FedSpace: An Efficient Federated Learning Framework at Satellites and Ground StationsFedSpace：衛星與地面站之間的高效聯邦學習框架

1. Where this paper sits1. 這篇論文的位置

FedSpace is the bridge between the small analytical letter of Razmi et al. ^[90] and the ISL-based OrbitalBrain ^[94] . It shares three authors, Hsieh, Arzani and Chandra, with the latter. Thus, the two papers are best read as the same group's view of one problem before and after inter-satellite links. FedSat shows predictable contacts replacing synchronous rounds on ten satellites.

FedSpace studies aggregation schedules for a real fleet with irregular contacts. Nearly two hundred satellites visit a dozen fixed ground stations at varying times. Section 9.3 uses the sixteen-fold scheduling improvement over synchronous training as evidence for visibility-limited learning. Every link stays fixed, isolating the scheduling effect.

Because it is a preprint with a single version and a venue still awaiting confirmation, every number in this digest carries the preprint tag.

FedSpace 是 Razmi 等人的短篇分析論文^[90]與以 ISL 為基礎的 OrbitalBrain^[94]之間的橋樑，而且它與後者共享三位作者（Hsieh、Arzani 與 Chandra），所以兩篇最好當成同一個團隊在有星間鏈路之前與之後對同一問題的看法來讀。FedSat 在十顆衛星上證明可預測的接觸能取代同步回合，FedSpace 則問：擁有真實機隊的營運者，在接觸模式隨時間變化時（十幾座固定地點的地面站、將近兩百顆在各異時間經過的衛星）該如何排程彙總。第 9.3 節把它當作「衛星聯邦學習受可見性主導，頻寬則維持固定」最有力的證據，因為它相對同步訓練的十六倍改善完全來自排程，所有鏈路維持原設定。由於它是只有一個版本、發表場所仍待確認的預印本，本導讀中的每個數字都帶有預印本標記。

2. The problem, measured2. 問題，用量測說話

The system model is Planet's actual infrastructure. The 12 ground stations act together as a single federated-learning server. During contact, a satellite downloads the global model. It trains on local imagery between contacts and uploads at the next contact.

Contact requires satellite elevation above a station-specific minimum. Slots have T₀ = 15 minutes. ^{[91, §2.1]} This turns connectivity into a deterministic, time-varying matrix computable in advance from orbits and Earth rotation ^{[91, §2.2]} . Two numbers from that matrix define the difficulty. Synchronous learning waits for every satellite each round and takes 45.8 days to reach the target.

Fully asynchronous learning aggregates each update immediately. It stays below the target because updates from older models reverse some global-model progress. ^{[91, §1]}

系統模型就是 Planet 的真實基礎設施。12 座地面站合起來當作單一聯邦學習伺服器，衛星在接觸時下載全局模型，離開接觸範圍時用自己的影像訓練，到下一次接觸時上傳更新^{[91, §2.1]}。當衛星對某座地面站的仰角超過最小角度時就算連上，論文把時間切成 T₀ = 15 分鐘的時槽，於是連線性變成一個確定性、隨時間變化、可以事先從軌道與地球自轉算出的矩陣^{[91, §2.2]}。這個矩陣中的兩個數字界定了難度。

同步聯邦學習每一回合都要等所有衛星，在這個時刻表上需要 45.8 天才達到目標準確率。完全非同步的學習在每個更新到達時就彙總，卻始終低於目標，因為在舊模型上算出的更新會把全局模型往回拉^{[91, §1]}。

3. The design3. 設計

FedSpace's design is a choice of when to aggregate, made in advance. Frequent aggregation gives downloading satellites a fresh global model. Incoming updates then reflect models that may have undergone several replacements. The number of global rounds that have passed between a satellite's download and the aggregation of its update is its staleness s. Rare aggregation limits staleness. Returning satellites can still receive only the previous global model, wasting contact time through idleness. ^{[91, §2.3]}

The two costs pull in opposite directions. Their balance depends on exactly which satellites will be in contact in the coming slots. This is the information an orbit predictor supplies. Predicted connectivity determines the aggregation schedule that maximizes estimated convergence rate.

