Scale-aware Recognition in Satellite Images under Resource Constraints

Thank you for your review and comments. We hope to address your concerns and answer your questions below:

1. We use an off-the shelf LLM, specifically ChatGPT 3.5 for inferring scales of concepts. LLMs being utilized as tools to infer semantic information in the field of remote sensing has previously generally been focused on improving the understanding of satellite images. Here, instead we leverage the knowledge of these large models in the correct way by prompting using in-context examples. Details of these examples are described in section 6.3 (in the supplementary material). Here we compare our LLM approach to other baselines, and we will include additional results using other off the shelf LLMs (Gemini, Claude). We are happy to provide more details if you have any further questions.

We compared our LLM-based approach for inferring the best modality per concept with other baseline approaches. For each approach, we evaluate its accuracy in terms of determining the right modality. These baselines include either always choosing the HR modality and always choosing LR modality, or selecting between the two based on the average area of each concept, as provided via OpenStreetMaps (OSM). Here one can see that the LLM performs with 100% accuracy for our set of concepts, but it also capable of easily being extended to far more concepts.

2. Figure 3 visualizes the images with varying levels of disagreement. We agree that including a quantitative analysis would strengthen our results and will revise the paper to include details regarding the correlation between the two sets of model disagreement scores. The correlation coefficient between the disagreement scores of the HR (NAIP) vs KD with the LR model is 0.9322, which signifies a strong positive correlation. The correlation coefficient between the disagreement scores of the HR (NICFI) vs KD with the LR model is 0.998, which also signifies a strong positive correlation.

Questions:

1. The models are not limited to locations only seen during training. Eq. 5 pertains to calculating this disagreement between LR and KD for a location. We assume we can acquire LR images for all locations at inference time. We can then calculate this disagreement score using the LR and KD models with the LR image.

2. We discuss testing the LLM on unseen classes in both sec. 4.2.4 on page 9 as well as more in depth on page 2 of the supplementary materials, with Table 5 tabulating these results on our validation set and comparing our technique to others.

The cost of querying an LLM with a few words is essentially free compared to the cost of acquiring HR satellite imagery: one does not need to host a LLM locally, we simply query existing LLMs. We used ChatGPT 3.5 for our experiments, which is free, and we only need to query the LLM once for any given concept, since once we know which resolution is optimal, we don’t need to query again. As for why we do not train a network to estimate the HR and LR scores, we want to ensure we can generalize to unseen concepts.

Thank you for your insightful comments and review, we address your questions and concerns below:

Weaknesses:

1. Yes indeed our approach relies on in-context examples. LLMs have surprisingly broad knowledge on the world and may be able to extrapolate more information than we would expect, but one could provide additional in-context examples or further prompting to allow the LLM to generalize to more niche concepts.

We use an off-the shelf LLM, specifically ChatGPT 3.5, here. The LLM semantically understands the scale of different concepts, for example if one were to ask “how large is a tennis court”, most LLMs are able to answer with high levels of accuracy. With concept scale information, we can also ask the LLM whether or not the concept would be visible in a given resolution of a satellite image. A good way to think about it is that if the LLMs can answer “Is a tennis court better suited to a 10m per pixel resolution image or a 1m per pixel resolution image, given that a swimming pool is suited to 1m and a coastline is suited to 10m”. Then for any concept, regardless of how commonly it is searched for within satellite imagery, the LLM would know roughly its size and then be able to infer its visibility for a given resolution.

An alternative/future work can explore prompting the LLM with the true scale and specific resolutions of the satellite. We will discuss this limitation further in the weaknesses section of our paper.

2. A lot of the use cases of satellite imagery require accuracy. With any in-painting or super resolution techniques, there is room for the model to hallucinate. Take for example researchers in the fields of anthropology [1], who report to Congress or the United Nations about specific incidents of cultural erasure. For the concepts they are keeping an eye on, there’s essentially 0 tolerance for hallucinations in these cases as this work can have major socio-political implications.

[1] ©2023, C. H. W. (2024b, June 1). Between the wars. ArcGIS StoryMaps. https://storymaps.arcgis.com/stories/e1c69b7dd46f4c839dffc0fab9248368

Questions:

1. The idea of using a super resolution model in place of the KD model for model disagreement based sampling is an insightful suggestion! We test this hypothesis using a SoTA super resolution (SR) model [2] and present the results below.

