Crop Mapping without Labels: Investigating Temporal and Spatial Transferability of Crop Classification Models Using a 5-Year Sentinel-2 Series and Machine Learning

Abstract

Multitemporal crop classification approaches have demonstrated high performance within a given season. However, cross-season and cross-region crop classification presents a unique transferability challenge. This study addresses this challenge by adopting a domain generalization approach, e.g., by training models on multiple seasons to improve generalization to new, unseen target years. We utilize a comprehensive five-year Sentinel-2 dataset over different agricultural regions in Slovakia and a diverse crop scheme (eight crop classes). We evaluate the performance of different machine learning classification algorithms, including random forests, support vector machines, quadratic discriminant analysis, and neural networks. Our main findings reveal that the transferability of models across years differs between regions, with the Danubian lowlands demonstrating better performance (overall accuracies ranging from 91.5% in 2022 to 94.3% in 2020) compared to eastern Slovakia (overall accuracies ranging from 85% in 2022 to 91.9% in 2020). Quadratic discriminant analysis, support vector machines, and neural networks consistently demonstrated high performance across diverse transferability scenarios. The random forest algorithm was less reliable in generalizing across different scenarios, particularly when there was a significant deviation in the distribution of unseen domains. This finding underscores the importance of employing a multi-classifier analysis. Rapeseed, grasslands, and sugar beet consistently show stable transferability across seasons. We observe that all periods play a crucial role in the classification process, with July being the most important and August the least important. Acceptable performance can be achieved as early as June, with only slight improvements towards the end of the season. Finally, employing a multi-classifier approach allows for parcel-level confidence determination, enhancing the reliability of crop distribution maps by assuming higher confidence when multiple classifiers yield similar results. To enhance spatiotemporal generalization, our study proposes a two-step approach: (1) determine the optimal spatial domain to accurately represent crop type distribution; and (2) apply interannual training to capture variability across years. This approach helps account for various factors, such as different crop rotation practices, diverse observational quality, and local climate-driven patterns, leading to more accurate and reliable crop classification models for nationwide agricultural monitoring.

Transferability of Machine Learning Models for Crop Classification in Remote Sensing Imagery Using a New Test Methodology: A Study on Phenological, Temporal, and Spatial Influences

“…Referring back to the literature review, while the problems of transferability are known, they have not been systematically analyzed [24,25,[27][28][29][30]. Using the testing methodology presented here (Table 6), spatial transferability problems can now be systematically investigated in hopes of gaining insight into the causes and effects of transferability.…”

“…Furthermore, shifts in crop phenology, field characteristics, or ecological site conditions in the previously unseen area may reduce the classification performance of machine learning classifiers that often overfit to training sites [24]. Using Sentinel-2 data, high variability in transferability performance was found in the models [25]. Efforts have been made to transfer the models from regions with abundant in situ data to regions with limited in situ data using augmentation and transfer learning (TTL).…”