Impact of sampling techniques on crop type mapping using multi-temporal composites from Harmonized Landsat-Sentinel images

Abstract

Crop type mapping provides key spatial information for agricultural monitoring, management, production, and food security. Satellite multi-spectral imagery and machine learning algorithms have been used for large-area crop-type mapping, which necessitates training sample selection for high-quality mapping results. However, questions persist regarding the optimal sampling techniques and sample size to employ during application development. The novelty of this study focuses on investigating how sampling techniques and sample sizes influence the accuracy and reliability of satellite-based crop type mapping. This topic is often overlooked by practitioners and can have a drastic influence on the resulted maps. This study investigated how different sampling techniques and sample sizes influence crop type mapping in the Yazoo – Mississippi Delta (YMD), a complex agroecological landscape of northwestern Mississippi, United States. We derived a set of temporal composites from 2022 Harmonized Landsat Sentinel-2 (HLS) images and utilized the Cropland Data Layer (from the United States Department of Agriculture National Agricultural Statistics Service) as the ground reference to train and validate the classification model. The influence of four sampling techniques (grid, random, stratified, and cluster techniques) and sample sizes (0.05%, 0.1%, 0.50%, and 1.00% of the total pixels of the HLS tile and a minimum size of 100 samples per class) was evaluated with Artificial Neural Network classification of six primary crops in the YMD, including corn, cotton, rice, sorghum, soybeans, and wheat. Our results demonstrated that the stratified sampling technique achieved the highest performance across all crop-type classes, with an average F1-score of 0.81. In contrast, the grid sampling technique yielded the lowest performance, with an F1-score of 0.63. The cluster sampling technique also presented good results, with an average F1-score of 0.79. Classes representing small areas in the YMD region, such as sorghum and wheat, were not adequately captured by the grid sampling technique (F1-scores ranging from 0 to 0.2) or the random sampling technique (F1-scores ranging from 0.10 to 0.65). Allocating the samples considering a proportion between the area represented by the classes, and increasing the number of samples improved the model’s performance. The highest accuracy was obtained using 0.05% and 1.00% of the total number of pixels in the image tile as training samples, with an average F-score ranging from 0.81 to 0.84, respectively. When an equal number of samples was distributed per class, the results revealed significant mapping confusion among crop-type classes, particularly for other land use classes, which achieved an F1-score of 0. Our findings suggest that the optimal sampling technique involves allocating samples per image layer based on the proportional area of each class to determine the number of samples. Also, HLS-derived temporal composites are effective for crop type mapping, which may assist producers and stakeholders in more effective management of agricultural activities.

Citation

Aires, Uilson R.V.; Martins, Vitor S.; Ferreira, Lucas B.; Huang, Yanbo; Heintzman, Lucas; Ouyang, Ying. 2025. Impact of sampling techniques on crop type mapping using multi-temporal composites from Harmonized Landsat-Sentinel images. Computers and Electronics in Agriculture, 237: 110676. https://doi.org/10.1016/j.compag.2025.110676