Active Learning-Based Species Range Estimation

We propose a new active learning approach for efficiently estimating the geographic range of a species from a limited number of on the ground observations. We model the range of an unmapped species of interest as the weighted combination of estimated ranges obtained from a set of different species. We show that it is possible to generate this candidate set of ranges by using models that have been trained on large weakly supervised community collected observation data. From this, we develop a new active querying approach that sequentially selects geographic locations to visit that best reduce our uncertainty over an unmapped species' range. We conduct a detailed evaluation of our approach and compare it to existing active learning methods using an evaluation dataset containing expert-derived ranges for one thousand species. Our results demonstrate that our method outperforms alternative active learning methods and approaches the performance of end-to-end trained models, even when only using a fraction of the data. This highlights the utility of active learning via transfer learned spatial representations for species range estimation. It also emphasizes the value of leveraging emerging large-scale crowdsourced datasets, not only for modeling a species' range, but also for actively discovering them.

Paper

References (47)

Scroll for more · 35 remaining

Similar papers

Peer review

Reviewer RtuW6/10 · confidence 4/52023-06-28

Summary

This paper tackles the problem of species distribution modeling from limited observation data. An active learning approach is proposed that makes the assumption that an unmapped species can be represented by a weighted combination of candidate species models (previously trained offline on a disjoint dataset). A sampling strategy is described that uses the aggregate uncertainty of the candidate models to select the next location to query. For representing location, a pre-existing feature representation is derived from a model trained on weakly labeled presence-only data. Results on two benchmark datasets outperform baselines and an ablation study is presented.

Strengths

- Addresses an important and interesting problem. Species distribution modeling has received a great deal of attention in the literature lately, especially related to fine-grained visual categorization. This work is of the first to explore how active learning learning can be applied to this problem to estimate range models from limited amounts of observation data. - The proposed approach has technical novelty. The key idea is to take advantage of an existing set of species range models as a hypothesis set. During training, a range model for an unseen species is estimated by a weighted average of the candidate models. To capture fine-details, this candidate set is augmented with a model trained on the accumulated observations of the new species. In addition, a modified form of uncertainty sampling is also used which leverages the candidate models. - Extensive evaluation, including an ablation study. The results indicate that design decisions directly lead to performance improvements, for example the hypothesis set sampling approach and the inclusion of the online model in the candidate set. - Well written, easy to follow. - Sufficient detail provided that the proposed method is likely reproducible.

Weaknesses

- The proposed approach is a relatively straightforward application of active learning beyond the introduction of the candidate models. - The approach is highly dependent on the hypothesis set (candidate models) being diverse such that a new species can be represented by some combination of them (L190). To partially address this, the authors add a model trained on the accumulated observations to the hypothesis set. - The evaluation is largely limited to simple baselines in the form of a self-ablation (incrementally adding to an active learning baseline). Only one recent paper is compared against (not active learning-based), with diverging results across the two benchmark datasets. That said, it appears there are no obvious other baselines to compare against.

Questions

My initital rating is a weak accept. This paper is exceptionally well written and was fun to read. The use of active learning for species range modeling has real-world utility. The proposed approach, while simple, has novelty in its construction and the results improve over baselines. I think this paper will be of interest to the community. Suggestions: - An interesting byproduct of the proposed approach would be to show which species contributed (and their range models) the most in the weighted average. - The annotations on Fig. 4 are hard to see without zooming in.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Limitations sufficiently addressed in the manuscript.

