This dataset was created using a large number of Earth remote sensing satellite images from Google Earth Engine. Firstly, collect training samples based on prerequisite knowledge, existing land cover baseline maps, and time weighted dynamic time distortion method. Secondly, the multi-year phenological features obtained from time-series land satellite images will be input into a random forest classifier to obtain the annual cultivated land probability for each pixel. Thirdly, the LandTrendr change detection algorithm is used to divide the probability time series of cultivated land into several segments, and record the breakpoints and corresponding change years within them. Fourthly, based on the established discrimination rules, determine the annual cultivated land type from the LandTrendr segmentation results. Finally, perform classification post-processing to better improve the farmland map.
Landsat TM/ETM+/OLI (Landsat 5/7/8) (1986-2021).
Land Cover Dataset (CLCD) (2020 and 2021 data).
The Space Shuttle Radar Topography Mission (SRTM) digital elevation dataset.
The trajectory based method combines machine learning and change detection techniques to draw annual dynamic maps of farmland. The annual cultivated land in this study is defined as a piece of land of at least 0.25 hectares (with a minimum width of 30 meters) that has been sown/planted and harvested at least once within 12 months after the sowing or planting date. This definition is consistent with the standards developed by the Joint Experiment on Crop Assessment and Monitoring (JECAM) network and adopts a shared arable land scope that conforms to the FAO Land Cover Meta Language. In this study, a key criterion for identifying annual farmland is that vegetation signals in remote sensing images must exhibit significant changes within 12 months, reflecting planting and harvesting activities.
The data processing process is as follows: (1) training data generation; (2) Feature space construction; (3) Slope probability estimation; (4) Time division; (5) Annual farmland surveying and mapping; (6) Post classification processing; (7) Accuracy assessment and comparison.
This study proposes an economically efficient and high-resolution dynamic monitoring scheme for farmland by combining automatic training sample generation, random forest supervised classification, and LandTrendr time segmentation algorithm. Using the complete archive of Landsat imagery and the GEE cloud computing platform, we innovatively created annual agricultural distribution maps of China from 1986 to 2021 at a resolution of 30 meters. The CACD annual map obtained from this has achieved an encouraging accuracy F1 score of 0.79 ± 0.02, which is superior to other products such as CLCD, CLUD, GLAD, and GFSAD. In addition, the validation of third-party sample sets, regression with provincial statistical data, and spatial detail comparisons between multiple products demonstrate the rationality of CACD in depicting the spatial distribution and temporal trends of farmland dynamics. In 2021, the total cultivated land area in China was 1725200 ± 212400 square kilometers, representing an increase of 30300 square kilometers (1.79%).
This work is licensed under a
Creative
Commons Attribution 4.0 International License.
CSTR:11738.11.NCDC.ZENODO.DB6649.2024
10.5281/zenodo.7936885