Jan Dirk Wegner

Nico Lang, Walter Jetz, Konrad Schindler et al.

The worldwide variation in vegetation height is fundamental to the global carbon cycle and central to the functioning of ecosystems and their biodiversity. Geospatially explicit and, ideally, highly resolved information is required to manage terrestrial ecosystems, mitigate climate change, and prevent biodiversity loss. Here, we present the first global, wall-to-wall canopy height map at 10 m ground sampling distance for the year 2020. No single data source meets these requirements: dedicated space missions like GEDI deliver sparse height data, with unprecedented coverage, whereas optical satellite images like Sentinel-2 offer dense observations globally, but cannot directly measure vertical structures. By fusing GEDI with Sentinel-2, we have developed a probabilistic deep learning model to retrieve canopy height from Sentinel-2 images anywhere on Earth, and to quantify the uncertainty in these estimates. The presented approach reduces the saturation effect commonly encountered when estimating canopy height from satellite images, allowing to resolve tall canopies with likely high carbon stocks. According to our map, only 5% of the global landmass is covered by trees taller than 30 m. Such data play an important role for conservation, e.g., we find that only 34% of these tall canopies are located within protected areas. Our model enables consistent, uncertainty-informed worldwide mapping and supports an ongoing monitoring to detect change and inform decision making. The approach can serve ongoing efforts in forest conservation, and has the potential to foster advances in climate, carbon, and biodiversity modelling.

Mehmet Ozgur Turkoglu, Alexander Becker, Hüseyin Anil Gündüz et al.

The ability to estimate epistemic uncertainty is often crucial when deploying machine learning in the real world, but modern methods often produce overconfident, uncalibrated uncertainty predictions. A common approach to quantify epistemic uncertainty, usable across a wide class of prediction models, is to train a model ensemble. In a naive implementation, the ensemble approach has high computational cost and high memory demand. This challenges in particular modern deep learning, where even a single deep network is already demanding in terms of compute and memory, and has given rise to a number of attempts to emulate the model ensemble without actually instantiating separate ensemble members. We introduce FiLM-Ensemble, a deep, implicit ensemble method based on the concept of Feature-wise Linear Modulation (FiLM). That technique was originally developed for multi-task learning, with the aim of decoupling different tasks. We show that the idea can be extended to uncertainty quantification: by modulating the network activations of a single deep network with FiLM, one obtains a model ensemble with high diversity, and consequently well-calibrated estimates of epistemic uncertainty, with low computational overhead in comparison. Empirically, FiLM-Ensemble outperforms other implicit ensemble methods, and it and comes very close to the upper bound of an explicit ensemble of networks (sometimes even beating it), at a fraction of the memory cost.

Yuanwen Yue, Damien Robert, Jianyuan Wang et al.

Modern neural architectures for 3D point cloud processing contain both convolutional layers and attention blocks, but the best way to assemble them remains unclear. We analyse the role of different computational blocks in 3D point cloud networks and find an intuitive behaviour: convolution is adequate to extract low-level geometry at high-resolution in early layers, where attention is expensive without bringing any benefits; attention captures high-level semantics and context in low-resolution, deep layers more efficiently. Guided by this design principle, we propose a new, improved 3D point cloud backbone that employs convolutions in early stages and switches to attention for deeper layers. To avoid the loss of spatial layout information when discarding redundant convolution layers, we introduce a novel, training-free 3D positional encoding, PointROPE. The resulting LitePT model has $3.6\times$ fewer parameters, runs $2\times$ faster, and uses $2\times$ less memory than the state-of-the-art Point Transformer V3, but nonetheless matches or even outperforms it on a range of tasks and datasets. Code and models are available at: https://github.com/prs-eth/LitePT.

Patrick Ebel, Vivien Sainte Fare Garnot, Michael Schmitt et al.

Clouds and haze often occlude optical satellite images, hindering continuous, dense monitoring of the Earth's surface. Although modern deep learning methods can implicitly learn to ignore such occlusions, explicit cloud removal as pre-processing enables manual interpretation and allows training models when only few annotations are available. Cloud removal is challenging due to the wide range of occlusion scenarios -- from scenes partially visible through haze, to completely opaque cloud coverage. Furthermore, integrating reconstructed images in downstream applications would greatly benefit from trustworthy quality assessment. In this paper, we introduce UnCRtainTS, a method for multi-temporal cloud removal combining a novel attention-based architecture, and a formulation for multivariate uncertainty prediction. These two components combined set a new state-of-the-art performance in terms of image reconstruction on two public cloud removal datasets. Additionally, we show how the well-calibrated predicted uncertainties enable a precise control of the reconstruction quality.

