傳統雨量推估高度依賴雨量計與氣象雷達等地面觀測設備，但此 類儀器的部署成本高且空間覆蓋侷限，並且容易受到地形限制。相 較之下，氣象衛星提供了不受地形及海陸分布之高範圍與高時空解 析度等優勢，使其成為雨量推估的重要補強來源。本研究旨在利用 日本同步衛星向日葵八號之多頻道時序影像，結合 Spatial Spectral Transformer（SSFormer）、卷積長短期記憶網路（ConvLSTM）與 LinkNet 解碼器，建構一套適用於台灣地區的降雨等級推估模型，並 評估其作為使用衛星影像來補強地面降雨推估的可行性。 研究主要分成三個部分。首先，將連續三時刻（T−2、T−1、T） 的多頻道衛星影像進行標準化與區塊化處理，以建構時序輸入資料。 第二步我們採用 SSFormer 與 ConvLSTM 進行光譜與時間特徵建模， 並於解碼端採用 LinkNet 的多尺度上採樣策略。此外，本研究額外加 入邊界強化與有序回歸相關機制，並搭配一主幹三分支的損失設計， 使模型在處理降雨連續性及等級邊界時能更精確。模型最終以 Focal Loss 與多任務損失共同訓練，並以 Macro-F1、mIoU 與邊界 F1 等指 標進行模型評估。 實驗結果顯示，本研究提出之 SSFormer-ConvLSTM-LinkNet 模 型在降雨等級分類與邊界偵測上皆優於既有方法，尤其在邊界區域的 錯判率有明顯改善。在模擬衛星觀測中斷的情境下，模型仍能維持穩 定的推估能力，顯示其具備作為衛星資料備援方案的潛力。本研究證 實了僅依賴向日葵八號影像即可達成相當可靠的降雨等級預測。未來 可進一步優化模型以降低計算成本，並探索其在不同區域與氣候條件 下的適用性，尤其是時間性的資料集不平衡問題。;Traditional rainfall estimation relies heavily on ground-based ob- servation systems such as rain gauges and weather radars. However， these instruments are costly to deploy， provide limited spatial cover- age， and are often constrained by complex terrain. In contrast， meteo- rological satellites offer wide-area coverage， and high spatiotemporal resolution， making them an important complementary source for rain- fall estimation. This study aims to develop a rainfall intensity classification model for Taiwan using multi-channel temporal imagery acquired by the Himawari- 8 satellite. The proposed framework integrates a Spatial Spectral Trans- former (SSFormer)， a Convolutional Long Short-Term Memory (Con- vLSTM) network， and a LinkNet decoder. Furthermore， the feasibility of using satellite imagery to complement weather radar observations for rainfall monitoring and classification is investigated. The research framework consists of three main components. First， multi-channel satellite images from three consecutive timestamps (T−2， T−1， T) are standardized and divided into spatial patches to construct the temporal input sequence. Second， spectral and temporal fea- tures are modeled using SSFormer and ConvLSTM， while the decoder adopts LinkNet＇s multi-scale upsampling strategy. In addition， this study incorporates boundary enhancement and ordinal regression mech- anisms， along with a three-branch loss design， enabling the model to better capture rainfall continuity and class boundaries. The final model is trained using a combination of Focal Loss and multi-task losses， and its performance is evaluated through metrics such as Macro-F1， mIoU， ii and boundary F1. Experimental results show that the proposed SSFormer-ConvLSTM- LinkNet model outperforms existing approaches in both rainfall clas- sification and boundary detection， with significantly reduced misclas- sification along rainfall transition zones. Under simulated satellite- outage scenarios， the model maintains stable estimation performance， demonstrating its potential as a fallback solution when satellite data are unavailable. This study verifies that reliable rainfall intensity pre- diction can be achieved using Himawari-8 imagery alone. Future work may further optimize computational efficiency and investigate model generalization across different regions and climatic conditions， partic- ularly addressing temporal class imbalance issues.
