紅樹林是提供各種生態和社會經濟效益的關鍵生態系統，但它們受到森林砍伐和城市化等人類活動的威脅。傳統的紅樹林損失監測方法依賴於勞動密集和耗時的現場調查或高解析度衛星圖像分析，通常在空間覆蓋和時間分辨率上存在限制。 LSST-Former架構整合了FCN、基於Transformer的模型和少樣本學習技術的優勢，以應對使用小數據集進行紅樹林損失檢測的挑戰。 Transformer已經在捕捉長程依賴性和從序列數據中學習方面取得了顯著成功，而少樣本學習使模型能夠很好地對未見過的類別或任務進行泛化，並且具有有限的訓練示例。通過結合這些方法，LSST-Former旨在有效地從異構類別中學習。我們的實驗結果展示了LSST-Former相對於現有的深度學習架構（如隨機森林、支援向量機、U-Net、LinkNet、Vision Transformer、SpectralFormer、MDPrePost-Net 和SST-Former）的優越性能，凸顯了其在紅樹林保護和管理工作中實際應用的潛力。;Mangroves are crucial ecosystems that provide various ecological and socio-economic benefits, but they are under threat from anthropogenic activities such as deforestation and urbanization. Traditional methods for monitoring mangrove loss rely on labor-intensive and time-consuming field surveys or high-resolution satellite imagery analysis, which are often limited in spatial coverage and temporal resolution. The LSST-Former architecture integrates the strengths of both FCN, Transformer-based models, and few-shot learning techniques to address the challenges of mangrove loss detection with small datasets. Transformers have demonstrated remarkable success in capturing long-range dependencies and learning from sequential data, while few-shot learning enables models to generalize well to unseen classes or tasks with limited training examples. By combining these approaches, LSST-Former aims to learn from heterogeneous classes effectively. Our experimental results showcase the superior performance of LSST-Former compared to existing deep learning architectures such as random forest, Support Vector Machine, U-Net, LinkNet, Vision Transformer, SpectralFormer, MDPrePost-Net, and SST-Former, highlighting its potential for practical applications in mangrove conservation and management efforts.