Ayushi Sharma, Clemens Mosig, Lukas Drees et al.

Forests worldwide are increasingly threatened by climate change and disturbances such as fire, pests, and pathogens, creating an urgent need for scalable monitoring of tree cover and tree mortality. Aerial imagery from drones and aircraft is a key data source for detailed and large-scale mapping of tree crowns and mortality. However, related progress is limited by the lack of globally representative, harmonized datasets for joint segmentation of tree cover and mortality. We introduce two novel, open, machine-learning-ready datasets to enable joint segmentation of tree cover and tree mortality from centimeter-scale aerial imagery for the first time at global scales. With DTE-aerial-train, we provide a training dataset comprising 385K image patches of size 1024x1024 pixels, with resolutions ranging from 2.5 to 20 cm. It includes multi-class expert-annotated and -audited pseudo-labels for tree cover and mortality. With DTE-aerial-bench, we provide a geographically balanced benchmark test set of 25 globally distributed orthoimages totaling 525 patches with high-quality expert annotations for both tree cover and mortality. Both the training and benchmark datasets span tropical, temperate, boreal, and dryland biomes and cover a wide range of forest structures and mortality patterns. Using the benchmark test set for evaluation, we establish strong reference baselines that improve mortality segmentation across all biomes and scales with significant gains in challenging regions, such as boreal forests, where the F1 score increases from 0.40 to 0.58 with around 45% relative improvement. All data, models, and code will be publicly released under permissive open-source licenses. An interactive visualization of the benchmark dataset is available at deadtrees.earth/releases/dte-aerial-bench.

Alexander Becker, Rodrigo Caye Daudt, Dominik Narnhofer et al.

Recent approaches to arbitrary-scale single image super-resolution (ASR) use neural fields to represent continuous signals that can be sampled at arbitrary resolutions. However, point-wise queries of neural fields do not naturally match the point spread function (PSF) of pixels, which may cause aliasing in the super-resolved image. Existing methods attempt to mitigate this by approximating an integral version of the field at each scaling factor, compromising both fidelity and generalization. In this work, we introduce neural heat fields, a novel neural field formulation that inherently models a physically exact PSF. Our formulation enables analytically correct anti-aliasing at any desired output resolution, and -- unlike supersampling -- at no additional cost. Building on this foundation, we propose Thera, an end-to-end ASR method that substantially outperforms existing approaches, while being more parameter-efficient and offering strong theoretical guarantees. The project page is at https://therasr.github.io.

Yuchang Jiang, Anton Raichuk, Xiaoye Tong et al.

Monitoring tree crop expansion is vital for zero-deforestation policies like the European Union's Regulation on Deforestation-free Products (EUDR). However, these efforts are hindered by a lack of highresolution data distinguishing diverse agricultural systems from forests. Here, we present the first 10m-resolution tree crop map for South America, generated using a multi-modal, spatio-temporal deep learning model trained on Sentinel-1 and Sentinel-2 satellite imagery time series. The map identifies approximately 11 million hectares of tree crops, 23% of which is linked to 2000-2020 forest cover loss. Critically, our analysis reveals that existing regulatory maps supporting the EUDR often classify established agriculture, particularly smallholder agroforestry, as "forest". This discrepancy risks false deforestation alerts and unfair penalties for small-scale farmers. Our work mitigates this risk by providing a high-resolution baseline, supporting conservation policies that are effective, inclusive, and equitable.

Olivier Dietrich, Merlin Alfredsson, Emilia Arens et al.

Natural disasters demand rapid damage assessment to guide humanitarian response. Here, we investigate whether medium-resolution Earth observation images from the Copernicus program can support building damage assessment, complementing very-high resolution imagery with often limited availability. We introduce xBD-S12, a dataset of 10,315 pre- and post-disaster image pairs from both Sentinel-1 and Sentinel-2, spatially and temporally aligned with the established xBD benchmark. In a series of experiments, we demonstrate that building damage can be detected and mapped rather well in many disaster scenarios, despite the moderate 10$\,$m ground sampling distance. We also find that, for damage mapping at that resolution, architectural sophistication does not seem to bring much advantage: more complex model architectures tend to struggle with generalization to unseen disasters, and geospatial foundation models bring little practical benefit. Our results suggest that Copernicus images are a viable data source for rapid, wide-area damage assessment and could play an important role alongside VHR imagery. We release the xBD-S12 dataset, code, and trained models to support further research.

Stefania Russo, Nathanaël Perraudin, Steven Stalder et al.

