鲁中南山丘区耕地地力的遥感反演模型与应用

耕地地力是耕地生产能力的重要表征，地力的遥感快速准确反演是耕地资源利用管理的客观需求。该研究针对鲁中南山丘区，选择东平县和滕州市2个代表性县市，利用东平县的TM影像构建与筛选光谱指标，通过经典统计分析（一元线性回归、曲线回归、多元逐步线性回归）和机器学习（BP神经网络、极限学习机）方法构建与优选反演模型，进而在滕州市进行模型的验证和应用，同时对不同时相的反演模型进行了比较分析。结果表明：5类光谱指标与耕地地力综合指数均有显著的相关性，其中改进型光谱指数的r均>0.684，能更好地反映耕地地力状况；最佳反演模型为经典统计分析方法中的改进指数组-多元逐步线性回归（IIG-MLSR）模型（R2=0.684，RMSE=5.674）和机器学习算法的改进指数组-BP神经网络（IIG-BPNN）模型（R2=0.746，RMSE=5.089）；模型在山丘区具有较好的普适性，地力反演与评价的结果具有相似的空间分布特征、相近的耕地面积比例和较高的空间契合度。其中2个最佳模型的高中低3级耕地的面积比例差普遍低于5.55个百分点，空间契合度分别为84.50%和88.76%；模型动态反演分析结果显示，2007—2016年滕州市耕地地力不断提升，高级地由67.30%增加至80.72%；不同时相模型比较结果显示，多时相遥感反演耕地地力具备可行性，4月份冬小麦返青拔节期是反演的最佳时相，10月份裸土时相次之，8月份夏玉米时相最差。该研究提出了山丘区耕地地力快速定量遥感反演的有效方法，对完善遥感反演指标与模型，提高评价效率有参考价值。

Remote sensing inversion and application for soil fertility of cultivated land in the hilly areas of central-south Shandong of China

Soil fertility of a cultivated land is an important indicator of cultivated land productivity. It is necessary to obtain the rapid and accurate inversion of cultivated land fertility via remote sensing for the better utilization and management of land resource. In this study, a new inversion model was constructed and optimized using the classical statistical analysis (SLR, CR, and MLSR), and machine learning (BPNN and ELM). An effective way was also proposed for rapid quantitative remote sensing inversion of cultivated land fertility in hilly areas. Dongping County and Tengzhou City were selected as two representative counties and cities in the hilly area of center southern Shandong Province, China. In Dongping County, the TM image during the turning green and jointing stage was used to construct and screen spectral indexes of cultivated land fertility. Tengzhou City was selected to verify the spatial universality of inversion model for the soil fertility of a cultivated land. Furthermore, the remote sensing inversion model was used to quantitatively monitor the spatial-temporal dynamic status of cultivated land fertility in Tengzhou City in 2007, 2011, and 2016. The prediction accuracy of inversion models was compared in different periods. The results showed that there were significant correlations between the five kinds of spectral indexes in a remote sensing and the integrated fertility index (IFI), among which the correlation coefficients of improved spectral index were greater than 0.684, indicating better reflecting the status of cultivated land fertility. The best inversion model was the IIG-MLSR model (Rv2=0.684, RMSE=5.674) in the classical statistical analysis, while, the IIG-BPNN model (Rv2=0.746, RMSE=5.089) in the machine learning. The obtained model demonstrated excellent universal applicability in hilly areas, where there were similar spatial distribution characteristics between the inversion and evaluation on the cultivated land fertility, and the similar proportion of cultivated land and high spatial compatibility. In the two best models, the difference in the area ratio of the high, middle, and low levels of cultivated land fertility inversion and cultivated land fertility evaluation was generally less than 5.55%, where the spatial fit was 84.50% and 88.76%, respectively. The dynamic inversion analysis showed that the cultivated land fertility of Tengzhou City increased continuously in recent 10 years (from 2007 to 2016). The area proportion of high-level land increased from 67.30% to 80.72%, whereas, that of middle-level and low-level land decreased. The multi-temporal remote sensing inversion of cultivated land fertility was feasible, compared with the remote sensing inversion models in different time periods. The optimal time phase to invert the cultivated land fertility was the turning green and jointing stage of winter wheat in April, followed by bare soil in October, and the worst in summer maize in August. The remote sensing inversion index and model can be used to effectively increase the evaluation efficiency of cultivated land fertility. At the same time, this finding can provide a positive reference for the related research of cultivated land quality.