JOINT LAND COVER AND CROP TYPE MAPPING USING MULTI-TEMPORAL SENTINEL-2 DATA FROM VARIOUS ENVIRONMENTAL ZONES IN GREECE

Christina Karakizi, Z. Kandylakis, A. D. Vaiopoulos, Κωνσταντίνος Καράντζαλος2021 · DOI

Abstract

In this work, we elaborate on the gained insights from various classification experiments towards detailed land cover mapping over four representative regions of different environmental characteristics in Greece. In particular, the proposed methodology exploits Sentinel-2 data at an annual basis, for the joint classification of 35 land cover and crop type classes. A number of pre-processing steps were employed on the satellite data, in order to address atmospheric and geometric effects, as well as clouds and pertinent shadows. Several classification set-ups were designed and performed using either time series of spectral features or temporal features. The latter consisted of statistical metrics, derived from the spectral time series, and therefore were significantly reduced in dimension. Experiments using the Random Forest algorithm were performed by building several per-tile models, as well as cross-regional models based on training data from all considered regions/tiles. Overall classification accuracy rates exceeded 90% for most experiments. Further analysis on the experimental results highlighted that crop types were classified more accurately when using the spectral time series features, compared to the temporal ones. Classification accuracy for non-crop classes proved much less affected by the type of employed features. The inclusion of auxiliary data layers was beneficial in all cases, both for overall and for per-class accuracy metrics. Qualitative evaluation on the predicted maps further affirmed the efficiency of the developed methodology.

Introduction

Mapping land cover is a staple of remote sensing, but doing so accurately across large, environmentally diverse regions remains a challenge. This research addresses two primary hurdles in the field:

  1. The Dimensionality Curse: Multi-temporal satellite data creates massive datasets that are computationally expensive to process.
  2. Phenological Complexity: Crops change color and structure throughout the growing season. Mapping 18 different crop types alongside 17 general land cover classes (like "dense urban fabric" or "coniferous forest") requires a model that can distinguish between a field of maize and a field of sunflowers, even if they look similar at a single point in time.

The study aims to find the best balance between data richness and computational efficiency while testing how well models trained in one region generalize to others.

Methods

The researchers employed a Random Forest (RF) classifier, a reliable "workhorse" in remote sensing due to its ability to handle high-dimensional data and provide feature importance rankings. They tested two distinct feature extraction approaches:

  • Spectral Time Series (TS): Using raw reflectance values from 10 spectral bands plus three indices (NDVI, NDWI, NDBI) across all available dates in 2019, resulting in up to 350 features per pixel.
  • Temporal Features (TF): Statistical metrics (mean, median, standard deviation, and various percentiles) for each band over the year, compressing the data into just 90 features per pixel.

Additionally, they integrated auxiliary data, including digital elevation models (EU-DEM) and existing thematic layers (Corine Land Cover 2018), to provide the model with context like altitude and historical land use.

Datasets

The study focused on four Sentinel-2 tiles covering approximately 47,000 km² of Greece, representing three distinct climatic zones: Mediterranean North, Mediterranean South, and Mediterranean Mountains.

Study area map showing the four Sentinel-2 tiles across different environmental zones in Greece.
  • Satellite Data: Sentinel-2 L1C imagery from 2019 with less than 20% cloud cover.
  • Reference Data: A high-quality ground-truth dataset manually digitized using VHR (Very High Resolution) imagery from Google Earth, Bing, and PlanetScope, supplemented by the Greek Paying Agency’s Land Parcel Identification System (LPIS).
  • Nomenclature: A detailed 35-class system, including 18 specific crop types (e.g., cotton, rice, tobacco, olives, vineyards).

Key Results

The experimental results demonstrated high performance across the board, with overall accuracy (OA) exceeding 90% for most experiments.

Comparative performance metrics (OA, avF1, avF1crops, avF1non-crops) for the four study tiles under different feature sets.
  • Phenology is Key for Crops: A significant performance gap was observed between TS and TF for agricultural classes. While general land cover (forests, water, urban) was identified by both methods, crop types were significantly more accurate when using the full time series (TS). For crop-specific metrics (avF1crops), TS outperformed TF by up to 8%, suggesting that the timing of growth stages is a vital identifier that statistical averages tend to smooth over.
  • Context Matters: The inclusion of auxiliary data (elevation and thematic data) improved crop classification by 1-5% and non-crop classification by 3-8%. This helped the model differentiate between spectrally similar classes, such as natural grasslands versus agricultural fodder.
  • The Power of Aggregation: Creating a single cross-regional model using data from all four tiles often outperformed per-tile models for rare classes by pooling data and learning a more generalized representation.

Conclusion and Future Work

While the methodology proved robust, sample size imbalance remains a hurdle. Rare classes—such as pome trees and legumes—had low sample sizes in certain tiles, leading to poor F1 scores. Future work could explore advanced deep learning architectures like LSTMs or Transformers, specifically designed to handle the sequential nature of time series data, to further improve the mapping of these difficult minority classes.

Summary updated

Abstract

In this work, we elaborate on the gained insights from various classification experiments towards detailed land cover mapping over four representative regions of different environmental characteristics in Greece. In particular, the proposed methodology exploits Sentinel-2 data at an annual basis, for the joint classification of 35 land cover and crop type classes. A number of pre-processing steps were employed on the satellite data, in order to address atmospheric and geometric effects, as well as clouds and pertinent shadows. Several classification set-ups were designed and performed using either time series of spectral features or temporal features. The latter consisted of statistical metrics, derived from the spectral time series, and therefore were significantly reduced in dimension. Experiments using the Random Forest algorithm were performed by building several per-tile models, as well as cross- regional models based on training data from all considered regions/tiles. Overall classification accuracy rates exceeded 90% for most experiments. Further analysis on the experimental results highlighted that crop types were classified more accurately when using the spectral time series features, compared to the temporal ones. Classification accuracy for non-crop classes proved much less affected by the type of employed features. The inclusion of auxiliary data layers was beneficial in all cases, both for overall and for per-class accuracy metrics. Qualitative evaluation on the predicted maps further affirmed the efficiency of the developed methodology.
BibTeX
@inproceedings{karakizi2021joint,
  author = {Christina Karakizi and Z. Kandylakis and A. D. Vaiopoulos and Κωνσταντίνος Καράντζαλος},
  title = {{JOINT LAND COVER AND CROP TYPE MAPPING USING MULTI-TEMPORAL SENTINEL-2 DATA FROM VARIOUS ENVIRONMENTAL ZONES IN GREECE}},
  booktitle = {˜The},
  year = {2021},
  doi = {10.5194/isprs-archives-xliii-b3-2021-319-2021},
}
Figure