Towards Text-Only Remote Sensing Scene Classification via Semantic Proxy
Organizations: Faculty of Geosciences and Engineering, Southwest Jiaotong University, Chengdu 611756, China. · Society Hub, The Hong Kong University of Science and Technology (Guangzhou), China, Guangzhou 511453, China. · School of Artificial Intelligence, Wuhan University, Wuhan 430072, China.
Abstract
Remote sensing scene classification (RSSC) is traditionally formulated as a vision-centric problem, where inference relies on rich visual representations extracted from images. In this work, we revisit this assumption by introducing text-only RSSC, a new problem setting that investigates whether remote sensing scenes can be recognized using only textual semantic representations. To enable this study, we propose SeProRS, which reformulates RSSC as a text-only classification task through semantic proxy generation. Specifically, a vision-language model (VLM) is first employed to convert remote sensing images into textual descriptions, termed semantic proxies, which are then encoded by a text embedding model for scene classification. This decoupled pipeline provides a unified and reproducible benchmark for evaluating text-only scene understanding. Furthermore, we investigate representative prompting strategies and proxy generators, and analyze their impact on classification performance, text embedding behavior, cross-generator transferability, and efficiency. Extensive experiments on three well-known RSSC datasets show that semantic proxies can preserve substantial discriminative information, while also revealing inherent limitations of purely text-based scene understanding. This paper does not aim to replace image-based RSSC, but instead establishes a new research direction by quantifying the feasibility and boundaries of semantic proxy-based scene recognition. The proposed benchmark and analysis provide a foundation for future research on efficient and modality-reduced remote sensing scene understanding. The source code and datasets will be accessible at https://github.com/sstary/SSRS.
Figures & tables
| Representation | Granularity | Image-specific | Main Function |
| Class label | Category-level | No | Defines the prediction target |
| Class prompt | Category-level | No | Provides textual class prior |
| Generic description | Instance-level | Yes | Describes visible scene content |
| Semantic proxy | Instance-level | Yes | Preserves task-relevant evidence for text-only classification |
| Dataset | Classes | Total images | Spatial resolution | Training/Test ratio | Train/Test images |
| UCM | 21 | 2,100 | 0.3 m | 50% / 50% | 1,050 / 1,050 |
| 80% / 20% | 1,680 / 420 | ||||
| AID | 30 | 10,000 | 0.5-8 m | 20% / 80% | 2,000 / 8,000 |
| 50% / 50% | 5,000 / 5,000 | ||||
| NWPU-RESISC45 | 45 | 31,500 | 0.2-30 m | 10% / 90% | 3,150 / 28,350 |
| 20% / 80% | 6,300 / 25,200 |
| ID | Generator | Access | Model Scale | Parameters | Role |
| G1 | LLaVA-OneVision-Qwen2-0.5B [ 38 ] | local | Small | 0.5B | lightweight open VLM |
| G2 | Qwen2.5-VL-3B [ 39 ] | local | Medium | 3B | efficient open VLM |
| G3 | Gemini-2.5-Flash [ 40 ] | API | Commercial | Not disclosed | closed-source strong VLM |
| Item | Setting |
| Input | Remote sensing image and task-agnostic descriptive prompt. |
| Label control | No class name, ground-truth label, or candidate label list is provided to the VLM. |
| Generators | LLaVA-OneVision-Qwen2-0.5B, Qwen2.5-VL-3B, Gemini-2.5-Flash. |
| Prompt types | Flat Prompt and Structured Prompt. |
| Flat Prompt | Generates one concise paragraph describing scene-level semantics. |
| Structured Prompt | Generates global description, object-level attributes, and functional cues. |
| Method | 50% training | 80% training |
| Vision-based Methods | ||
| HFAM | ||
| CGINet | ||
| UPetu | ||
| MSCN | ||
| UniRepLKNet | ||
| Method | 20% training | 50% training |
| Vision-based Methods | ||
| HFAM | ||
| CGINet | ||
| UPetu | ||
| MSCN | ||
| UniRepLKNet | ||
| Method | 10% training | 20% training |
| Vision-based Methods | ||
| HFAM | ||
| CGINet | ||
| UPetu | ||
| MSCN | ||
| UniRepLKNet | ||
| Train/Test | G1 | G2 | G3 |
| G1 | |||
| G2 | |||
| G3 |
| Type | Method | Max OA (%) | Params (M) | Train Time (s/it) | Test Time (s/it) | Max Mem. (MB) |
| Vision | DINOv3-ViT-B/16 Adapter | 98.58 | 7.12 | 0.3025 | 0.2861 | 8543 |
| Vision | UniRepLKNet-T | 97.88 | 31.07 | 0.4557 | 0.3030 | 15793 |
| Text | SeProRS | 95.20 | 0.54 | 0.0701 | 0.0935 | 1266 |
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| Item | Setting |
| Text encoder | bge-base-en-v1.5 |
| Encoder update | Frozen |
| Maximum text length | 192 tokens |