Reviewer yX8t7/10 · confidence 3/52023-07-07

Summary

**Paper Summary** This paper tackles the data-driven estimation of geographical habitat area for a given species; that is, given a certain geographic location (i.e., a "point on the map"), the goal is to predict how likely a certain species is present in this location. Prior efforts train models in an "offline" setting, assuming a finite training data set that indicates (for a fixed set of different species) the presence (or absence) of the species. Datasets that signal the presence (and possibly absence) of a large number of species are readily available. This paper points out that dataset expansion for species not covered by existing datasets would be cumbersome- we would need to collect presence indication signals for new (or not yet recorded) species worldwide by possibly uniformly sampling presence indication evidence, which can then, in turn, be used to update/re-train existing models. This paper investigates an interesting angle of this general problem: can we leverage *existing* (already collected) datasets to obtain a "weak" classifier for a new species, which can then guide the data collection efforts to regions, such that with as few samples we can decrease the overall presence entropy *as much as possible*? This is akin to active learning, where the idea is to label as-informative-as-possible recorded data points. Different to that, in this case, the model suggests geographic regions in which we should sample the data for the new species (i.e., obtain presence/absence signals), such that the range estimates for new species will be as accurate as possible with as few data points (i.e., manual collections) as possible. The overall idea behind the proposed work is to leverage the similarity between the (new) query species and existing species: this is based on the intuition that "similar species" (in terms of similarity in the feature space) will likely habitat similar territories. Such an approach can then be applied iteratively by incorporating newly-obtained presence/absence signals, suggesting increasingly more informative locations for the data collection. The idea behind the approach is simple and intuitive: assuming an ensemble of (trained) linear models for a set of species, a model for a new species can be estimated as a weighted combination of existing linear models (where the model weight is proportional to the similarity in the feature space between the new (query) species and species database). This model then provides a "heatmap" over geographic locations, and guides the data collection -- new data points can be collected in the areas where the model is most uncertain about the new species' presence. **Core Contributions/Insights** The core contribution of this work is the insight that techniques similar to active learning can be used to streamline the data collection process. This is based on the key insight that existing datasets contain a sufficiently large number of species samples, such that a linear combination of already trained linear classifiers provide a reasonably good predictor for the presence of new species. **Key Results** The paper convincingly demonstrates that the proposed hypothesis ensembling outperforms logistic regression, and adding an online-trained model to the ensemble further improves the performance. I cannot comment on whether any relevant baselines are missing, as I am not actively working in species range estimation or active learning, and I am, therefore not familiar with related work. I hope my colleagues can comment on this. Experimental validation also discusses the impact of the number of existing species (models) for model averaging on the final performance and establishes that the proposed approach is useful when a large number of specie presence models is readily available. **Potential for Impact** Specie habitat estimation has the potential to assist scientists in monitoring how the habitats of species are changing over time, which is especially important in the era in which global warming presents an existential threat to several species. Therefore, data-driven methods that can provide reasonable estimates based on (I assume difficult to obtain) evidence of species have important practical implications. As monitoring (and consequentially, data collection) is a continual process, developing models that can guide data collection efforts using active learning techniques makes sense. **Rating Justification** I think this paper presents an interesting, well-motivated, and well-executed work with direct practical relevance to monitoring species habitats. At this point, I see no obvious objections to why this paper should not be accepted. This is a low-confidence review written by a reviewer entirely unfamiliar with bird-species range estimation and only weakly familiar with the field of active learning. **Post Rebuttal Comments** I want to thank the authors for their thorough responses. All my questions were addressed, and I see other reviewers are also positive about this paper. I am upgrading my rating to accept.

Strengths

* Application relevance: this work has the potential to guide future data collection efforts for species range estimation. Monitoring species range is important for understanding how the earth's biosphere evolves under rapidly changing environmental conditions. * Paper presentation: this paper is very well written and summarizes well both the application domain (species range estimation), as well as existing methods that are commonly being used, together with simple and intuitive taxonomy of prior efforts. Core contributions are presented clearly, well-motivated, and experimentally validated. As someone reading a paper on species range estimation for the first time, I really appreciate Sec. 3, which summarizes how this problem is typically tackled in the literature. * Motivation: the paper makes a convincing case for investigating which areas should be used to collect data in the next training cycles using active learning techniques. * Results: I find the paper's claims convincingly validated.

Weaknesses

In addition to the weaknesses discussed in the limitation section, it would be fair to point out that the proposed approach is most likely *only* applicable to new species that are sufficiently similar to species for which data is readily available. I assume this approach would fail in the case of potentially newly-discovered species that would hypothetically be quite unlike anything that was observed before. As I am not familiar with the biological aspect of this task, I am not sure if this is a practical concern.

Questions

Did the authors perform any kind of analysis on whether the intuition that most related species are a good indicator for the presence of new species holds? I.e., for a query species, are highly-weighted classifiers indeed representing most related species, and their presence is a good indicator? Did the authors observe any failure cases? For example, the presence of domestic cat, which is related to wild cats, is probably not a good presence indicator for, e.g., lion.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The paper thoroughly discusses the limitations in a designated section. I find this section insightful.

Reviewer B3F84/10 · confidence 4/52023-07-09

Summary

The paper explores a method for geographical range estimation of a species. One of the primary challenges in range estimation is the scarcity and costliness of data collection, particularly for rare species. To address this issue, the paper introduces an active learning framework as a solution. The proposed method offers two key advantages. Firstly, it utilizes a joint feature representation that combines information from multiple species, enhancing the accuracy of the range estimation. Secondly, it employs an active learning scheme to efficiently collect and learn from data to estimate a geographical range of new species. The paper also proposes a linear combination of learned classifiers to develop a new classifier for each new species. The effectiveness of the proposed method is evaluated on two datasets: the IUCN dataset and the S&T dataset.

