Spatial distribution information extraction of macadamia forest based on GF-2 remote sensing image

【Objective】 This study aimed to quickly and accurately obtain the spatial distribution information of macadamia forests based on GF-2 remote sensing image. It provided reference for the effective utilization of GF-2 remote sensing image to study the distribution of macadamia forests in the southwestern mountainous areas, as well as for the extraction of other land cover types in hilly and mountainous regions. 【Method】 The study area was located in Nansan Town, Zhenkang County, Lincang City, Yunnan Province. GF-2 image and digital elevation model（DEM） were used as data sources. An object-oriented approach was employed to extract 90 dimensional feature variables, including spectral, texture, shape and terrain features. Eight feature combination schemes（A1 to A8） were designed. The importance of the features was measured using the mean decrease in impurity（MDI） method, and the best feature combination was selected. Random forest（RF）, support vector machine（SVM）, and decision tree（DT） algorithms were used for the extraction of macadamia nut forests. The study explored the influence of different feature types and classification algorithms on the accuracy of macadamia nut forest extraction.【Result】Compared to the exhaustive segmentation parameter method, the combination of the scale parameter estimation（ESP） tool and the neighborhood difference absolute value and standard deviation ratio（RMAS） method was more efficient and objective in determining the optimal segmentation scale for specific land cover types. By comparing scheme A8 with scheme A7, it was found that adding terrain as a feature in scheme A8 reduced the overall feature dimensionality, particularly in the texture features, with only 4 texture features retained. The contribution of different feature types to the macadamia nut forest identification was ranked as follows: spectral features ＞ terrain features ＞ texture features ＞ shape features. In terms of classification algorithms, random forest outperformed support vector machine and decision tree in overall accuracy（OA）, user accuracy（UA）, producer accuracy（PA） and Kappa coefficient. Scheme A8, which integrated all features, achieved the best classification results, all higher than those of other schemes. Among the combinations of spectral, texture, shape, and terrain features, the random forest classification method achieved the best accuracy, with an OA of 95.8%, PA of 87.7%, and UA of 94.3%. The spatial distribution of macadamia nut forests showed that the largest plantation area was in the slope range of 15°-20°, covering 2.9 km². The macadamia nut forest area was mainly distributed in the southeast-facing slopes and at altitudes of 900-1200 m. 【Conclusion】 The combination scheme of terrain, texture, shape, and terrain after feature selection along with the random forest algorithm can effectively identify the distribution of macadamia nut forests. GF-2 remote sensing data and the object-oriented method have potential applications for mapping and resource monitoring of macadamia forests in the southern mountainous hilly regions and can be used for the identification of other land cover types in the region.