Abstract

Crop monitoring is an important aspect of agricultural management, as it provides insights into cultivated area, crop health, growth patterns, and yields potential. Mapping cultivated areas and identifying crop types was historically conducted through field surveys and manual mapping, which are time-consuming and labor-intensive. Remote sensing classification has transformed large-scale land cover mapping, including crop identification. This work aims to: (1) compare the performance of Dynamic Time Warping (DTW) and two machine learning families (artificial neural networks and decision trees) for crop classification using Sentinel-2 data; (2) assess whether reflectance data, spectral indices, or both yield better classification results; and (3) evaluate the effect of hyperparameters on model performance. Among the DTW variants evaluated, dynamic time warping without a time constraint performed the best, with an overall accuracy of 0.921 using the combination of both reflectance and spectral indices. Most machine learning methods outperformed DTW. Although the convolutional neural network reached the highest single accuracy (0.948), the transformer was selected as the best model overall (accuracy of 0.944), as it combined a comparable accuracy with the lowest sensitivity to hyperparameter variations, making it a reliable option when testing machine learning architectures applied to crop mapping. This work also provides insights for model architecture development based on an exhaustive hyperparameter search for the machine learning models.

IPC Classification

Keywords