Deep learning has become a key enabler of remote-sensing–based environmental monitoring because it can rapidly extract information from large volumes of satellite imagery. However, the appearance of targets and their surrounding background is often ...
Deep learning has become a key enabler of remote-sensing–based environmental monitoring because it can rapidly extract information from large volumes of satellite imagery. However, the appearance of targets and their surrounding background is often non-stationary: it changes with acquisition geometry, environmental forcing, seasonality, and site-specific conditions. As a result, robust deployment typically requires large and diverse labeled datasets. In practice, such labels are still most often produced by expert visual interpretation, which is labor-intensive and costly. This limitation becomes more critical in two operational settings. First, for many satellite missions—especially rapidly expanding small-satellite constellations with comparatively short design lifetimes—waiting to accumulate “sufficient” imagery and labels before building a model can reduce the usable lifetime of the algorithm, and may confine its applicability to retrospective analysis of past data. Second, for event-driven hazards such as oil spills (and similar sudden disasters), timely decision support requires a deployable model to be available immediately when an incident occurs; if model development is delayed until large training archives are collected, rapid response and information delivery may not be feasible. These constraints motivate an automated approach that can (i) bootstrap a usable model from limited initial supervision and (ii) keep the model current as new images arrive, without repeated full manual labeling cycles.
To address this need, this thesis proposes a self-evolving deep learning framework composed of three modules: (1) satellite-image analysis, (2) automatic training-data generation, and (3) model upgrade via iterative retraining. The framework is validated on three representative environmental-monitoring tasks—marine oil-spill detection, tidal-flat waterline extraction, and glacier-terminus delineation—selected because they share two practical characteristics that make static, one-shot model development fragile. First, all three tasks operate in highly variable, heterogeneous, and rapidly changing environments (e.g., tide- and turbidity-driven shoreline variability, speckle- and mélange-driven SAR variability at glacier fronts, and sea-state/geometry-driven variability for oil slicks and look-alikes). Second, precisely because of this variability, automatic training-data generation is intrinsically difficult: simple global rules often fail to produce consistently reliable labels across time and space, which is why many operational pipelines still require repeated expert intervention. These shared challenges provide a stringent testbed for evaluating whether a practical “update-as-you-observe” learning workflow can be realized for remote-sensing monitoring.
A central methodological contribution is Module-2, which implements guided post-classification to transform model outputs into new labels with sufficient quality for retraining. Rather than accepting raw predictions as pseudo-labels, Module-2 applies spatially constrained updating (e.g., buffers around expected boundaries or candidate regions), integrates task-specific auxiliary evidence (e.g., spectral-index guidance for optical waterlines), and uses conservative decision rules and morphological refinement to limit error propagation. This design explicitly targets the main failure mode in these monitoring problems—boundary ambiguity and scene-dependent nuisance variability—while reducing the need for sustained manual labeling as the data distribution broadens. To quantify model evolution consistently, the study adopts task-appropriate metrics, and for boundary-sensitive tasks further uses a composite score defined as the arithmetic mean of F1-score, Intersection over Union (IoU), Figure of Merit (FOM), and Mean Distance Error (MDE).
Across all three tasks, iterative self-evolving yields consistent net performance gains with modest cycle-to-cycle fluctuations. For tidal-flat waterline extraction, 16 self-evolving iterations improved the composite score from 0.82 to 0.89 (≈ +7%). For glacier-terminus delineation, 20 iterations improved the composite score from about 0.82 to about 0.87 (≈ +5%). For oil-spill detection, 23 iterations improved the full-scene F1-score from 0.84 to 0.90 (≈ +6%).
Importantly, the framework reduces reliance on fully manual labeling while maintaining strong performance: models trained with a smaller seed set plus automatically generated labels achieve accuracy comparable to—and in our experiments higher than—static baselines trained once using fully manual datasets of the same total size. Finally, the self-evolved models are applied to time-series monitoring case studies, including tidal-flat mapping and area-change estimation, glacier-terminus time-series construction and external comparison, and oil-slick distribution mapping from large SAR archives. Overall, these results demonstrate that an automated label-growth-and-retraining cycle is a practical and scalable way to keep remote-sensing monitoring models accurate and up to date under non-stationary conditions, while supporting reliable and temporally consistent environmental monitoring products for both long-term change analysis and event-driven response.