Publications by Benjamin Alexander Kellenberger

Fine-grained population mapping from coarse census counts and open geodata

Muhammad Imran, Benjamin Alexander Kellenberger, Devis Tuia

Fine-grained population maps are needed in several domains, like urban planning, environmental monitoring, public health, and humanitarian operations. Unfortunately, in many countries only aggregate census counts over large spatial units are collected, moreover, these are not always up-to-date. We present POMELO, a deep learning model that employs coarse census counts and open geodata to estimate fine-grained population maps with 100m ground sampling distance. Moreover, the model can also estimate population numbers when no census counts at all are available, by generalizing across countries. In a series of experiments for several countries in sub-Saharan Africa, the maps produced with POMELO are in good agreement with the most detailed available reference counts: disaggregation of coarse census counts reaches R2 values of 85–89%; unconstrained prediction in the absence of any counts reaches 48–69%.

2022

Mapping forest in the Swiss Alps treeline ecotone with explainable deep learning

Benjamin Alexander Kellenberger, Thiên-Anh Claris Nguyen, Devis Tuia

Forest maps are essential to understand forest dynamics. Due to the increasing availability of remote sensing data and machine learning models like convolutional neural networks, forest maps can these days be created on large scales with high accuracy. Common methods usually predict a map from remote sensing images without deliberately considering intermediate semantic concepts that are relevant to the final map. This makes the mapping process difficult to interpret, especially when using opaque deep learning models. Moreover, such procedure is entirely agnostic to the definitions of the mapping targets (e.g., forest types depending on variables such as tree height and tree density). Common models can at best learn these rules implicitly from data, which greatly hinders trust in the produced maps. In this work, we aim at building an explainable deep learning model for forest mapping that leverages prior knowledge about forest definitions to provide explanations to its decisions. We propose a model that explicitly quantifies intermediate variables like tree height and tree canopy density involved in the forest definitions, corresponding to those used to create the forest maps for training the model in the first place, and combines them accordingly. We apply our model to mapping forest types using very high resolution aerial imagery and lay particular focus on the treeline ecotone at high altitudes, where forest boundaries are complex and highly dependent on the chosen forest definition. Results show that our rule-informed model is able to quantify intermediate key variables and predict forest maps that reflect forest definitions. Through its interpretable design, it is further able to reveal implicit patterns in the manually-annotated forest labels, which facilitates the analysis of the produced maps and their comparison with other datasets.

Elsevier2022

, , ,

Inexpensive and accessible sensors are accelerating data acquisition in animal ecology. These technologies hold great potential for large-scale ecological understanding, but are limited by current processing approaches which inefficiently distill data into relevant information. We argue that animal ecologists can capitalize on large datasets generated by modern sensors by combining machine learning approaches with domain knowledge. Incorporating machine learning into ecological workflows could improve inputs for ecological models and lead to integrated hybrid modeling tools. This approach will require close interdisciplinary collaboration to ensure the quality of novel approaches and train a new generation of data scientists in ecology and conservation.

2022

Training Techniques for Presence-Only Habitat Suitability Mapping with Deep Learning

Benjamin Alexander Kellenberger, Devis Tuia

The goal of habitat suitability mapping is to predict the lo-cations in which a given species could be present. This is typically accomplished by statistical models which use envi-ronmental variables to predict species observation data. The relationship between the environmental characteristics of a location and the species that live there is likely to be quite complex, so deep learning models would seem natural to use. In practice, there are biases in the training data which present obstacles to standard deep learning approaches. First, large-scale species observation collections typically consist of presence-only data, which means we only have locations where a species has been observed (not where it has been confirmed to be absent). Second, the class distribution tends to be long-tailed. In this work we examine training tech-niques to mitigate these challenges: (i) a method for sharing species information between nearby observations and (ii) a curriculum learning strategy to reduce class imbalance early in training. These methods enable us to outperform state-of-the-art results on the GeoLifeCLEF 2020 dataset and suggest fruitful directions for future work.

