DGL-RSIS: Decoupled RS Segmentation

• DGL-RSIS is a training-free framework that decouples local class semantics and global context for robust remote sensing segmentation.
• It incorporates context-aware cropping and RS-specific prompt engineering to improve mask proposals and boost mIoU by up to 5 points.
• Empirical results demonstrate superior performance, with mIoU of 21.55% on iSAID and 21.50% on RRSIS-D compared to other methods.

Searching arXiv for DGL-RSIS and closely related remote sensing VLM segmentation work. DGL-RSIS is a training-free framework for remote sensing (RS) image segmentation that decouples visual and textual inputs and performs visual-language alignment at both the local semantic and global contextual levels. It was introduced to address two persistent transfer difficulties for vision-LLMs (VLMs) in RS imagery: the limited category diversity of RS datasets and the domain gap between natural and RS imagery. The framework combines a Global-Local Decoupling module, context-aware cropping, RS-specific knowledge enrichment for text inputs, Cross-Scale Grad-CAM, and mask selection to support both open-vocabulary semantic segmentation (OVSS) and referring expression segmentation (RES) without task-specific training or fine-tuning on RS data (Li et al., 30 Aug 2025).

1. Problem setting and conceptual scope

DGL-RSIS is situated in the problem of adapting pre-trained VLMs such as CLIP from natural-image understanding to RS segmentation. The central difficulty is that RS imagery differs substantially from natural imagery in viewpoint, scale variation, and texture, while RS datasets also exhibit fewer and often coarser class labels. DGL-RSIS addresses these constraints by separating local class semantics from global contextual modifiers and aligning them with different visual representations.

The framework is explicitly dual-task. For OVSS, it assigns semantic labels to class-agnostic mask proposals through mask-level image-text matching. For RES, it segments image regions specified by free-form referring expressions that may contain both class information and global spatial cues. The same decoupling principle governs both tasks: local semantics are associated with object- or region-level masks, whereas global context is handled through Grad-CAM-based localization and subsequent mask selection.

A recurrent ambiguity concerns the phrase “training-free.” In DGL-RSIS, this denotes the absence of additional supervised RS training, fine-tuning, or task-specific optimization. The framework instead uses pre-trained VLM knowledge together with unsupervised mask proposal methods, RS-specific prompting, and deterministic matching and selection procedures. This suggests a design centered on transfer rather than adaptation in the usual optimization-based sense (Li et al., 30 Aug 2025).

2. Global-local decoupling and task reformulation

The architectural core of DGL-RSIS is the Global-Local Decoupling (GLD) module. On the textual side, a referring expression is parsed with NLP techniques to separate class nouns from modifiers. The decoupling is written as

$T_{\text{ref}} = \text{NLP}(T),$

followed by

$$T_{\text{ref}} = T_{\text{cls}} \cup T_{\text{mod}}.$$

Here, $T_{\text{cls}}$ captures local semantic content such as object or class nouns, while $T_{\text{mod}}$ captures global modifiers such as spatial qualifiers.

On the visual side, the input RS image is decomposed into class-agnostic regions by an unsupervised mask proposal network:

$M_{\text{ca}} = \text{MPN}(I).$

This partitions the image into a global image context and a set of local mask proposals. The overall segmentation problem is thereby reformulated as a mask-level image-text matching problem, with the local masks serving as candidate semantic units and the full image remaining available for contextual reasoning.

The significance of GLD is methodological. Rather than forcing a single visual embedding to simultaneously encode category identity and spatial context, DGL-RSIS separates these factors before alignment. In the RS setting, where top-down layout and contextual arrangements often dominate recognition, this decoupling provides a structured alternative to direct end-to-end similarity matching (Li et al., 30 Aug 2025).

3. Local visual-textual alignment for open-vocabulary segmentation

The local alignment path supports OVSS by assigning each class-agnostic mask proposal a semantic label through VLM similarity. For a mask proposal $m_i$, the local visual feature is extracted by cropping and masking:

$$F_{\text{local}} = E_v(\mathcal{T}_{\text{crop}}(I \odot m_i)).$$

Similarity to candidate text features is computed as

$$P_{\text{local}} = \text{SoftMax}(F_{\text{local}} F_{\text{cls}}^T),$$

and the predicted class is obtained by

$y_{\text{local}} = \arg\max(P_{\text{local}}).$

The final OVSS output is then assembled as

$M_{\text{ov}} = \bigcup_{i=1}^M m_i \cdot y_i.$

A central difficulty in RS patch extraction is that naïve cropping often either removes discriminative context or retains excessive irrelevant background. DGL-RSIS therefore introduces a context-aware cropping strategy based on a minimum bounding rectangle (MBR) and buffer expansion:

$$T_{\text{ref}} = T_{\text{cls}} \cup T_{\text{mod}}.$$0

The stated objective is to preserve proper boundaries and contextual support for mask-guided visual features. This is particularly relevant for RS scenes, where object identity can depend strongly on surrounding land-cover configuration.

The textual pathway is also modified to compensate for domain mismatch. DGL-RSIS uses RS-specific prompt templates such as “Top view of a {CLASS}” and augments a class text bank with synonyms, visual interpretation descriptions, and negative or background land-cover classes. The text bank $$T_{\text{ref}} = T_{\text{cls}} \cup T_{\text{mod}}.$$1 is enriched by $$T_{\text{ref}} = T_{\text{cls}} \cup T_{\text{mod}}.$$2, $$T_{\text{ref}} = T_{\text{cls}} \cup T_{\text{mod}}.$$3, and $$T_{\text{ref}} = T_{\text{cls}} \cup T_{\text{mod}}.$$4, and for each class $$T_{\text{ref}} = T_{\text{cls}} \cup T_{\text{mod}}.$$5 the augmented text set is written as

$$T_{\text{ref}} = T_{\text{cls}} \cup T_{\text{mod}}.$$6

This local matching mechanism constitutes the OVSS component of the framework and provides open-vocabulary classification without RS-specific training (Li et al., 30 Aug 2025).

4. Global alignment for referring expression segmentation

The RES pathway is designed for expressions whose meaning depends not only on class identity but also on global modifiers. DGL-RSIS addresses this with Cross-Scale Grad-CAM, which separates attention generated by local class tokens from attention generated by global modifiers. A difference map between the two is computed as

$$T_{\text{ref}} = T_{\text{cls}} \cup T_{\text{mod}}.$$7

and the enhanced global activation is then formed by

$$T_{\text{ref}} = T_{\text{cls}} \cup T_{\text{mod}}.$$8

The final Grad-CAM is obtained by averaging this enhanced global activation with the activation of the full referring expression.

This construction is motivated by a limitation of vanilla Grad-CAM in multimodal alignment: averaging across all tokens can dilute the effect of modifiers such as directional or relational phrases. By isolating modifier-driven activations and contrasting them with class-token activations, DGL-RSIS attempts to make global context operational rather than auxiliary. The resulting maps are intended to be both more focused and more interpretable.

A subsequent mask proposal selection stage converts pixel-level activation maps into a mask-level segmentation result. First, peak Grad-CAM coordinates are extracted and used to identify activated masks:

$$T_{\text{ref}} = T_{\text{cls}} \cup T_{\text{mod}}.$$9

A weighted similarity score between each candidate mask and the global Grad-CAM is then computed and normalized by mask area; the top-scoring mask becomes the global candidate $T_{\text{cls}}$0. The final RES prediction is obtained by intersecting this global candidate with the locally assigned semantic masks:

$T_{\text{cls}}$1

This intersection enforces consistency between global spatial context and local class semantics, which is the central design principle of the framework (Li et al., 30 Aug 2025).

5. Empirical performance and ablation behavior

DGL-RSIS is evaluated on iSAID for OVSS and RRSIS-D for RES. On iSAID, it reports an mIoU of 21.55% and outperforms several training-free baselines, including SegEarth-OV at 18.4%, MaskCLIP at 16.91%, and GEM at 7.56%. On RRSIS-D, it reports an mIoU of 21.50%, exceeding IteRPrimE at 20.96%, CaR at 18.39%, and Global-Local CLIP at 15.01% (Li et al., 30 Aug 2025).

Setting	DGL-RSIS	Comparison reported
OVSS on iSAID	21.55% mIoU	SegEarth-OV 18.4%; MaskCLIP 16.91%; GEM 7.56%
RES on RRSIS-D	21.50% mIoU	IteRPrimE 20.96%; CaR 18.39%; Global-Local CLIP 15.01%

The ablation evidence is aligned with the architecture. Context-aware cropping using MBR plus buffer expansion improves mIoU by up to approximately 5 points relative to naïve cropping. Prompt engineering with “Top view of ...” and text-bank augmentation further improves accuracy, and the paper reports that each augmentation contributes additional gains. For global alignment, Cross-Scale Grad-CAM together with mask selection improves performance over single-scale Grad-CAM. The framework also reports that the choice of mask proposal model is dataset-dependent and should be considered under an mIoU-versus-speed tradeoff.

The reported empirical profile is notable for two reasons. First, DGL-RSIS is described as outperforming previous training-free methods on both OVSS and RES benchmarks. Second, it is also reported to outperform supervised ZegFormer and CLIPSeg on unseen categories. This suggests that the framework’s principal strength lies not in RS-specific optimization, but in retaining zero-shot generality while compensating for RS-specific semantic and contextual mismatches (Li et al., 30 Aug 2025).

6. Interpretation, research context, and disambiguation

Within RS segmentation research, DGL-RSIS belongs to a line of work that attempts to overcome data and supervision limitations by transferring general multimodal priors into geospatial imagery. A separate research direction addresses related difficulty through unsupervised domain adaptation, often under class-symmetry assumptions; one multi-source class-asymmetry method proposes expert branches, collaborative learning with cross-domain mixing, pseudo-label generation, and a multiview-enhanced knowledge integration module for RS semantic segmentation (Gao et al., 2023). DGL-RSIS differs in operating without RS-specific training and in using explicit visual-textual decoupling rather than optimization-based adaptation (Li et al., 30 Aug 2025).

Its methodological emphasis is interpretability through decomposition. The framework does not treat the referring expression as a single monolithic prompt, nor the image as a single undifferentiated embedding. Instead, it enforces separate pathways for local class semantics and global contextual modifiers, and joins them only at the point of mask-level decision. A plausible implication is that this design is especially compatible with scenarios in which annotation is limited, category vocabularies are fluid, and decisions need to remain traceable to specific masks and activation maps.

The acronym “DGL” is also used in unrelated research areas. In machine learning systems, it denotes the “Deep Graph Library” (Wang et al., 2019). In formal methods, it denotes “Differential Game Logic” (Platzer, 2014), with a subsequent uniform-substitution calculus developed in (Platzer, 2018). DGL-RSIS refers instead to a remote sensing image segmentation framework whose defining contribution is the decoupling of global spatial context and local class semantics for training-free OVSS and RES in RS imagery (Li et al., 30 Aug 2025).