sentinel-2
- Countries
- United StatesIndia
- AI models
- AlphaEarth
From the original article
1 min read at arxiv.orgEarth Observation (EO)-based agricultural monitoring is shifting from pixel-level spectral analysis toward representation-driven modeling, yet systematic evaluation of learned representations across spatial, temporal, and domain-generalization dimensions remains limited.
By Priyanka Goyal · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.