IEEE2022

21 000 birds in 4.5 h: efficient large‐scale seabird detection with machine learning

Benjamin Alexander Kellenberger, Devis Tuia

We address the task of automatically detecting and counting seabirds in unmanned aerial vehicle (UAV) imagery using deep convolutional neural networks (CNNs). Our study area, the coast of West Africa, harbours significant breeding colonies of terns and gulls, which as top predators in the food web function as important bioindicators for the health of the marine ecosystem. Surveys to estimate breeding numbers have hitherto been carried out on foot, which is tedious, imprecise and causes disturbance. By using UAVs and CNNs that allow localizing tens of thousands of birds automatically, we show that all three limitations can be addressed elegantly. As we employ a lightweight CNN architecture and incorporate prior knowledge about the spatial distribution of birds within the colonies, we were able to reduce the number of bird annotations required for CNN training to just 200 examples per class. Our model obtains good accuracy for the most abundant species of royal terns (90% precision at 90% recall), but is less accurate for the rarer Caspian terns and gull species (60% precision at 68% recall, respectively 20% precision at 88% recall), which amounts to around 7% of all individuals present. In sum, our results show that we can detect and classify the majority of 21 000 birds in just 4.5 h, start to finish, as opposed to about 3 weeks of tediously identifying and labelling all birds by hand.

2021

Counting using deep learning regression gives value to ecological surveys

Benjamin Alexander Kellenberger, Devis Tuia

Many ecological studies rely on count data and involve manual counting of objects of interest, which is time-consuming and especially disadvantageous when time in the field or lab is limited. However, an increasing number of works uses digital imagery, which opens opportunities to automatise counting tasks. In this study, we use machine learning to automate counting objects of interest without the need to label individual objects. By leveraging already existing image-level annotations, this approach can also give value to historical data that were collected and annotated over longer time series (typical for many ecological studies), without the aim of deep learning applications. We demonstrate deep learning regression on two fundamentally different counting tasks: (i) daily growth rings from microscopic images of fish otolith (i.e., hearing stone) and (ii) hauled out seals from highly variable aerial imagery. In the otolith images, our deep learning-based regressor yields an RMSE of 3.40 day-rings and an R2 of 0.92. Initial performance in the seal images is lower (RMSE of 23.46 seals and R2 of 0.72), which can be attributed to a lack of images with a high number of seals in the initial training set, compared to the test set. We then show how to improve performance substantially (RMSE of 19.03 seals and R2 of 0.77) by carefully selecting and relabelling just 100 additional training images based on initial model prediction discrepancy. The regression-based approach used here returns accurate counts (R2 of 0.92 and 0.77 for the rings and seals, respectively), directly usable in ecological research.

2021

, ,

When applied to new datasets, acquired at different time moments, with different sensors or under different acquisition conditions, deep learning models might fail spectacularly. This is because they have learned from the data distribution observed during training and, as such, do not generalize out of that domain naturally. This chapter introduces methodologies designed to tackle this problem and provide deep learning models able to adapt to new data distributions, i.e. domain adaptation. Domain adaptation works by either adapting the representation to the new data distribution, modifying the inputs or performing smart sampling. But independently of the strategy, they lead to updated models, able to process effectively the new data without needing observation from it (or a very limited amount).

Wiley2021

How to find a good image- text embedding for remote sensing visual question answering?

Christel Marie Chappuis, Benjamin Alexander Kellenberger, Sylvain Lobry, Devis Tuia

2021

Mapping vulnerable populations with AI

Muhammad Imran, Benjamin Alexander Kellenberger, Devis Tuia, John Edgar Vargas Munos

2021

Self-Supervised Pretraining and Controlled Augmentation Improve Rare Wildlife Recognition in UAV Images

Rui Gong, Benjamin Alexander Kellenberger, Devis Tuia, Xiaochen Zheng

Automated animal censuses with aerial imagery are a vital ingredient towards wildlife conservation. Recent models are generally based on deep learning and thus require vast amounts of training data. Due to their scarcity and minuscule size, annotating animals in aerial imagery is a highly tedious process. In this project, we present a methodology to reduce the amount of required training data by resorting to self-supervised pretraining. In detail, we examine a combination of recent contrastive learning methodologies like Momentum Contrast (MoCo) and Cross-Level Instance-Group Discrimination (CLD) to condition our model on the aerial images without the requirement for labels. We show that a combination of MoCo, CLD, and geometric augmentations outperforms conventional models pretrained on ImageNet by a large margin. Crucially, our method still yields favorable results even if we reduce the number of training animals to just 10%, at which point our best model scores double the recall of the baseline at similar precision. This effectively allows reducing the number of required annotations to a fraction while still being able to train highaccuracy models in such highly challenging settings.

IEEE2021