在低軌衛星遙測平台快速發展的推動下，遙測影像已成為地球觀測任 務中不可或缺的資源。然而，要將深度學習模型真正部署到這類衛星平台 上，困難之處並不只在於辨識精度：衛星端本身的運算能力有限、對地面 下傳的頻寬受限，而且當衛星通過不同地區上空時，所觀測到的地貌也會 持續改變；既有研究多半只單獨處理其中一項限制，少有同時兼顧者。為 此，本研究提出資源高效率的語義衛星通訊框架 REL-Net，在分割學習 （Split Learning）架構下同時處理上述三項限制。分割學習將模型拆成兩 半，衛星端只執行前半段的編碼器，其餘的解碼器與分類器交由地面站完 成，因此衛星端的運算負擔大幅降低，原始影像也不必離開衛星。針對傳 輸量，本研究提出深度壓縮特徵傳輸（Deep Compressed Feature Transfer， DCFT）模組：先為每一個特徵通道評估其區分不同地物類別的能力，只 挑選貢獻最大的少數通道進行傳送，再以 Lloyd–Max 量化器將這些數值 以較少的位元表示，使每張影像所需傳送的位元組數大幅下降；同時以深 度 Q 網路（Deep Q-Network， DQN）依據當下可用的頻寬，決定編碼器各 層應採用哪一種特徵擷取方式，其獎勵設計同時考慮新地區的分割精度、 傳輸量，以及舊地區知識的損失程度。針對跨地區的持續適應，本研究 提出兩階段雙頭式持續增量地物分類（Continual Incremental Land-Cover Classification， CICC）流程：第一階段先在來源資料集上完成訓練，第二 階段再為新地區加上一個新的分類頭，同時凍結舊地區的分類頭，並以第 一階段的模型作為參照，使模型在學會新地區的同時不至於遺忘舊地區。 於 DeepGlobe 與 LoveDA 兩組資料集上的實驗顯示，相較於多種地物分類 模型，本研究所提架構位於「精度—運算量」權衡的柏拉圖前緣（Pareto front），亦即沒有任何比較對象能在兩項指標上同時更好，能在明顯降低 運算成本與傳輸負擔的前提下維持具競爭力的分割效能。實驗另外得到兩 項觀察：其一，第一階段並非訓練得越久越好，訓練程度適中的模型保有 較大的調整空間，其在第二階段的遺忘程度反而遠低於過度訓練的模型； 其二，兩個遷移方向所產生的遺忘程度差異極大，顯示評估此類系統時必 須同時觀察分割精度、遺忘量與傳輸量三項指標，僅以單一精度指標判斷 容易造成誤判。;With the rapid advancement of low-Earth-orbit satellite sensing plat- forms， remote sensing imagery has become an indispensable resource for Earth observation. Deploying deep models on such satellites， however， is difficult for reasons that go beyond accuracy: the onboard processor has lim- ited computing power， the downlink to the ground has limited bandwidth， and the land cover the satellite observes keeps changing as it passes over new regions. Most existing works address only one of these issues at a time. This thesis proposes REL-Net， a resource-efficient framework for semantic satellite communication that addresses all three within a split learning ar- chitecture. Split learning divides the network in two: the satellite runs only the shallow encoder， while the ground station runs the decoder and the clas- sifier， so that the onboard computing cost stays small and the raw image never leaves the satellite. To reduce what must be transmitted， we design a Deep Compressed Feature Transfer (DCFT) module that scores every fea- ture channel by how well it separates the land-cover classes， sends only the highest-scoring channels， and represents their values with few bits through a Lloyd–Max quantizer， which sharply lowers the bytes sent per image. A Deep Q-Network (DQN) then decides， according to the available bandwidth， which feature-extraction branch each encoder stage should use; its reward accounts jointly for the accuracy on the new region， the transmitted volume， and the knowledge lost on the old region. For adaptation across regions， we propose a two-stage dual-head Continual Incremental Land-Cover Classifica- tion (CICC) procedure: the model is first trained on the source dataset， and a second classification head is then added for the new region while the old head is frozen and the Stage 1 model is kept as a reference， so that the new region is learned without erasing the old one. Experiments on DeepGlobe and LoveDA show that， against a broad set of land-cover classification models， the pro- posed system lies on the Pareto front of the accuracy–computation trade-off， meaning that no compared model is better on both axes at once， and that it markedly reduces computation and transmission while keeping competitive accuracy. Two further observations are reported. Longer Stage 1 training is not always better， since a moderately trained checkpoint keeps more room to adjust and forgets far less in Stage 2 than an over-trained one; and forgetting differs greatly between the two transfer directions， which shows that accu- racy， forgetting， and bandwidth must be read together rather than accuracy alone.
