大規模景觀圖像斑塊特征增強算法仿真

摘" 要： 不同景觀斑塊特征存在一定的差異，整體增強會導致斑塊重疊和模糊等問題。為此，提出一種大規模景觀圖像斑塊特征增強算法。計算大規模景觀圖像斑塊形狀指數、多樣性指數、破碎性指數、最大斑塊指數以及優勢度指數，以此反映景觀圖像內斑塊組成和結構特征，并度量景觀斑塊特征；再將所有指數計算結果組成斑塊特征集，輸入多分支注意力機制卷積神經網絡中，依據網絡通道注意力機制增強圖像斑塊特征分辨率；最后，將增強結果作為局部特征融合網絡的輸入，通過該網絡的卷積操作生成各個通道的局部斑塊圖，獲取局部特征、斑塊特征的位置和細節信息，完成斑塊特征二次增強。仿真實驗結果表明：所提出的增強算法的梯度損失和結構相似性損失函數值均在0.10以下，說明其能夠有效處理斑塊邊緣之間的模糊效應，并且可靠區分不同的景觀斑塊分布空間。

關鍵詞： 大規模景觀圖像； 斑塊特征； 增強算法； 網絡通道注意力機制； 卷積神經網絡； 特征分辨率

中圖分類號： TN911.73?34； TP391" " " " " " " " " "文獻標識碼： A" " " " " " " " " 文章編號： 1004?373X（2024）12?0086?05

Simulation of large scale landscape image patch feature enhancement algorithm

YANG Bixiang

（Beijing Institute of Technology， Zhuhai 519088， China）

Abstract： There are a certain differences in the characteristics of different landscape patches， and overall enhancement can lead to issues such as patch overlap and blurring. Therefore， a large?scale landscape image patch feature enhancement algorithm is proposed. The shape index， diversity index， fragmentation index， maximum patch index， and dominance index of large?scale landscape image patches are calculated to reflect the composition and structural characteristics of patches in the landscape image， and measure the characteristics of landscape patches. All index calculation results are composed into a patch feature set and input into the multi branch attention mechanism convolutional neural network， so as to enhance the resolution of image patch features by means of the network channel attention mechanism. The enhancement results are used as input of the local feature fusion network， and local patch maps of each channel are generated by means of the convolution operation of the network， obtaining local features， the position and detail information of patch features， and completing the secondary enhancement of patch features. The simulation experimental results show that the gradient loss and structural similarity loss functions of the enhanced algorithm are both below 0.10， indicating that it can effectively handle the fuzzy effects between patch edges and reliably distinguish different landscape patch distribution spaces.

Keywords： large scale landscape images; patch characteristics; enhanced algorithms; network channel attention mechanism; convolutional neural networks; feature resolution

0" 引" 言

隨著遙感技術和圖像處理技術的發展，獲取大規模景觀圖像數據變得更加容易，但由于景觀多樣性和復雜性，利用傳統方法進行特征提取和分析仍然具有挑戰性[1?2]。針對這一問題，研究大規模景觀圖像斑塊特征增強算法具有重要意義，可以有效地提取和增強景觀圖像中的斑塊特征，為后續的分類、目標檢測、環境監測等應用提供更可靠的基礎[3]。……