基于Sentinel-2的农安县2019~2025年主要农作物分类及种植格局变化

Main Crop Classification and Planting Pattern Changes in Nong’an County from 2019 to 2025 Based on Sentinel-2

摘要: 准确掌握县域作物种植结构及其连续变化，对粮食生产管理与种植结构调整评估具有重要意义。本研究以吉林省长春市农安县为研究区，基于Google Earth Engine平台，利用2019~2025年Sentinel-2 SR Harmonized影像构建21个特征变量，以2025年1636个样点训练随机森林分类器(100棵树)，经5折分层OOF交叉验证，并结合Dynamic World年度耕地掩膜，逐年提取水稻、玉米、大豆和其他作物分布。结果表明：模型总体精度88.5%、Kappa系数0.85；玉米占比由56.8%升至70.9%，其他作物占比由37.8%降至23.8%，大豆和水稻分别稳定在3%~5%和2%左右；玉米净增829.11 km²，其他作物净减444.01 km²，“其他作物转为玉米”(658.11 km²)与“大豆转为玉米”(55.30 km²)为主要转换路径。在稳定耕地分析区内，40.4%的像元作物类型未发生变化，连续7年玉米面积498.30 km²、占稳定农田的79.4%，玉米与大豆相关轨迹合计仅占3.6%。研究表明，Sentinel-2多时相特征与随机森林可实现县域作物连续制图；研究期内农安县种植结构呈玉米占比上升、其他作物占比下降，种植格局稳定性较高。

Abstract: Accurate information on county-level crop planting structure and its continuous changes is of great significance for grain production management and for evaluating planting-structure adjustment. Taking Nong’an County, Changchun City, Jilin Province, China as the study area, this study, based on the Google Earth Engine platform, used Sentinel-2 SR Harmonized imagery from 2019 to 2025 to construct 21 feature variables, trained a random forest classifier (100 trees) with 1636 samples from 2025, evaluated it through 5-fold stratified out-of-fold (OOF) cross-validation, and, combined with the annual Dynamic World cropland mask, extracted the distributions of rice, maize, soybean and other crops year by year. The results showed that the overall accuracy of the model was 88.5% with a Kappa coefficient of 0.85; the proportion of maize rose from 56.8% to 70.9%, while that of other crops declined from 37.8% to 23.8%, with soybean and rice remaining stable at about 3%~5% and 2%, respectively; maize net increased by 829.11 km² and other crops net decreased by 444.01 km², and “other crops to maize” (658.11 km²) and “soybean to maize” (55.30 km²) were the major conversion pathways. Within the seven-year common valid area, 40.4% of the pixels did not change in crop type; continuous seven-year maize covered 498.30 km², accounting for 79.4% of the stable farmland, and maize–soybean related trajectories accounted for only 3.6%. The study indicates that multi-temporal Sentinel-2 features combined with random forest can achieve continuous county-level crop mapping; during the study period, the planting structure of Nong’an County showed a rising share of maize and a declining share of other crops, with relatively high stability of the planting pattern.