| Projection adapter | LayerNorm, Linear , GELU, Dropout |
| Classifier head | LayerNorm, Linear , GELU, Dropout, Linear |
| Loss | Cross-entropy |
| Item | DINOv3 Adapter | UniRepLKNet-T |
| Input size | ||
| Pretraining | DINOv3 ViT-B/16 | ImageNet |
| Training epochs | 200 | 200 |
| Batch size | 128 | 128 |
| Optimizer | AdamW | AdamW |
| Base learning rate |
| Dataset | Image | VLM | Prompt | Semantic proxy |
| UCM | agricultural00.tif | Flat | The image depicts a large, well-organized agricultural field with rows of crops. The field is divided into multiple sections by straight lines that run parallel to each other, indicating the use of precision farming techniques. The crops appear to be in good health and are likely being cultivated for their potential yield or market value. There are no visible structures or people within the field, suggesting it may be an open space used for crop cultivation rather than a farmstead. | |
| UCM | agricultural00.tif | Structured | The image shows a large, well-maintained agricultural field with neatly arranged rows of crops. The field is bordered by a dirt road and there are trees visible in the background. Buildings can be seen lining the perimeter of the field, indicating residential or commercial structures. Roads run parallel to the fields, providing access for vehicles and farming equipment. Vegetation includes grasses and other plants that cover the soil, contributing to the overall health of the crop. The scene appears to be an active agricultural area, likely used for growing crops such as grains or vegetables. The presence of roads suggests easy transportation for farm equipment and supplies. The buildings could be storage facilities or offices related to the operations of the farm. | |
| UCM | agricultural00.tif | Flat | The remote sensing image depicts a large agricultural field with rows of crops, likely wheat or corn, stretching across the landscape. The field is bordered by a dense forest on the left side, indicating a transition between agricultural land and natural vegetation. The functional context suggests this area is used for crop cultivation, possibly for commercial purposes such as grain production. | |
| UCM | agricultural00.tif | Structured | The image depicts a large agricultural field with rows of crops, likely wheat or corn, stretching across the landscape. The field is surrounded by a mix of trees and open areas, indicating a rural setting. The functional use of this area is primarily agricultural, designed for crop cultivation and harvesting. | |
| UCM | agricultural00.tif | Flat | An expansive agricultural scene characterized by numerous, precisely aligned green crop rows covering the majority of the area, hinting at large-scale cultivation or a nursery, with a small patch of denser, irregular vegetation in the upper left corner and a bright, rectilinear artificial structure at the bottom right, collectively depicting an organized rural or peri-urban agricultural landscape. | |
| UCM | agricultural00.tif | Structured | The scene depicts a predominantly agricultural landscape characterized by extensive cultivation. It features numerous parallel crop rows, a small patch of natural vegetation, and a potential artificial structure at the periphery. This area is primarily used for agricultural production, indicating a farming context. |
| Dataset | Image | VLM | Prompt | Semantic proxy |
| AID | airport_1.jpg | Flat | The image is an aerial view of a busy airport tarmac, showcasing multiple airplanes parked at gates. The tarmac is marked with various lines and symbols indicating taxiways and other operational areas. In the center of the image, there’s a large jet bridge connecting one terminal to another, facilitating passenger boarding and deplaning. Adjacent to this jet bridge, several smaller aircraft are stationed on the ground, likely awaiting their turn for departure or having just arrived. The overall scene represents a typical day at an airport, with planes ready for takeoff and passengers waiting in anticipation. | |
| AID | airport_1.jpg | Structured | The image shows an aerial view of a busy airport terminal with multiple jet bridges connecting the terminals. There are several airplanes parked at gates, and various buildings and roads can be seen surrounding the airport. Buildings include large hangars, terminals, and control towers, while roads are visible as well as parking areas for vehicles. Vegetation is present in the form of grassy patches near the terminal buildings. The scene represents transportation-related activities, with airplanes being the primary mode of transport. The presence of jet bridges indicates that passengers can board or disembark from aircraft directly to the terminal. The overall setting suggests a commercial airport environment. | |
