在越南湄公河三角洲，監測稻田、區分雙季稻田和三季稻田以及稻蝦共作田，並追蹤水產養殖、濕地和紅樹林區域，是環境管理、農業實踐和其他領域面臨最關鍵和最具挑戰性的問題。本研究結合先進的地理空間資料和遙測技術，以及創新的機器學習和深度學習方法，以應對這些挑戰。我們評估了多種模型，包括無監督、監督和半監督方法，並使用了各種衛星衍生指數。除了傳統模型外，我們還採用了深度卷積嵌入式聚類（DCEC）和基於置信度感知聚類的半監督軟標籤選擇（SoC4SS-FGVC）等新方法。在6024個樣本的實驗中，卷積神經網路（CNN）模型表現出最佳精確度，達到0.97。在樣本量僅 500 的情況下，監督模型的限制顯露無疑，準確率顯著下降至約 0.75 – 0.85，而 SoC4SS-FGVC 模型則取得了顯著的高精度（500 個樣本點下準確率達 0.92）。非監督模型的準確率則低得多，介於 0.40 和 0.50 之間。產生的土地覆蓋圖揭示了分類方面的挑戰，包括由於過度擬合和欠擬合導致的相似土地利用類型之間的混淆。土地覆蓋的均一性和有限的地面實測資料限制了監督模型的表現。相反地，儘管計算量更大，但半監督技術仍展現出良好的應用前景。與官方統計數據、專家意見和先前研究的交叉驗證表明，SoC4SS-FGVC 模型顯著提高了稻田的檢測和分離以及紅樹林的監測效果。這些發現為具有特定土地利用模式區域的土地覆蓋分類提供了寶貴的見解。;In the Mekong River Delta of Vietnam， monitoring paddy rice， distinguishing between double- and triple-cropping paddy rice fields and rice-shrimp land， and tracking aquaculture， wetlands， and mangrove areas are the most critical and challenging issues due to the need for environmental management， agricultural practices， and other sectors. This study addresses these challenges by combining advanced geospatial data and remote sensing techniques with innovative machine learning and deep learning methods. We evaluated several models， including unsupervised， supervised， and semi-supervised approaches， using various satellite-derived indices. Beyond traditional models， we employed novel approaches such as Deep Convolutional Embedded Clustering (DCEC) and the semi-supervised Soft Label Selection with Confidence-Aware Clustering (SoC4SS-FGVC). The CNN model showed the best accuracy at 0.97 when experimenting with 6024 samples. At small sample sizes of 500， the supervised models exhibited their constraints as the accuracies dropped significantly to about 0.75 – 0.85， while the SoC4SS-FGVC model achieved remarkably high-performance accuracy (0.92 with 500 sample points). Unsupervised models yielded much less accurate results， with scores between 0.40 and 0.50. The generated land cover maps revealed classification challenges， including confusion among similar land use types caused by overfitting and underfitting. The uniformity of land cover and limited ground-truth data restrict the performance of supervised models. Conversely， semi-supervised techniques show promise despite higher computational demands. Cross-validation with official statistics， expert input， and prior research demonstrates that the SoC4SS-FGVC model significantly improves paddy rice field detection and separation and mangrove monitoring. These findings offer valuable insights into land cover classification in areas with specific land-use patterns.