Update weight c(s) = (s+1)^−α retains stale contributions while reducing their influence. ^{[91, §2.3–2.4]} Orbital predictions adapt the schedule to uneven fleet connectivity. The server aggregates as fresh updates approach and waits when few satellites are visible.

FedSpace 的設計是事先決定何時彙總。頻繁彙總讓下載全局模型的衛星拿到新鮮的模型，但每個到達的更新所依據的模型在此期間已被替換好幾次，而衛星下載模型到其更新被彙總之間經過的全局回合數，就是它的陳舊度 s。很少彙總能讓陳舊度保持低，但已經回報過的衛星再次進入接觸時僅能取得上一版全局模型，這段浪費掉的接觸時間就是閒置^{[91, §2.3]}。這兩種成本方向相反，而它們的平衡點取決於接下來幾個時槽中到底哪些衛星會進入接觸，這正是軌道預測器能提供的資訊。因此 FedSpace 從預測的連線性計算出一個彙總時刻表，使收斂速率的估計值最大化，並以 c(s) = (s+1)^−α 為每個更新加權，讓陳舊的更新仍有貢獻、只是貢獻較少^{[91, §2.3–2.4]}。由於時刻表是從軌道推導出來的，它會自動適應真實機隊隨時間變化的接觸分布：在很多新鮮更新即將到達時彙總，在視野內衛星很少時暫緩。

4. Evaluation setup4. 實驗設定

The evaluation is simulation driven by a real constellation's geometry. Connectivity comes from the orbits of Planet's 191 satellites and the locations of its 12 ground stations, discretized into 15-minute slots. Training uses real satellite images from Christie et al. (2018), authors of the fMoW dataset. Images use IID and non-IID partitions.

Baselines are synchronous FL, asynchronous FL and FedBuff. FedBuff aggregates after a fixed number of updates. ^{[91, §2.2]} All methods are compared by the time to reach the same target accuracy ^{[91, §1]} .

The model uses ground links only. It states that "inter-satellite communication is currently not possible in LEO satellites". Its citation is Handley's HotNets 2019 study of ground relays for Starlink's laser-free first phase. ^{[91, §2.1]} Each slot marks a satellite contact as active or inactive. The fact base leaves link rate, pass duration and model architecture unspecified. Thus, the evaluation can be read only as a statement about scheduling under a contact pattern, with pass capacity in bytes outside its scope.

評估是由真實星座幾何驅動的模擬。連線性來自 Planet 191 顆衛星的軌道與其 12 座地面站的位置，離散成 15 分鐘時槽。訓練資料是 Christie 等人（2018，fMoW 資料集的作者）研究中的真實衛星影像，以 IID 與非 IID 兩種方式分配給各衛星^{[91, §2.2]}。基準是同步 FL、非同步 FL 與 FedBuff（一種在收到固定數量更新後才彙總的緩衝式非同步方法），所有方法以達到相同目標準確率所需的時間比較^{[91, §1]}。網路方面，模型僅採地面鏈路，理由是「LEO 衛星目前不可能進行衛星間通訊」這句話，論文為此引用的是 Handley 在 HotNets 2019 研究 Starlink 採對地鏈路的第一階段之地面中繼的論文^{[91, §2.1]}。連線性模型以「衛星接觸為啟用或停用」的時槽表示，而鏈路資料速率、過頂時長及模型架構仍待事實庫補充，所以這個評估只能讀成在某種接觸模式下的排程結論，一次過頂的位元組容量則位於評估範圍之外。