Generating and performing inference using the SR data proved to be significantly more computationally expensive. Here we report inference time, one can see the process is significantly slower than using the KD model on the LR images. This is due to the fact that the SR model accepts 32x32 pixel patches of Sentinel-2 images at a time. So for our test set of 5015 LR images of size 256x256 pixels, the SR model is performing inference on 320,960 patches. This process took 12,549.90 s (~3.5 hours). Additionally upon stitching these SR patches to recreate the coverage of the original LR image, we must split the stitched SR image into 100 components corresponding to the same organization of HR images. This leads to inference using the HR model on 501,500 images. And even after this, one would still need to run the KD model for regions not covered by the budget. Overall this significant jump in inference time makes using any superresolution model in place of our KD model less favorable.

Additionally, our KD model is specifically trained to identify cases where HR data would provide meaningful improvements in recognition accuracy. In contrast, SR models attempt to reconstruct all high-frequency details, regardless of their importance to the recognition task. This targeted approach makes our method more efficient at identifying truly necessary cases for HR image use.

Here we present the results using the SR images with the HR model to calculate the model disagreement score to sample HR imagery. LR + SR + LLM denotes the use of both the LLM and model disagreement components of our work, (using the SR model disagreement), without the use of any KD model. KD + SR + LLM denotes results with our full system, but with the model disagreement scores being obtained via the SR imagery.

While it does offer improvements in mAP for unseen classes, the inference time alone is 34 times longer than our technique, and this improvement only comes with the use of our KD model.

2. The impact of resolution ratio is an important one, we discuss this limitation on line 513. Intuitively, if the ratio is closer to 1, then one would expect there are fewer concepts that the models would disagree on as there will be considerable overlap between which concepts have visual signatures and vice versa as the ratio approaches 0. However in practice we found that in cases where the resolution ratio was 1:100 (as is the case for our NAIP benchmark) or 1:4 (as is the case in our NICFI benchmark), both KD model disagreement scores were highly correlated with their HR counterparts.

With the correlation coefficient between the disagreement scores of the HR (NAIP) vs KD with the LR model is 0.9322, which signifies a strong positive correlation and the correlation coefficient between the disagreement scores of the HR (NICFI) vs KD with the LR model is 0.998, which also signifies a strong positive correlation. To carry out this experiment we calculate the ranked Spearman correlation coefficient between the two sets of disagreement scores.

[2] Wolters, Piper, Favyen Bastani, and Aniruddha Kembhavi. "Zooming out on zooming in: Advancing super-resolution for remote sensing." arXiv preprint arXiv:2311.18082 (2023).

Thank you for your review and questions. We hope to address your concerns and answer your questions below:

Weaknesses:

1. We discuss the limitations of OpenStreetMaps in the limitations subsection of our Results (Line 518). If there are any other datasets the reviewer recommends we use to demonstrate the accuracy of our models, we are happy to do so as we believe it would strengthen our results.

As reported in the paper, we acknowledged that OpenStreetMaps data contains noise, due to its crowd-sourced nature. Unfortunately, there are not many sources for fine-grained labeling of satellite imagery. Therefore, we follow past work in using OSM for training and evaluation (Mall et al., 2024; Bastani et al., 2023). Moreover other standard sources of labeled satellite imagery such as Microsoft’s building footprint dataset work with similar noise levels (Bing Maps, 2018). Additionally, to show robustness to this noise, we performed an experiment using the GRAFT LR model, wherein we constructed bootstrap samples by sub-sampling within our original test set 1000 times. The results show that the 90% confidence interval is within 2.43% of the reported mAP.

2. Yes we can retrieve multiple targets simultaneously. Each model provides the scores for all query concepts at once as logits. So for all images, we run inference using the LR model just once. If some of the query concepts require HR imagery, we run inference on the images with the KD model (again just once). We then calculate the disagreement (which is shared across all concepts), and retrieve how many ever HR images the budget allows for and evaluate using the HR model only once.

3. We address time discrepancies on line 324 in the Benchmark subsection. “To handle time discrepancies we retrieved the LR image for a set of HR within the same month that the HR image was collected to avoid major disagreements between the two sets of images.”

4. We do not include runtime in tables 3 and 4 as these tables are used to demonstrate ablation results on LR imagery alone. We include runtimes for all of the techniques in tables 1 and 2 (page 8).