In this technical report we compare different deep learning models for prediction of water depth rasters at high spatial resolution. Efficient, accurate, and fast methods for water depth prediction are nowadays important as urban floods are increasing due to higher rainfall intensity caused by climate change, expansion of cities and changes in land use. While hydrodynamic models models can provide reliable forecasts by simulating water depth at every location of a catchment, they also have a high computational burden which jeopardizes their application to real-time prediction in large urban areas at high spatial resolution. Here, we propose to address this issue by using data-driven techniques. Specifically, we evaluate deep learning models which are trained to reproduce the data simulated by the CADDIES cellular-automata flood model, providing flood forecasts that can occur at different future time horizons. The advantage of using such models is that they can learn the underlying physical phenomena a priori, preventing manual parameter setting and computational burden. We perform experiments on a dataset consisting of two catchments areas within Switzerland with 18 simpler, short rainfall patterns and 4 long, more complex ones. Our results show that the deep learning models present in general lower errors compared to the other methods, especially for water depths $>0.5m$. However, when testing on more complex rainfall events or unseen catchment areas, the deep models do not show benefits over the simpler ones.

Nando Metzger, Mehmet Özgür Türkoglu, Rodrigo Caye Daudt et al.

Forecasting where and when new buildings will emerge is a rather unexplored topic, but one that is very useful in many disciplines such as urban planning, agriculture, resource management, and even autonomous flying. In the present work, we present a method that accomplishes this task with a deep neural network and a custom pretraining procedure. In Stage 1, a U-Net backbone is pretrained within a Siamese network architecture that aims to solve a (building) change detection task. In Stage 2, the backbone is repurposed to forecast the emergence of new buildings based solely on one image acquired before its construction. Furthermore, we also present a model that forecasts the time range within which the change will occur. We validate our approach using the SpaceNet7 dataset, which covers an area of 960 km^2 at 24 points in time across two years. In our experiments, we found that our proposed pretraining method consistently outperforms the traditional pretraining using the ImageNet dataset. We also show that it is to some degree possible to predict in advance when building changes will occur.

Kaan Karaman, Yuchang Jiang, Damien Robert et al.

Accurate Above-Ground Biomass (AGB) mapping at both large scale and high spatio-temporal resolution is essential for applications ranging from climate modeling to biodiversity assessment, and sustainable supply chain monitoring. At present, fine-grained AGB mapping relies on costly airborne laser scanning acquisition campaigns usually limited to regional scales. Initiatives such as the ESA CCI map attempt to generate global biomass products from diverse spaceborne sensors but at a coarser resolution. To enable global, high-resolution (HR) mapping, several works propose to regress AGB from HR satellite observations such as ESA Sentinel-1/2 images. We propose a novel way to address HR AGB estimation, by leveraging both HR satellite observations and existing low-resolution (LR) biomass products. We cast this problem as Guided Super-Resolution (GSR), aiming at upsampling LR biomass maps (sources) from $100$ to $10$ m resolution, using auxiliary HR co-registered satellite images (guides). We compare super-resolving AGB maps with and without guidance, against direct regression from satellite images, on the public BioMassters dataset. We observe that Multi-Scale Guidance (MSG) outperforms direct regression both for regression ($-780$ t/ha RMSE) and perception ($+2.0$ dB PSNR) metrics, and better captures high-biomass values, without significant computational overhead. Interestingly, unlike the RGB+Depth setting they were originally designed for, our best-performing AGB GSR approaches are those that most preserve the guide image texture. Our results make a strong case for adopting the GSR framework for accurate HR biomass mapping at scale. Our code and model weights are made publicly available (https://github.com/kaankaramanofficial/GSR4B).

Elena Plekhanova, Damien Robert, Johannes Dollinger et al.

With the exacerbation of the biodiversity and climate crises, macroecological pursuits such as global biodiversity mapping become more urgent. Remote sensing offers a wealth of Earth observation data for ecological studies, but the scarcity of labeled datasets remains a major challenge. Recently, self-supervised learning has enabled learning representations from unlabeled data, triggering the development of pretrained geospatial models with generalizable features. However, these models are often trained on datasets biased toward areas of high human activity, leaving entire ecological regions underrepresented. Additionally, while some datasets attempt to address seasonality through multi-date imagery, they typically follow calendar seasons rather than local phenological cycles. To better capture vegetation seasonality at a global scale, we propose a simple phenology-informed sampling strategy and introduce corresponding SSL4Eco, a multi-date Sentinel-2 dataset, on which we train an existing model with a season-contrastive objective. We compare representations learned from SSL4Eco against other datasets on diverse ecological downstream tasks and demonstrate that our straightforward sampling method consistently improves representation quality, highlighting the importance of dataset construction. The model pretrained on SSL4Eco reaches state of the art performance on 7 out of 8 downstream tasks spanning (multi-label) classification and regression. We release our code, data, and model weights to support macroecological and computer vision research at https://github.com/PlekhanovaElena/ssl4eco.