Strengths

The paper is written clearly. The description of the problem setup and the technical approach is intuitive, allowing readers to grasp the concepts easily. The paper makes a significant contribution to the field of species range estimation. It achieves a notable performance increment and thoroughly explores the problem with a comprehensive analysis. Furthermore, the paper provides a comprehensive analysis of the proposed method, particularly focusing on the active learning scheme. This includes multiple stages such as initial representation learning, acquisition function, and model training. The authors thoroughly examine and analyze the proposed active learning framework, providing a comprehensive understanding for each components.

Weaknesses

While the paper thoroughly analyzes its proposed method and provides empirical results for the specific problem of range estimation, it does not explicitly highlight the unique machine learning challenges of this problem or how they connect to the technical contribution of the paper. Range estimation can be seen as sharing commonalities with other common machine learning problem settings. For instance, the presence-absence data setting can be considered a supervised classification problem, while the presence-only data setting resembles one-class learning or semi-supervised learning, where the labeled set only contains positive samples. The paper primarily focuses on the application aspect by developing an efficient active learning framework for the range estimation problem. However, it does not provide a detailed explanation of why the chosen solution, which employs widely used machine learning techniques, was devised among numerous alternative approaches. The paper would benefit from explicitly stating the unique machine learning challenges posed by the range estimation problem and how the proposed method effectively addresses those challenges. Additionally, a more comprehensive discussion of the motivation behind the chosen approach would enhance the clarity and understanding of the paper.

Questions

What is the unique machine learning challenge of the range estimation problem that distinguishes it from other machine learning problems or applications? Equation (3) averages the classifier parameters. How does the model perform if averaging is applied to the output rather than to the parameters? Why did the authors decide to apply the average on parameters rather than output? Equation (5) uses average over outputs while equation (3) averages over parameters. Why should uncertainty computation (Eq. 5) be different from inference computation (Eq. 3)? Is there an intuition behind this? What happens if the pretrained classifiers are not used, but instead train a new species classifier from scratch on S_t? This is an important baseline but is missing in the paper. It is interesting that nn_rand performs way better than the other two features; loc_feats and env_feats. What would be the reason for this? Is nn_rand trained on S_t? Or is nn_rand fixed as the random initial weights?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

2 fair

Limitations

The paper addresses the limitations of the proposed method and its broader impacts.

Reviewer yFB16/10 · confidence 3/52023-07-13

Summary

This paper proposes an active learning framework for estimating the geographical range of species from sparse observations. The core idea is to exploit a large set of pretrained range estimation models from many different species. This paper also provides extensive experiments including detailed ablation studies and comparisons with baselines, demonstrating the effectiveness of the proposed active learning approach.

Strengths

1. It is technically sound to exploit a large set of pretrained range estimation models from diverse species to reduce the required observations for training a new model. 2. This paper provides detailed ablation studies and comparisons with existing methods and baselines and convincingly demonstrates superior performance. 3. This paper is well-written and easy to follow.

Weaknesses

1. The technical novelty is limited. It is reasonable yet straightforward to exploit a set of pretrained range estimation models for training a model for a target species. 2. The exact reason why it is meaningful to learn a linear combination of pretrained range estimation models needs further clarification. Is this because the range of different species is highly correlated and the learned weights indicate some correlation among different species, or such a linear combination indicates some general habitat candidates? A baseline probably stronger than the LR_random (Fig. 2) is to train a general classifier on joint presence data from all available species. Such a general model may indicate whether a given location is a good habitat for birds or other animals. 3. Is it visualize the weights in the linear combination for some target species? Do the weight visualizations indicate any correlations among species? If so, is such a correlation related to species taxonomy? It might also be useful to incorporate the species taxonomy as input to the model.

Questions

1. Could you elaborate on the details of LR_HSS? For HSS (hypothesis set selection)? If logistic regression instead of weighted averaging is applied during training, how to apply HSS (Eqn. 5) at the test time? It looks like the weights per hypothesis H is required during active querying. 2. Typos: #199-#200, WA_HHS should be WA_HSS.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

1. Limited technical novelty. The core idea of this paper is to exploit a set of pretrained range estimation models to reduce the number of observations required for training a new model for target species. This problem studied in this work is very specific to the range estimation problem. 2. A stronger baseline that trains a general classifier should be considered as described above. 3. The role of weighted averaging needs more clarification.

Reviewer WCfR6/10 · confidence 4/52023-07-14

Summary

This paper proposes to: 1) Use crowdsourced data to obtain species range maps for a large set of data-rich species, and then 2) Learn a linear combination of these maps to represent the range of other, data-poor, species, while minimizing the number of required presence/absence queries via active learning.

Strengths

The paper is well motivated, well structured, reads well and provides a good overview of the existing literature. The proposed method is simple and intuitive, and the performance improvements quite substantial.

Weaknesses

Although I'm quite positive about the paper, I do have some small doubts about whether a more general context placement would be required for a better fir in NeurIPS. I'm no expert, but it seems to me that the authors could connect the work to methods beyond species range estimation in order to connect better with the ML community. After all, the proposed method seems to perform reconstruction with a predefined basis. Otherwise, it seems like a paper more oriented exclusively towards the SDM community.