5. What was measured5. 量到了什麼

6. Limits and open questions6. 限制與待解的問題

• A preprint with one version. The sole release is arXiv v1, dated 2 February 2022. Venue and refereed status remain unconfirmed, so the preprint tag preserves its evidence status.只有一個版本的預印本。2022 年 2 月 2 日的 arXiv v1 是唯一版本，發表場所仍待確認，這些數字的審查狀態也仍待確認，引用時應加上預印本標記。
• "Not possible in LEO" needs a date. The ground-only premise cites a 2019 study of Starlink's laser-free first phase. Starlink has carried lasers since v1.5. Four years later, the same group's OrbitalBrain assumes about 100 Mbps ISLs. The claim describes Planet-class Earth-observation satellites of 2022, with its scope limited to that class of LEO spacecraft (fact base Part C, item 59).「LEO 不可能」需要標上日期。僅採地面鏈路的前提建立在 2019 年對 Starlink 採對地鏈路的第一階段的分析上，而 Starlink 自 v1.5 起就搭載雷射鏈路，同一團隊四年後的 OrbitalBrain 也假設約 100 Mbps 的 ISL。這個說法描述的是 2022 年 Planet 等級的地球觀測衛星，其適用範圍限於這類 LEO 衛星（事實庫 Part C 第 59 項）。
• A contact model with byte capacity outside its scope. Connectivity is binary per slot. The fact base leaves link rate, pass duration and model size open, so assessing model-size limits and antenna contention requires those additional inputs.接觸模型的範圍排除位元組容量。每個時槽的連線性是二元的，而鏈路速率、過頂時長及模型大小仍待事實庫補充，所以評估模型大小的傳輸上限及地面站天線競爭對排程的影響，仍需這些額外輸入。
• A small rounding mismatch. 45.8 − 43.1 days leaves 2.7 days. This gives about 17× rather than 16.5×, so one of the printed values is rounded. The digest quotes the paper's figures and the computed check side by side.一個小小的四捨五入落差。45.8 − 43.1 天剩下 2.7 天，得到約 17 倍，對照論文所述的 16.5 倍，所以某個印出的數值經過四捨五入。本導讀把論文的數字與計算檢核並列。
• Cross-paper numbers require their evaluation conditions. AsyncFLEO reimplements FedSpace on 40 satellites at 2,000 km with 16 Mb/s links. Its 72 h result applies to that separate setting.跨論文數字須附上各自的評估條件。AsyncFLEO 給 FedSpace 的 72 h 來自在 2,000 km、40 顆衛星、16 Mb/s 鏈路上的重新實作，此結果適用於重新實作的情境，原論文的量測條件須另行保留。

7. How it connects7. 與其他內容的關聯

• Section 9.3 compares 45.8 days with about 2.7 days under fixed links. This isolates visibility and scheduling effects. Panel 9.2 uses those times as reference ticks.第 9.3 節：FedSpace 的 45.8 天在鏈路固定的情況下縮到約 2.7 天，是該章「決定訓練時間的是可見性，頻寬則維持固定」的核心證據，面板 9.2 也以這兩個值作為參考刻度。
• Section 9.5: FedSpace sits at the bottom of the link-rate ladder with FedSat, using ground links only.第 9.5 節：FedSpace 與 FedSat 一起位於鏈路速率之梯的最底層，假設僅採地面鏈路。
• Section 9.1, Section 9.2 and Section 9.4: the CDN and inference services that need links outside FedSpace's ground-only model.第 9.1 節、第 9.2 節與第 9.4 節：需要 FedSpace 模型以外之鏈路的 CDN 與推論服務。
• Section 3.3: gateways and ground stations, here the only aggregation points, and the source of the contact pattern.第 3.3 節：閘道與地面站，在這裡是唯一的彙總點，也是接觸模式的來源。
• Section 10.6: the constellation-wide regime where federated learning fits, as opposed to the formation cluster.第 10.6 節：聯邦學習所屬的星座尺度區間，相對於編隊叢集。
• FedSat [90] is the analytical precursor on ten satellites. OrbitalBrain [94] is the same group's successor with ISLs. Handley [7] is the ground-relay paper FedSpace cites for its ground-only premise. Orbital Edge Computing [60] describes the same ground-station bottleneck for inference.FedSat[90]是在十顆衛星上的分析前作。OrbitalBrain[94]是同一團隊加入 ISL 的後繼者。Handley[7]是 FedSpace 為其ISL 排除 前提所引用的地面中繼論文。軌道邊緣運算[60]從推論角度描述了同樣的地面站瓶頸。

Key numbers關鍵數字