Questions:

Deployment: Indeed, one would need to build appropriate storage systems for practical deployment. But storage systems for data and detection models like this are already being done manually by anthropologists, archaeologists, soil and crop scientists, amongst many others. We offer an efficient alternative to the manual overhead, and we believe there is limited additional cost over how retrieval in satellite imagery is currently being done. Our focus is on tackling retrieval and diving deeper into how the size and scale of a concept is best suited to different resolutions, and how this information can be used to achieve better retrieval results given cost constraints.

We further believe that the systems challenge notwithstanding, the problem of effective retrieval in the face of the cost of data is a pervasive challenge in many applications. Many researchers in the fields of anthropology [1], among others, manually search through large amounts of HR satellite imagery to pinpoint sites of interest and then purchase HR imagery to monitor these sites of interest a few times a year (2-12) due to high costs. Our technique would greatly help in both the initial search for points of interest and for monitoring for these sites in between the HR image acquisitions, and the storage space and retrieval needs of the LR imagery is far lower than that of the HR imagery. We hope that including the information regarding the scale of the datasets and the inference runtime help alleviate concerns regarding overall process time and deployment.

[1] ©2023, C. H. W. (2024b, June 1). Between the wars. ArcGIS StoryMaps. https://storymaps.arcgis.com/stories/e1c69b7dd46f4c839dffc0fab9248368

[2] Mall, Utkarsh, et al. "Remote Sensing Vision-Language Foundation Models without Annotations via Ground Remote Alignment." ICLR 2024

[3] Bastani, Favyen, et al. "Satlaspretrain: A large-scale dataset for remote sensing image understanding." ICCV 2021

[4] Microsoft. (2018). Microsoft/USBUILDINGFOOTPRINTS: Computer generated building footprints for the United States. GitHub. https://github.com/microsoft/USBuildingFootprints

Thank you for your insightful review and feedback. We hope to address your concerns and answer your questions below:

Weaknesses:

1. The final result is impacted by budget, in the sense that we are able to have improved retrieval results with far fewer HR images. We can look to Table 2, our model (bottom row) outperforms the solely HR model. We are able to get an improved accuracy on a tight budget since we only sample HR images when necessary. The reason we are able to do so is (a) some concepts are better in LR and (b) for the remaining concepts we are able to recognize which regions require HR. Of course if the budget decreases further, at some point, the accuracy will also decrease. We demonstrate this in the table below with results using our method for varying budgets.

2. We will add further details on each of these components and release the code upon acceptance. If there are any specific clarifications we are happy to address them! With regards to the LLM, we use an off-the shelf LLM specifically ChatGPT 3.5 which is free to the public and we use the prompts specified in the supplemental materials.

3. We are happy to clarify the specifics of our retrieval process, as described in section 3.6 (lines 272-281). Given a concept, our LLM determines which resolution is best suited. If the best suited resolution is LR, we perform inference using the LR model and all LR imagery only. If the best suited resolution is HR, we ask the user for their HR budget. Then we perform model disagreement using LR imagery alone and the LR and KD models. With these disagreement scores and the user budget, we then sample the top locations, obtain HR images and perform inference using the HR model for these locations. The retrieval results for the remaining area, that it is preferable to use HR for, but cannot be fit in the budget, is scored using the KD model on LR imagery.

Questions:

1. Our model disagreement component only comes into play when we are searching for concepts we know are better detected in HR. So for these concepts, if there is any disagreement between the HR and LR models, regardless of the direction of the disagreement, we know to sample HR imagery.

2. As we mentioned earlier, we use an off-the shelf LLM specifically ChatGPT 3.5, and provide in-context examples as described in section 6.3 (in the supplementary material). Here we compare our LLM approach to other baselines, and we will include additional results using other off the shelf LLMs (Gemini and Claude). We are happy to provide more details if you have any further questions.

We compared our LLM-based approach for inferring the best modality per concept with other baseline approaches. For each approach, we evaluate its accuracy in terms of determining the right modality. These baselines include either always choosing the HR modality and always choosing LR modality, or selecting between the two based on the average area of each concept, as provided via OpenStreetMaps (OSM). Here one can see that the LLM performs with 100% accuracy for our set of concepts, but it is also capable of easily being extended to far more concepts.

3. In some senses, yes, we can use the largest possible LR image, however in practice we fix the image size and use more tiles.

4. You are right! Thank you for pointing this out to us, this section heading is a typo. We apologize for any confusion and will correct it to “retrieval” in the revision.