Questions

As reflected in the weaknesses and limitations sections: 1) What other fields of application and methods in the ML community would have connections to the presented methodology? 2) Would this method eventually be feasible in reality?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Most limitations are properly addressed. The only thing I missed would be the feasibility of obtaining presence/absence queries. Is this something that has been done? Obtaining a presence given that the species is present is one thig, but obtaining an absence could be very challenging. Would this method eventually be feasible in reality?

Reviewer yX8t2023-08-12

Thank you!

Thanks for the thorough response! I am happy with the rebuttal, and I didn't see any other major weaknesses raised by my colleagues. I upgraded my rating to accept.

Reviewer WCfR2023-08-14

I appreciate the authors' response, as I think it would be important to touch these two aspects in the paper.

Authorsrebuttal2023-08-20

We agree, we are very happy to add this - thanks for the suggestion. These will be easy additions to make to the final paper and will potentially even motivate future work in other related problem domains.

Reviewer RtuW2023-08-15

Thanks for the response. I have read the other reviews and all author rebuttals. As the majority of reviewers are in agreement, I will be retaining my initial rating.

Reviewer B3F82023-08-18

Thank you for the clarification.

Thank you for clarifying the questions. 1. Regarding the rand_nn experiment in Figure 3, how is nn_rand feat obtained? Is nn_rand feat output of an untrained neural network with location input? (or env input or env+location?) 2. [B3F8-3] The computational trade-off between parameter averaging and output averaging comes from the size of observed samples versus (cost for parameter averaging) the size of the unexplored area, i.e., the number of location samples to be predicted for species existence (cost for output averaging). A brief discussion of the computational cost would enhance the readability of the paper. Please correct me if there is any misunderstanding on my side. 3. [B3F8-5] Thank you for the clarification but I am still unclear how the classifier trained from scratch performs. As I understand, the feature extractor is frozen, so classifiers are only updated. Three parts are introduced in this paper, 1) linear combination of hypothesis models using equation (3), 2) Active sampling using equation (5), 3) online training with an updated set of annotated samples S^t. I wonder how the model would perform with only 2) and 3) without 1). In other words, how would a new species classifier perform when trained from scratch but using active sampling and online training? This is not a critical question, but I would like to know the importance of training from linearly combined weights instead of starting from scratch.

Authorsrebuttal2023-08-20

response to reviewer

Thank you for your responses and for your engagement with our paper. **Q1 - [B3F8-6]** Your first explanation is correct. “nn rand feats” are the output of an untrained neural network that takes *location* as input. We will clarify this in the paper. **Q2 - [B3F8-3]** We agree with your comments and interpretation of the computational cost. We will add a discussion of this to the final paper so that readers can better understand the merits of our approach, e.g. the benefits of the compact representation stemming from eqn 3, as noted in our previous response. Thanks for the suggestion. **Q3 - [B3F8-5]** To clarify, the feature extractor is frozen in all experiments (with the exception of the AN_FULL_E2E baseline). This is motivated by two primary reasons: (i) the amount of labeled data that is obtained during the early stages of active learning is very small, i.e. just a few samples, whereas the backbone neural network has multiple layers and (ii) it is desirable for the model update step to be efficient so that it does not become a bottleneck between the active sampling steps. Point (i) is particularly important, as it would not be trivial to select the appropriate training hyperparameters for each time step when training end-to-end with limited data. On line 258 we describe a baseline that we believe is close to what you request, i.e. the AN_FULL_E2E backbone is trained in an end-to-end manner with all the labeled data. As can be observed in Fig 2, we outperform this strong baseline for the IUCN dataset and are approaching it in the case of S&T. This gives a sense of the impact of training end-to-end versus using a linear combination of weights. Furthermore, Fig 3 (top right) illustrates the performance decrease of our method when the linear combination is not a good approximation for a test species’ range owning to a reduced candidate set size. Also related is LR_HSS in Fig 2 which utilizes eqn 5 for active sampling and has online training (of just the classifier) using S^t, but does not generate a model using eqn 3. In this case sampling locations are determined by the weighted hypothesis set, but unlike WA_HSS, the classifier is trained on S^t using standard logistic regression (i.e. from scratch). This model performs poorly on both the IUCN and S&T datasets. We agree that yours is an interesting question, and we will further address it in the final text. Also, please do not hesitate to follow up with any additional questions if we have not fully answered your questions or if we have misinterpreted them.

Reviewer yFB12023-08-18

Follow-up

Thanks for the detailed clarification! Upon reading the rebuttal and other reviewers' comments, I'll raise my rating from borderline accept to weak accept.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC