Vision-language models (VLMs) have sparked growing interest in zero-shot Earth Observation (EO) downstream tasks, with further gains enabled by remote-sensing-adapted models. We examine this setting across 17 VLM variants and 12 remote sensing (RS) datasets under Meta-Prompting for Visual Recognition (MPVR), and show that zero-shot performance remains highly sensitive to textual design choices, from the meta-prompts used to guide the LLM in generating class descriptions to the descriptions themselves. We explore why semantically rich LLM-generated class descriptions do not translate into consistent gains over simple domain-adapted CLIP-style descriptions. While LLM descriptions are more semantically expressive, they can also introduce noise in the text embedding space, reducing robustness in downstream tasks. We support this observation through a text log-likelihood analysis in the whitened CLIP feature space, comparing LLM-generated and template-based descriptions. Building on this finding, we study query embedding calibration and show that lightweight calibration of the query space consistently yields strong improvements in zero-shot classification and retrieval. Overall, our results provide practical insight into the trade-off between semantic richness and robustness, and identify embedding calibration as a simple and effective tool for improving zero-shot remote sensing performance.
Remote sensing vision-language models are increasingly expected to support open-ended reasoning over Earth Observation data and a variety of tasks. Most recent progress in this area has been driven by remote-sensing-specific architectural designs, often introducing new encoders, alignment modules, or task-specific fusion mechanisms. In this work, we challenge the necessity of such architectural specialization. We show that a generally capable vision-language model can achieve competitive or state-of-the-art performance at challenging remote sensing benchmarks, provided that it is trained at sufficient scale across diverse data and tasks. Our model uses a single language policy that can either answer directly in text or invoke a localization tool for segmentation and grounding. To train this heterogeneous behaviour, we employ a multi-task reinforcement learning framework with adaptive task rewards covering multiple-choice VQA, free-form VQA, captioning, detection, and segmentation across a large variety of input types. Our approach achieves competitive results across a broad set of benchmarks, including high-resolution, multi-temporal, multi-modal and multi-view tasks. Further, as training data scales, our experiments show consistent improvements across most tasks both in and out of distribution, which correlate with per-task data diversity. These findings suggest that, for remote sensing VLMs, data scale is more important than architectural novelty.
Stefan Maria Ailuro, Mario Markov, Mohammad Mahdi +2
A robust Multimodal Large Language Model (MLLM) for Earth Observation should maintain consistent interpretation and reasoning under realistic input variations. However, current Remote Sensing MLLMs fail to meet this requirement. Trained on carefully curated clean datasets, they learn brittle mappings that do not generalize to noisy conditions in operational Earth Observation. Consequently, their performance degrades when confronted with imperfect inputs in deployment. To quantify this vulnerability, we construct a realistic set of multimodal perturbations, including visual degradations such as cloud and fog cover, together with diverse human-centric textual variations ranging from colloquialisms to vague or omitted instructions. Empirical evaluations show that these perturbations significantly impair the visual-semantic reasoning capabilities of leading RS foundation models. To address this limitation, we introduce RemoteShield, a robust Remote Sensing MLLM trained to maintain consistent outputs across realistic input variations. During training, each clean sample is paired with its image-text perturbed variants to form a semantic equivalence cluster. Rather than directly fitting noisy samples, RemoteShield is optimized through preference learning over clean and perturbed conditions within the same cluster. By comparing model responses to clean and corrupted inputs, the model is encouraged to favor stable responses over perturbation-induced failures. This cross-condition alignment helps the model focus on underlying task semantics despite visual degradations and textual noise. Experiments on three Earth Observation tasks show that RemoteShield consistently delivers stronger robustness and cross-condition consistency than representative baselines under realistic multimodal perturbations.
Open-vocabulary segmentation labels arbitrary categories from a text query without per-class training, yet on remote sensing imagery it underperforms on categories it handles reliably elsewhere. We find that much of this gap traces to the text query rather than to the segmentation model. Because these models are not specialized for overhead imagery, the class name that serves as the query is often a weak address into the vision-language embedding space. We show that a better name repairs part of the gap, while the remaining failures call for an address that the tested natural-language rephrasings do not provide. We recover that address from a few examples through textual inversion on a frozen model, keeping inference text only. On a representative benchmark this raises the mean intersection over union on the affected categories from 3.9 to 39.4, and across eight remote sensing datasets it improves over few-shot methods that instead inject visual prompts at inference.