A Lightweight Pipeline for Crop Time-Series Parcel Classification via Self-Supervision

This study proposes and benchmarks a lightweight pipeline for large-scale parcel-level crop classification using time-series data. The key contributions include a scalable SITS download strategy, a novel parcel-level crop classification dataset for the USA States of Texas and California, and a benchmark of classification techniques tested under varied data availability scenarios, highlighting the pipeline’s potential for enhancing agricultural monitoring systems. The methodology extracts basic descriptive statistics integrating with agricultural and cloud filtering, leveraging field boundaries delineation and USDA Cropland Data Layer datasets. Experimental results show that pretrained and fine-tuned models, such as SITS-BERT, outperform Random Forest and Support Vector Machine approaches, achieving an F1 score of 98.2% and overall accuracy of 99.3% on the Texas dataset for the more abundant training data scenarios. The pipeline has a projected computational speed-up of at leastat least 2, 500× compared to pixel-based methods for the tested datasets. The download pipeline is available on AgriGEE.lite library in https://pypi.org/project/agrigee-lite. The full benchmark results and related code are available at https://github.com/mateuspinto/light-crop-classification.