Robust and timely within-season conterminous United States crop type mapping using Landsat Sentinel-2 time series and the transformer architecture

Abstract

The timeliness and accuracy of existing methods for crop type mapping within the growing season (i.e., within-season crop type mapping) are often limited by their ability to fully utilize dense time series observations, for example, because they use monthly composites. This study presents a novel and robust methodology that leverages every single good-quality observation in the time series and a trained model capable of achieving improved mapping accuracy using 30 m NASA Harmonized Landsat-8 and Sentinel-2 (HLS) data for any time within-season across the conterminous U.S. (CONUS). A single pixel time series Transformer-based deep learning model was trained using seven years (2016–2022) of United States Department of Agriculture (USDA) Cropland Data Layer (CDL) 30 m crop classifications, with > 1 million training pixels systematically sampled across the CONUS. The model was trained using only HLS data as input predictors to classify 37 crop types. The model can handle the irregular HLS time series without the need for temporal compositing by using a fixed-length time series input to accommodate sequences of any length, with time positions lacking good-quality observations masked out during classification via the Transformer's attention masking mechanism. The model was trained using different early growing season HLS time series portions defined up to a randomly selected end date each year and randomly removing some observations to account for variations in time series acquisition dates among years and locations. CONUS results for 2023 derived by classifying the HLS time series with progressively more growing season HLS data were presented and validated against the 2023 CDL. F1-scores > 0.8 were achieved for 15 crop classes by June 30, 20 classes by July 31, and 23 classes by August 31, 2023. The major commodity crop winter wheat could be classified with F1-scores > 0.8 by Jan 6th, 2023. The five major summer commodity crops could be classified with F1-scores > 0.8 by May 11th (corn), May 26th (cotton and rice), May 31st (soybean), and Jun 25th (spring wheat), 2023. The model achieved similar accuracy one month earlier than the USDA In-season Cropland Data Layer (ICDL) product. July 2023 regional crop area estimates for nine central U.S. agricultural states showed higher consistency with the end-of-season CDL (0.96 ≤ R² ≤ 1.00) than those from the July ICDL (0.86 ≤ R² ≤ 0.96). The model's efficiency, as well as potential future improvements in accuracy and efficiency are discussed. The training data, application codes, and trained model, are publicly available.