| AID | airport_1.jpg | Flat | The image depicts an expansive airport terminal with multiple runways and taxiways, featuring several aircraft parked at gates. The terminal is surrounded by various ground support equipment and vehicles, indicating active operations. The surrounding area includes parking lots and possibly maintenance facilities, suggesting a well-equipped airport infrastructure. | |
| AID | airport_1.jpg | Structured | The image depicts an expansive airport terminal with multiple runways and taxiways, surrounded by various infrastructure elements including parking lots and service areas. The terminal is characterized by its long, linear structure with numerous windows and doors, indicative of passenger boarding and disembarking points. The functional aspect of this scene is primarily transportation, serving as a hub for air travel with clear demarcations for aircraft movement and ground operations. | |
| AID | airport_1.jpg | Flat | An expansive airport facility is depicted, featuring a prominent, elongated passenger terminal building with numerous aircraft parked at gates along its length, surrounded by vast concrete aprons, intricate taxiways, and portions of runways, alongside a distinct landscaped circular structure near the main terminal, all indicative of a busy aviation hub designed for air travel and transport. | |
| AID | airport_1.jpg | Structured | The scene depicts a large-scale developed area characterized by extensive paved surfaces and structures, indicative of an airport. Numerous aircraft are visible parked alongside a long, multi-sectioned terminal building, surrounded by an intricate network of runways, taxiways, and aprons. This area serves as a critical hub for air transportation, facilitating the movement and servicing of aircraft and passengers. |
| Dataset | Image | VLM | Prompt | Semantic proxy |
| NWPU-RESISC45 | airplane_001.jpg | Flat | The image shows a scene of three airplanes parked on an airport tarmac. The planes are positioned in a staggered formation, with the middle plane slightly ahead of the other two. Each airplane is white and appears to be a commercial jet airliner. The tarmac surface is a light grey color, typical for such environments. There are no visible markings or identifiers that would indicate the specific location or airline. The overall scene type is remote sensing, as it captures aerial imagery rather than ground-level observations. | |
| NWPU-RESISC45 | airplane_001.jpg | Structured | The image shows a large, open area with three airplanes parked on the tarmac. The planes are white and appear to be commercial airliners. There is also some vegetation visible in the background. The airplanes are likely used for commercial purposes such as air travel or cargo transport. The tarmac appears to be made of concrete and there are no other objects or structures visible in the immediate vicinity. The scene suggests that it could be an airport or a transportation hub where aircraft are parked before takeoff or after landing. The presence of multiple airplanes indicates that this location serves a significant number of passengers or cargo operations. | |
| NWPU-RESISC45 | airplane_001.jpg | Flat | The image depicts an aerial view of an airport runway with three airplanes in flight. Two aircraft are positioned closer to the camera, appearing larger due to their proximity, while the third is further away, smaller in size. The planes are likely commercial jets, given their size and design, and they are flying at a low altitude, suggesting they might be preparing for landing or taking off. The runway is clearly marked, indicating its functional use for aircraft operations. | |
| NWPU-RESISC45 | airplane_001.jpg | Structured | The image depicts an airport with multiple airplanes on the tarmac. The ground is paved with asphalt, indicating a well-maintained runway area. Airplanes are present, suggesting active air traffic at this location. | |
| NWPU-RESISC45 | airplane_001.jpg | Flat | An airbase or large airport scene is presented, characterized by multiple large, propeller-driven transport aircraft prominently parked on a paved tarmac or apron, strongly indicating a military or heavy logistics operational facility. | |
| NWPU-RESISC45 | airplane_001.jpg | Structured | This is an aerial scene dominated by an airport or airfield environment. Multiple propeller aircraft are visible parked on large, untextured concrete or asphalt surfaces. The scene clearly indicates a transportation function, specifically for air travel and aircraft operations. |