Transfer Learning for Crop classification with Cropland Data Layer data (CDL) as training samples.

Area of Science:

• Earth and Planetary Sciences
• Remote Sensing
• Agricultural Science

Background:

• Acquiring ground-truth training samples for crop mapping using remotely sensed data is challenging in many regions.
• This limitation hinders accurate crop classification, particularly in data-scarce areas.

Purpose of the Study:

• To propose and evaluate a transfer learning (TL) workflow for crop classification in regions lacking sufficient training data.
• To leverage crop phenology patterns from existing datasets (Contiguous U.S.A.) for improved global crop mapping.

Main Methods:

• Utilized high-confidence pixels from the Cropland Data Layer (CDL) in the U.S.A. as training samples.
• Generated 30-m, 15-day composited NDVI time series from harmonized Landat-8 and Sentinel-2 (HLS) data.
• Trained Random Forest (RF) classification models and applied them to test regions in China (Hengshui) and Canada (Alberta), comparing with local training (LO).

Main Results:

• Transfer learning achieved high overall classification accuracies: 97.79% (HS), 86.45% (AB), and 94.86% (NE) using full growing season NDVI time series.
• Local training (LO) provided higher accuracies earlier in the season compared to TL.
• TL showed potential for similar crop growth environments but faced challenges with spectrally similar crops early in the season.

Conclusions:

• Transfer learning offers a promising approach for crop classification in data-limited regions by utilizing established models.
• The study demonstrates the feasibility of applying models trained on U.S. data to international regions with similar crop growth patterns.
• Optimizing TL for early-season classification and addressing spectral confusion between crops are areas for future research.