研究目的
To develop a semi-automatic methodology for joint land cover and crop type mapping with a highly detailed nomenclature of over 40 classes using multitemporal Landsat-8 data and SVM classification.
研究成果
The proposed semi-automatic methodology using SVM and multitemporal Landsat-8 data achieved high accuracy (OA 91.2%, kappa 90.6%) for joint land cover and crop type mapping with 47 classes. Water-related and single crop sub-type classes showed very high accuracy, while mixed and heterogeneous classes had lower performance. The approach is efficient but benefits from higher resolution data for improved results in complex areas.
研究不足
The classification accuracy was lower for mixed classes (e.g., urban green, mixed forest) and heterogeneous crop types due to the 30m spatial resolution of Landsat-8 data, which may not fully resolve fine details. Future work could use higher resolution data like Sentinel-2 and apply the methodology to more study areas.
1：Experimental Design and Method Selection:
The methodology employed a linear SVM classifier from the LIBSVM library for classification, using multitemporal surface reflectance Landsat-8 data. The rationale was to leverage machine learning for accurate mapping with a detailed class nomenclature.
2：Sample Selection and Data Sources:
The study area was northern Greece, covering 26,000 km2, selected for its heterogeneous landscapes. Reference data were produced through intensive manual annotation by experts using Landsat-8, Sentinel-2 images, Google Earth data, CLC2012 product, and C.A.P. geospatial data. Data were split into 70% for training and 30% for testing.
3：List of Experimental Equipment and Materials:
Landsat-8 satellite data (surface reflectance products), Sentinel-2 images, Google Earth imagery, CLC2012 product, C.A.P. geospatial data, and LIBSVM library for SVM implementation.
4：Experimental Procedures and Operational Workflow:
Acquired Landsat-8 data for 2016 from USGS EarthExplorer. Performed manual annotation to create reference polygons. Applied SVM classification with spectral bands and indices on a 16-dates multitemporal cube. Evaluated using confusion matrices and accuracy metrics.
5：Data Analysis Methods:
Calculated Overall Accuracy (OA), User's Accuracy (UA), Producer's Accuracy (PA), Kappa coefficient, and F-measure (F1) scores per class from confusion matrices.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
