應用多維遙感信息協同的森林樹種分類1)

謝天義 潘潔 孫玉琳 鄭光

(南京林業大學，南京，210037) (南京大學)

In order to explore the main remote sensing information affecting forest tree species classification, the difference of tree species classification based on different dimension information were analyzed. The research object is the forest tree species in the natural forest area of Panther Creek in the southeast of Portland, Oregon, United States. The principal component analysis and ant colony optimization method were used to select the feature bands, texture features and vegetation indexes of airborne hyperspectral images. The LiDAR data were used to extract the vertical structure parameters of forests, and the random forest method was used to classify tree species. The results showed that the average classification accuracy using one-dimensional information was 69.86%, of which the classification accuracy using texture features was the highest (77.40%), and the classification accuracy using vertical structure parameters was the lowest (62.44%). The average classification accuracy of two-dimensional information combination was 79.68%, and the classification accuracy of hyperspectral feature band and texture feature combination was the highest. The average classification accuracy of three-dimensional information combination was 85.00%, and the classification accuracy of hyperspectral feature band, texture feature and vegetation index was the highest. The classification accuracy of all four-dimensional information was higher than that of one-dimensional, two-dimensional and three-dimensional information, and the classification accuracy was 89.20%. It can be seen that collaborative multi-dimensional remote sensing information can effectively improve the classification accuracy of forest tree species, among which texture information and feature spectrum information extracted from hyperspectral data play an important role in forest tree species classification, while the collaborative vertical structure information further improves the accuracy of forest tree species classification.

森林是陸地生態系統的重要組成部分，對維持生態過程和維護生態平衡起著至關重要的作用[1-2]。森林樹種結構與分布則是森林資源可持續經營管理、森林生態系統服務評估以及生物多樣性維護的基礎[3-5]。

遙感作為目前森林資源調查的主要技術手段，在森林樹種分類研究中取得了諸多進展。其中，高光譜遙感在地物精細識別中的優勢為森林樹種分類精度的提高提供了可能[6]。Richter et al.[7]利用機載高光譜影像采用隨機森林和支持向量機的分類方法，對萊比錫河岸天然林進行樹種分類的精度為78.40%；申鑫等[8]利用LiCHy集成傳感器同期獲取的高空間分辨率和光譜分辨率數據，通過信息熵原理提取特征變量對4個典型樹種分類的總體精度達到62.90%。

高光譜遙感雖然提供了豐富的光譜信息，但僅限于冠層表面且存在“同譜異物和同物異譜”現象，在一定程度上影響了森林樹種識別的精度。激光雷達(LiDAR)作為一種新型的對地觀測技術，具有主動性、穿透性及可快速大面積地獲取三維信息等特點[9]，為獲取森林垂直結構特征提供了可能。吳艷雙等[10]采用機載高光譜和LiDAR數據構建多特征集合，對廣西高峰林場界牌分場樹種進行分類的精度最高達到83.70%；Zhao et al.[11]結合機載高光譜和LiDAR數據對東北某天然混交林樹種進行分類，平均分類精度達到85.33%。由此可見，結合LiDAR數據提取的垂直信息和高光譜影像數據的光譜信息能夠實現優勢互補，在提高森林樹種識別精度方面具有巨大潛力[12]。……