Abstract
The need for effective crop monitoring in large geographical scales has become increasingly important in recent years and constitutes a technological and scientific challenge for remote sensing applications. In Europe, member states of the European Union collect geospatial data in the framework of the Land Parcel Information System (LPIS) for agricultural management and subsidizing farmers. These data can be exploited as training datasets of machine learning classifiers for crop-type mapping applications. However, the way the LPIS data are being generated, concerning primarily errors in the farmers’ declarations in terms of crop-type labels, exact geometries, etc, constrains their direct use in such classification frameworks. In this study, we present and assess a methodology for filtering LPIS data based on geometric and spectral criteria in order to build a trustworthy training dataset for machine learning crop-type classifiers. A new nomenclature was developed, oriented towards the spectral discrimination of crop-type classes and sub-classes in Greece. The filtering methodology was applied at national scale for the agricultural year of 2019 and resulted in the selection of a sub-set of the LPIS parcels that were assessed as the most suitable and reliable to represent the different levels of the nomenclature. The developed filtering procedure was validated against actual crop-type labels derived from field visits, while the resulted filtered data were successfully utilized on various crop-type mapping experiments in Greece.
Introduction
Accurate crop-type mapping is essential for agricultural monitoring and policy enforcement. While the European Union's Land Parcel Information System (LPIS) offers millions of labeled samples, the data is inherently "noisy." Since only a small fraction of declarations are physically verified, the database often contains incorrect labels and imprecise boundaries. This paper introduces a robust pipeline to "clean" the LPIS dataset, creating high-quality training sets for machine learning by applying geometric constraints and multitemporal spectral filtering using Sentinel-2 data.
Methodology
The researchers developed a two-stage filtering pipeline applied at a national scale across Greece.
Nomenclature Refinement and Geometric Filtering
The initial Greek LPIS system, containing over 2,700 sub-categories, was reorganized into a hierarchical three-level system. Level 1 consists of 25 broad, spectrally distinct categories (e.g., Olive Groves, Cereals, Cotton) to facilitate better discrimination by machine learning models. To ensure spatial data quality, two main geometric constraints were applied: a 15-meter internal buffer to remove "edge pixels" and shape constraints to eliminate very small or overly elongated parcels.
Spectral Filtering
The core innovation of the study is the multitemporal spectral filter. Using Sentinel-2 NDVI (Normalized Difference Vegetation Index) time series, the researchers calculated a "median" growth profile for every crop type within each satellite tile.
Parcels were discarded if their growth curves deviated significantly from the median. The system employed "Smart Thresholding," using stricter standard deviation limits for over-represented crops (like Olives) and wider limits for rarer crops to maintain a balanced dataset. This "typicality" assumption posits that a parcel declared as a specific crop should follow the characteristic phenological cycle of that crop in its specific region.
Datasets
The study utilized the following data sources:
- LPIS Greece (2019): An initial dataset of over 6.1 million parcel entries.
- Sentinel-2 L2A: Multitemporal satellite data accessed via Google Earth Engine to construct NDVI profiles.
- Validation Data: A "ground truth" set of 9,635 points derived from Rapid Field Visits (RFVs) by the Greek Paying Agency (OPEKEPE) and independent laboratory field campaigns.
Results and Discussion
The filtering process significantly improved the quality of the training data:
- Data Reduction: The pipeline refined the initial ~3 million geometrically cleaned parcels down to approximately 933,000 "trustworthy" parcels (32% of the original set).
- Accuracy Boost: Validation against field visits showed that "Accepted" parcels achieved an Overall Accuracy of 92.79%, compared to 91.76% for unfiltered data.
- F1-Score Improvement: The average F1-score increased from 78.35% to 82.49%.
- Filter Validation: Parcels rejected by the algorithm had a significantly lower average F1-score (69.27%), confirming that the filter successfully isolated problematic data points.
Conclusion
While the methodology is highly effective, some limitations remain. Small sample sizes for certain classes can make accuracy metrics less robust. Additionally, strict filtering may occasionally discard correctly labeled parcels that exhibit atypical growth patterns due to unique local farming practices. However, this trade-off is necessary to build the "trustworthy" models required for operational crop monitoring. This pipeline is now positioned for integration into automated frameworks for annual, large-scale crop mapping.
Summary updated