基于KNN-FIFS的内蒙古根河森林郁闭度遥感估测研究

为探索国产高分一号宽幅（GF-1 Wide Field of View，GF-1 WFV）数据以及具有宽覆盖、红边波段（Red-Edge band，RE）的高分六号（GF-6）卫星数据在森林郁闭度（Forest Canopy Closure，FCC）定量反演中的潜力，本研究以GF-1 WFV多光谱数据为基础，添加哨兵2号（Sentinel-2A）红边波段，模拟GF-6红边波段特性，并提取相关纹理信息（Texture Information，TI）、植被指数（Vegetation Index，VI）和红边指数（Red- edge Index，RI），同时添加太阳入射角的余弦值cosi和1/cosi进一步探究了地形因素（Topographic Factors，TF）对FCC估测的影响，利用快速迭代特征选择的k-NN（k-Nearest Neighbor with Fast Iterative Features Selection，KNN-FIFS）模型，实现了内蒙古大兴安岭根河研究区FCC的定量反演，并对比逐步多元线性回归（Stepwise Multiple Linear Regressions，SMLR）和支持向量机（Support Vector Machine，SVM）估测结果。通过44块调查样地实测数据验证发现：基于GF-1 WFV估测的FCC与实测数据具有很好的一致性，R²=0.52，RMSE=0.08；GF-1 WFV+VI+TI估测结果为R²=0.56，RMSE=0.08；GF-1 WFV+RE+RI+TI的精度明显提高，R²=0.63，RMSE=0.07；GF-1 WFV+RE+RI+TI+TF的精度最高，R²=0.68，RMSE=0.07，并高于SMLR（R²=0.39，RMSE=0.10）和SVM（R²=0.49，RMSE=0.10）方法。KNN-FIFS方法比SMLR和SVM方法更适用于FCC遥感估测，且添加红边信息经地形校正后，能有效提高FCC的估测精度。

Estimation of Forest Canopy Closure by the KNN-FIFS Method in the Genhe of Inner Mongolia

Aiming at exploring the potentials of Gaofen-1 (GF-1) WFV data and Gaogen-6 (GF-6) satellite data in quantitative inversion of Forest Canopy Closure (FCC), based on GF-1data, the simulated GF-6 data by adding two Sentinel-2A red-edge bands (RE) into GF-1 WFV multispectral data and the extracted relevant Texture Information (TI) ,Vegetation Index (VI) and Red edge Index (RI) , a k- Nearest Neighbor with Fast Iterative Features Selection( KNN-FIFS) method, was used to estimate the Forest Canopy Closure (FCC) in the Genhe of the Great Khingan, Inner Mongolia. Besides that, the impact of terrain was further explored by adding Topographic Factors (TF) into the feature compositions. The verification using 44 field samples and the Leave-One-Out (LOO) method showed that: FCC estimation based on GF-1 WFV is in good agreement with measured data, with R² = 0.52, RMSE = 0.08; the GF-1 WFV+VI+TI’s has R² = 0.56, RMSE = 0.08; the GF-1 WFV+RE+RI+TI’s has been significantly improved with R²=0.63 and RMSE=0.07; and highest accuracy from the GF-1 WFV+RE+RI+TI+TF composition with R²=0.68 and RMSE=0.07 was superior to the results from both stepwise multiple linear regressions (SMLR) (R²=0.39, RMSE=0.10) and support vector machine (SVM) (R²=0.49, RMSE=0.10) methods. It indicated that the KNN-FIFS method is more reliable for FCC estimation than both SMLR and SVM methods, and the simulated GF-6 data with red-edge information can effectively improve the estimation accuracy of FCC, especially after topographic correction.