Crop Classification Based on Temporal Information Using Sentinel-1 SAR Time-Series Data

Lu Xu, Hong Zhang, Chao Wang, Bo Zhang, Meng LiuRemote 2018 · DOI

TL;DR

This paper introduces a novel crop classification method using Sentinel-1 SAR time-series data that leverages temporal backscattering patterns. By generating multiple "temporal models" for each crop type and employing a spectral similarity measure (SSV), the approach achieves high classification accuracies, outperforming traditional machine learning classifiers.

What Problem Does This Paper Address?

Accurate and timely crop monitoring is crucial for global food security and land resource management. While Synthetic Aperture Radar (SAR) satellites like Sentinel-1 offer all-weather, all-time observation capabilities, effectively utilizing their increasing temporal resolution for crop classification remains a challenge. Previous methods often focused on polarimetric information or treated temporal SAR data as independent features, failing to fully capture the dynamic phenological changes of crops. This paper aims to address this gap by developing a method that explicitly models and matches the temporal backscattering evolution of different crops.

Methods

The proposed method consists of three main steps:

  1. Data Preparation: Standard SAR preprocessing is applied to Sentinel-1 IW mode dual-pol SLC images (VH and VV polarization) to ensure data quality. This includes coregistration, filtering, radiometric calibration, and geocoding.
  2. Temporal Model Generation: This is the core innovation. For each crop class, K-means clustering is used to group pixels with similar temporal backscattering intensity curves. The mean intensity vector of each cluster center is then defined as a "temporal model," representing a specific phenological pattern for that crop. The use of multiple models per crop accounts for variations in growth patterns due to factors like variety, sowing time, and local conditions.
  3. Temporal Model Matching: For each unlabeled pixel, its temporal intensity vector is compared against all generated temporal models using the Spectral Similarity Value (SSV) measure. SSV combines Euclidean distance (ED) and spectral correlation similarity (SCS) to assess the similarity between intensity curves. The pixel is then assigned to the class whose temporal model exhibits the highest SSV.

Datasets

The study utilizes two Sentinel-1 IW mode dual-pol SLC datasets:

  • Site 1: Wuqing District, Tianjin, China: Nine images acquired from July 2 to October 6, 2017 (DOY 183-279). Dominant crops include summer corn, soybean, rice, grass, and lotus.
  • Site 2: Fuyu City, Jilin Province, China: Thirteen images acquired from May 23 to October 26, 2017 (DOY 143-299), with one date missing. Dominant crops include corn, peanut, rice, and soybean.

Ground truth data for both sites was collected through field studies and visual interpretation of optical imagery.

Key Insights and Ideas

  • Temporal Models Capture Phenology: The core idea is that crop growth stages have distinct temporal backscattering signatures. By creating "temporal models," the method directly represents these dynamic changes, which are more informative than static features.
  • Multiple Models per Crop are Crucial: Recognizing that a single temporal model is insufficient to capture the diversity within a crop class (due to variations in farming practices, genetics, etc.), the K-means clustering approach generates multiple models, significantly improving robustness.
  • SSV is a Superior Similarity Measure: The paper demonstrates that the SSV measure, adapted from hyperspectral analysis, is more effective than other spectral similarity measures (like ED, SCS, SAM, SID) for matching temporal SAR intensity curves, particularly for the VV band.
  • VH vs. VV Band Performance: The VH band generally shows better discrimination due to its sensitivity to volume scattering, which is more directly related to biomass. However, the SSV measure helps achieve comparable results in the VV band, which is more complexly influenced by scattering mechanisms.

Key Results

  • High Overall Accuracies: The proposed SSV-based temporal model matching method achieved the best overall accuracies (OA) and kappa coefficients (KA) in both VH and VV bands across both study sites, generally exceeding 90%.
  • Outperforms DT and NB: Compared to Decision Tree (DT) and Naive Bayes (NB) classifiers, the proposed method consistently achieved higher or comparable accuracies for most crops.
  • Site-Specific Models: Temporal models trained for one location were not directly transferable to another, highlighting the influence of local climate, acquisition conditions, and cultivation practices on SAR backscatter.
  • Corn Discrimination: The method showed excellent performance in discriminating corn, the dominant crop in both study areas, achieving high Producer's Accuracies (PA).

Limitations and Future Work

  • Data Dependency: The method's performance is closely tied to the number and temporal coverage of SAR images. A complete crop growth cycle observation is ideal.
  • Training Sample Selection: Like all supervised methods, the quality and representativeness of training samples are critical.
  • Generalization: Temporal models are location-specific. Future work will investigate the generalization of temporal models and explore incorporating polarimetric features for a joint model.
  • Climate and Environmental Factors: The authors suggest relating temporal models to climate factors (seed variety, latitude, altitude, irrigation, temperature) to characterize crop growth details more comprehensively and potentially build a temporal model library for yield estimation and market price discovery.

Summary updated

Abstract

With the increasing temporal resolution of space-borne SAR, large amounts of intensity data are now available for continues land observations. Previous researches proved the effectiveness of multitemporal SAR in land classification, but the characterizations of temporal information were still inadequate. In this paper, we proposed a crop classification scheme, which made full use of multitemporal SAR backscattering responses. In this method, the temporal intensity models were established by the K-means clustering method. The intensity vectors were treated as input features, and the mean intensity vectors of cluster centers were regarded as the temporal models. The temporal models summarized the backscatter evolutions of crops and were utilized as the criterion for crop discrimination. The spectral similarity value (SSV) measure was introduced from hyperspectral image processing for temporal model matching. The unlabeled pixel was assigned to the class to which the temporal model with the highest similarity belonged. Two sets of Sentinel-1 SAR time-series data were used to illustrate the effectiveness of the proposed method. The comparison between SSV and other measures demonstrated the superiority of SSV in temporal model matching. Compared with the decision tree (DT) and naive Bayes (NB) classifiers, the proposed method achieved the best overall accuracies in both VH and VV bands. For most crops, it either obtained the best accuracies or achieved comparable accuracies to the best ones, which illustrated the effectiveness of the proposed method.
BibTeX
@inproceedings{xu2018crop,
  author = {Lu Xu and Hong Zhang and Chao Wang and Bo Zhang and Meng Liu},
  title = {{Crop Classification Based on Temporal Information Using Sentinel-1 SAR Time-Series Data}},
  booktitle = {Remote},
  year = {2018},
  doi = {10.3390/rs11010053},
}
Figure