基于时序影像的农业活动因子提取与闽西耕地SOC数字制图

人类活动对表层耕地土壤有机碳(Soil Organic Carbon, SOC)影响强烈，但目前大范围复杂地貌地形区的耕地SOC数字制图对人为因素的空间刻画不足。本文以福建省西部耕地为研究对象，基于Sentinel-2/MSI时间序列数据提取轮作模式分类信息(Crop Rotation, CR)，以及可反映轮作模式信息的植被特征变换变量(Harmonic Analysis of Time Series, HANTS)，分别作为农业活动定性和定量因子，将常规气候和地形因子作为自然环境因子，并对不同类型环境变量进行组合(气候+地形、气候+地形+轮作模式、气候+地形+HANTS变量、气候+地形+轮作模式+HANTS变量)。基于随机森林模型(Random Forest, RF)对不同环境变量组合驱动的耕地表层SOC空间预测精度进行对比分析，探索以轮作模式为例的农业活动因子提高耕地表层SOC数字制图精度的可能性。结果表明，同时加入两种农业活动因子的RF模型表现最佳，其模型预测精度相较于纯自然环境变量驱动的模型有明显提高(R²提高了89.47%,RMSE和MAE分别下降了10...

Digital SOC Mapping in Croplands Using Agricultural Activity Factors Derived from Time-Series Data in Western Fujian

Human activities significantly affect the amount and spatial variation of top Soil Organic Carbon (SOC) in croplands. However, the spatial distribution of agricultural management practices has not been carefully considered in SOC mapping in croplands, especially for croplands in large-scale complex landforms. A case study was conducted in agricultural area in western Fujian Province. Sentinel-2/MSI NDVI time series data were used to derive two types of variables that contains crop rotation information. One is the Crop Rotation (CR) pattern type, which was regarded as qualitative factors of agricultural activities. The other are variables generated using Harmonic Analysis based on sentinel-2 NDVI time series data (HANTS), which were regarded as quantitative factors of agricultural activities. Two types of agricultural activities factors, as well as natural environmental variables were adopted as predictive environmental variables. Four different combinations of above variables according to different categories were formed respectively (i.e., climate factor + terrain factor, climate factor + terrain factor + crop rotation pattern, climate factor + terrain factor + HANTS variables, and climate factor + terrain factor + crop rotation pattern + HANTS variables). Random Forest (RF) models were developed based on four different combinations of above variables for predicting SOC. These RF models were compared to explore whether incorporating agricultural activity factors could improve the SOC mapping accuracy in croplands. Results showed that the combination of natural environment variables with both crop rotation type and variables derived through HANTS yielded the highest accuracy. Compared with the combination of natural environment variables, the prediction accuracy of the optimal model was significantly improved (R² increased by 89.47%, RMSE and MAE decreased by 10.66% and 12.05%, respectively). Two types of agricultural activity factors were both adopted in optimal model, especially CR significantly affected the SOC in croplands, ranking fourth in the importance of environmental variables of the optimal model. In all RF models, annual rainfall (Rainfall) ranked first in the importance of environmental variables. This indicated that climate factors play a dominant role in soil organic carbon digital soil mapping. The SOC content in croplands of the region predicted from the optimal model was (18.22±2.99) g/kg on average and varied in the range of 8.25~30.69 g/kg. The SOC content in double cropping rice and tobacco-rice planting area were higher than that in rice-vegetable planting area. The results provide a new vision for updating the environmental variables of SOC mapping in complex landform areas.