Abstract: [Objective] Aiming at the over-fitting problem caused by information redundancy from multi-source remote sensing data and their derived high-dimensional features, this study is to effectively pre-select the optimal feature combination to optimize the k-nearest neighbor (k-NN) for regional forest above-ground biomass (AGB) estimation.[Method] This study proposed a fast iterative features selection method for k-NN method (KNN-FIFS). This method iteratively pre-select the optimal features which determined by the minimum root mean square error (RMSE) between the measured forest AGB values and the k-NN estimates based on the leave-one-out (LOO) cross-validation. Based on KNN-FIFS, multi-source data, including Landsat-8 OLI and its vegetation indices, texture metrics, topographic factors, HV polarization of P-band synthetic aperture radar (SAR) data, and forest inventory data (P_HV), were used to estimate forest AGB over Daxing'an Mountain Genhe forest reserve located in Inner Mongolia. Afterwards, the model behaviors between KNN-FIFS and stepwise multiple linear regression (SMLR) method were compared.[Result] For KNN-FIFS method, the best configuration was that one with k of 3, the remotely sensed features using P_HV, second moment of 1^st and 2^nd short-wave infrared bands (S6,S7), homogeneity of 1^st short-wave infrared band (H6), correlation of coastal aerosol (Cr1), correlation of the near infrared (Cr5), dissimilarity of coastal aerosol (D1) and the enhanced vegetation index (EVI). This configuration generated the most accurate estimates with R²=0.77 and RMSE=22.74 t·hm^-2,which performed much better than SMLR with R²=0.53 and RMSE=32.37 t·hm^-2.[Conclusion] KNN-FIFS is a more suitable method for forest AGB estimation than SMLR. KNN-FIFS can efficiently select the optimal feature combination to estimate regional forest AGB by use of multi-source remote sensing data with high-dimensional information.

Key words: KNN-FIFS, feature selection, above-ground biomass(AGB)