Sentinel-2光譜與紋理特徵對土地覆蓋分類效能的改進

Enhancing land cover classification using spectral and texture features from Sentinel-2 data

即時並有效地掌握土地覆蓋資訊，對於地景規劃或自然資源管理至關重要。遙測影像分類技術作為當代獲取該資訊的主要手段，如何提升其分類效能，仍是遙測研究持續關注的課題之一。本研究以大肚臺地為例，首先，透過Sentinel-2（S2）原始波段的組合，生成可突顯地物光譜特徵的光譜指標，同時基於灰階共生矩陣，擷取可反映空間排列特徵的紋理指標，再分別令其與原始波段整合，作為隨機森林分類演算法的輸入資料集，對7種土地覆蓋類型進行分類，最後，評估並比較不同特徵組合的分類效果。研究結果顯示，儘管S2原始波段在地覆分類中具相當潛力，但用於辨別高異質性混淆植被或高變異性草本植被時，其效果仍顯不足。無論添加光譜或紋理特徵，均有助整體分類效能的提升，且兩類特徵對特定類別準確度的提升具互補性，合併添加可獲得最佳分類效果，總體準確度與Kappa係數，分別可達90.56%、0.88，相較於僅使用原始波段，約可提升3.35%、0.04，並在McNemar檢定下，達差異顯著水準（p≤0.001），各類別F1分數亦能獲不同程度的改善（0.64–11.06%）。綜合上述結果，本研究建議在選擇土地覆蓋分類輸入資料集時，除常見的光譜特徵外，亦可考慮額外添加紋理特徵，以最大程度提高分類準確度。

Timely and effective acquisition of land cover information is crucial for landscape planning and natural resource management. Remote sensing image classification is a primary method for acquiring land cover information, and enhancing the performance of classification still remains a key research topic to focus on. In our study, spectral features were generated from the combinations of Sentinel-2 (S2) original bands in Dadu Terrace area to highlight the spectral characteristics of land cover types. Simultaneously, texture features reflecting spatial arrangement patterns were extracted using the gray-level co-occurrence matrix. These features were then combined with the original bands to form input datasets for the random forest classification algorithm, which was used to classify seven land cover types. Finally, the classification performance of different feature combinations was evaluated and compared. The results indicate that although the S2 original bands display considerable potential in land cover classification, they are still insufficient for distinguishing heterogeneous mixed vegetation or highly variable herbaceous vegetation. Incorporating either commonly used spectral features or texture features improved overall classification performance and also enhanced classification accuracy from their complementary effects on the specific categories. Using the combination of spectral and texture features achieved the best results with an overall accuracy of 90.56% and a Kappa coefficient of 0.88, representing improvements of 3.35% and 0.04, respectively, over the use of original bands alone, and these improvements were statistically significant (p≤0.001), confirmed by the McNemar test. Additionally, the F1-scores across all categories also showed enhancement, ranging from 0.64 to 11.06%. In summary, this study recommends to use combination of texture features and spectral features to maximize classification accuracy when selecting input datasets for land cover classification.