| DC Field | Value | Language |
| dc.contributor.author | ElBATOUL, ZErga | - |
| dc.date.accessioned | 2026-10-06T14:43:29Z | - |
| dc.date.available | 2026-10-06T14:43:29Z | - |
| dc.date.issued | 2 | - |
| dc.identifier.uri | https://repository.esi-sba.dz/jspui/handle/123456789/979 | - |
| dc.description | Supervisor: Dr.Meziane Iftene /
Co-supervisor : Dr.Larabi Mohammed El Amin /
Co-supervisor : Dr.Chaib Souleymene | en_US |
| dc.description.abstract | Autonomous Earth Observation (EO) satellites must perform inference and adaptation
inside a strict operational envelope: RAM ≤ 1 MB for STM32H7-class deployment,
an OPS-SAT Operational SWaP Limit of ≤ 1.2 W, and a mission-critical throughput
target of ≤ 10 ms per image patch. Conventional deep networks exceed this envelope
and cannot support multi-year missions that encounter new terrain classes without
severe forgetting.
This thesis proposes Orbital Plasticity, a sensor-agnostic Few-Shot Continual
Learning framework for on-board satellite intelligence. The validated Level 2 system
combines PEFT-LoRA adaptation with five architecture-agnostic amendments: tiered
rank scaling, multi-pass checkpointing, uncertainty-aware dataset curation, adaptive
spiking prototype timing, and semantic drift guarding. NanoProtoNet remains the
Level 1 physical implementation target for STM32H7 deployment.
On EuroSAT RGB, the primary run demonstrates measured plasticity (ΔM′ ∈
{0.3000, 0.4000, 0.5000}), no measured forgetting (BWT = 0.0000), and a 0.60 MB
runtime footprint with 100 % low-power SPN residence at T = 50. A separate RGB
spectral-twin stress test exposes the Highway/River limitation (6.0 % diagnostic accuracy),
while the 13-band Sentinel-2 multispectral extension reaches 70.7 % final 5-way
accuracy, showing that NIR/SWIR evidence can harden ambiguous decision bounda***
Les satellites d’observation de la Terre autonomes font face à une triade de contraintes
opérationnelles: empreinte mémoire (≤1 Mo pour STM32H7), budget énergétique
(≤1.2 W OPS-SAT, limite SWaP opérationnelle), et débit de traitement (≤10 ms par
image). Cette thèse propose Orbital Plasticity, un cadre d’apprentissage continu en
peu d’exemples pour l’intelligence embarquée des satellites. Validé sur EuroSAT RGB,
le meilleur essai atteint une plasticité mesurée (ΔM′ ∈ {0.3000, 0.4000, 0.5000}), un
transfert rétrograde nul (BWT = 0.0000), et une empreinte d’exécution de 0.60 Mo.
Un essai de stress séparé sur les jumeaux spectraux est conservé comme cas d’échec
(ΔM′
moyen = −0.0200), motivant une escalade A1 plus agressive. Une configuration
multispectrale complète ensuite l’entrée RGB par 13 bandes Sentinel-2 et atteint 70.7 %
en 5-way final sur les classes spectralement confusables. | en_US |
| dc.language.iso | en | en_US |
| dc.subject | Few-Shot Continual Learning | en_US |
| dc.subject | Earth Observation | en_US |
| dc.subject | LoRA, | en_US |
| dc.subject | Spiking Neural Networks | en_US |
| dc.subject | Semantic Drift Guard | en_US |
| dc.subject | Edge Deployment | en_US |
| dc.subject | OPS-SAT | en_US |
| dc.subject | STM32H7 | en_US |
| dc.subject | Orbital Plasticity | en_US |
| dc.title | Orbital Plasticity Resource-Efficient Few-Shot Continual Learning for Autonomous Earth Observation Satellites | en_US |
| dc.type | Thesis | en_US |
| Appears in Collections: | Ingenieur
|