Scholarlink

<p>Under rapidly digitalizing conditions, automation and scalability in agricultural monitoring and crop type mapping are increasingly critical. This study proposes a phenology-driven, automated ground-truth generation pipeline for large agricultural regions with widespread double cropping and rapidly changing phenological dynamics, exemplified by the Harran Plain. To reduce the time and cost burden of field-based manual data collection, the workflow uses farmer-declared parcels from the Turkish Farmer Registration System (FRS) as input and exploits multi-temporal Sentinel-2 observations. Parcel-level Enhanced Vegetation Index (EVI) time series are subjected to a multi-stage validation procedure to systematically eliminate outlier and anomalous samples. Phenological agreement is first quantified using Dynamic Time Warping (DTW). Spectral–temporal anomalies are then identified via a ridge-regularized Mahalanobis discordance score. Because distance-based thresholding can generate false positives, a safeguard layer jointly evaluates complementary metrics (band_cover, R², and max_run_out) to retain phenologically coherent parcels while removing true anomalies. Following the multi-stage curve-based cleaning, 4,496 of 13,048 parcels (34.46%) were discarded as anomalous, whereas 8,552 parcels (65.54%) were retained for classification. Classification experiments conducted using this clean phenological set achieved an overall accuracy of 93.6%. Performance was strong for single-crop classes (cotton F1 = 97.4; wheat F1 = 92.3). Although performance decreased for double-cropping classes where phenological transitions are less distinct, sowing/harvest schedules are more heterogeneous, and declaration-driven label noise is more pronounced, results remained acceptable (wheat-cotton F1 = 91.8; wheat-maize F1 = 90.3; lentil-cotton F1 = 85.7). Overall, the proposed pipeline enhances inter-class separability by filtering samples that violate phenological integrity and offers an operational alternative to extensive field campaigns at large spatial scales. Rather than discarding declaration data, the framework transforms it into analysis-ready reference information through phenological and spectral–temporal quality control, enabling direct integration into crop type mapping infrastructures while minimizing time and cost requirements and enabling consistent, repeatable labeling across seasons.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Phenology-Driven Automatic Ground Truth Generation from Farmer-Declared Parcels for Agricultural Crop Classification Using Dynamic Time Warping

  • 期刊论文

  • 发表日期：

    2026年7月7日

查看全文
  • Fatih Fehmi Şimşek

摘要

Under rapidly digitalizing conditions, automation and scalability in agricultural monitoring and crop type mapping are increasingly critical. This study proposes a phenology-driven, automated ground-truth generation pipeline for large agricultural regions with widespread double cropping and rapidly changing phenological dynamics, exemplified by the Harran Plain. To reduce the time and cost burden of field-based manual data collection, the workflow uses farmer-declared parcels from the Turkish Farmer Registration System (FRS) as input and exploits multi-temporal Sentinel-2 observations. Parcel-level Enhanced Vegetation Index (EVI) time series are subjected to a multi-stage validation procedure to systematically eliminate outlier and anomalous samples. Phenological agreement is first quantified using Dynamic Time Warping (DTW). Spectral–temporal anomalies are then identified via a ridge-regularized Mahalanobis discordance score. Because distance-based thresholding can generate false positives, a safeguard layer jointly evaluates complementary metrics (band_cover, R², and max_run_out) to retain phenologically coherent parcels while removing true anomalies. Following the multi-stage curve-based cleaning, 4,496 of 13,048 parcels (34.46%) were discarded as anomalous, whereas 8,552 parcels (65.54%) were retained for classification. Classification experiments conducted using this clean phenological set achieved an overall accuracy of 93.6%. Performance was strong for single-crop classes (cotton F1 = 97.4; wheat F1 = 92.3). Although performance decreased for double-cropping classes where phenological transitions are less distinct, sowing/harvest schedules are more heterogeneous, and declaration-driven label noise is more pronounced, results remained acceptable (wheat-cotton F1 = 91.8; wheat-maize F1 = 90.3; lentil-cotton F1 = 85.7). Overall, the proposed pipeline enhances inter-class separability by filtering samples that violate phenological integrity and offers an operational alternative to extensive field campaigns at large spatial scales. Rather than discarding declaration data, the framework transforms it into analysis-ready reference information through phenological and spectral–temporal quality control, enabling direct integration into crop type mapping infrastructures while minimizing time and cost requirements and enabling consistent, repeatable labeling across seasons.