Yuchang Jiang, Vivien Sainte Fare Garnot, Konrad Schindler et al.

Data imbalance is ubiquitous when applying machine learning to real-world problems, particularly regression problems. If training data are imbalanced, the learning is dominated by the densely covered regions of the target distribution and the learned regressor tends to exhibit poor performance in sparsely covered regions. Beyond standard measures like oversampling or reweighting, there are two main approaches to handling learning from imbalanced data. For regression, recent work leverages the continuity of the distribution, while for classification, the trend has been to use ensemble methods, allowing some members to specialize in predictions for sparser regions. In our method, named UVOTE, we integrate recent advances in probabilistic deep learning with an ensemble approach for imbalanced regression. We replace traditional regression losses with negative log-likelihood, which also predicts sample-wise aleatoric uncertainty. Our experiments show that this loss function handles imbalance better. Additionally, we use the predicted aleatoric uncertainty values to fuse the predictions of different expert models in the ensemble, eliminating the need for a separate aggregation module. We compare our method with existing alternatives on multiple public benchmarks and show that UVOTE consistently outperforms the prior art, while at the same time producing better-calibrated uncertainty estimates. Our code is available at https://github.com/SherryJYC/UVOTE.

Yujia Liu, Anton Obukhov, Jan Dirk Wegner et al.

Computer-Aided Design (CAD) model reconstruction from point clouds is an important problem at the intersection of computer vision, graphics, and machine learning; it saves the designer significant time when iterating on in-the-wild objects. Recent advancements in this direction achieve relatively reliable semantic segmentation but still struggle to produce an adequate topology of the CAD model. In this work, we analyze the current state of the art for that ill-posed task and identify shortcomings of existing methods. We propose a hybrid analytic-neural reconstruction scheme that bridges the gap between segmented point clouds and structured CAD models and can be readily combined with different segmentation backbones. Moreover, to power the surface fitting stage, we propose a novel implicit neural representation of freeform surfaces, driving up the performance of our overall CAD reconstruction scheme. We extensively evaluate our method on the popular ABC benchmark of CAD models and set a new state-of-the-art for that dataset. Project page: https://www.obukhov.ai/point2cad}{https://www.obukhov.ai/point2cad.

Claudio Persello, Jan Dirk Wegner, Ronny Hänsch et al. · mila

The synergistic combination of deep learning models and Earth observation promises significant advances to support the sustainable development goals (SDGs). New developments and a plethora of applications are already changing the way humanity will face the living planet challenges. This paper reviews current deep learning approaches for Earth observation data, along with their application towards monitoring and achieving the SDGs most impacted by the rapid development of deep learning in Earth observation. We systematically review case studies to 1) achieve zero hunger, 2) sustainable cities, 3) deliver tenure security, 4) mitigate and adapt to climate change, and 5) preserve biodiversity. Important societal, economic and environmental implications are concerned. Exciting times ahead are coming where algorithms and Earth data can help in our endeavor to address the climate crisis and support more sustainable development.

Yujia Liu, Anton Obukhov, Jan Dirk Wegner et al.

We present a learning-based approach to reconstruct buildings as 3D polygonal meshes from airborne LiDAR point clouds. What makes 3D building reconstruction from airborne LiDAR hard is the large diversity of building designs and especially roof shapes, the low and varying point density across the scene, and the often incomplete coverage of building facades due to occlusions by vegetation or to the viewing angle of the sensor. To cope with the diversity of shapes and inhomogeneous and incomplete object coverage, we introduce a generative model that directly predicts 3D polygonal meshes from input point clouds. Our autoregressive model, called Point2Building, iteratively builds up the mesh by generating sequences of vertices and faces. This approach enables our model to adapt flexibly to diverse geometries and building structures. Unlike many existing methods that rely heavily on pre-processing steps like exhaustive plane detection, our model learns directly from the point cloud data, thereby reducing error propagation and increasing the fidelity of the reconstruction. We experimentally validate our method on a collection of airborne LiDAR data of Zurich, Berlin and Tallinn. Our method shows good generalization to diverse urban styles.

Carlos Gomes, Isabelle Wittmann, Damien Robert et al.

Over the past decades, there has been an explosion in the amount of available Earth Observation (EO) data. The unprecedented coverage of the Earth's surface and atmosphere by satellite imagery has resulted in large volumes of data that must be transmitted to ground stations, stored in data centers, and distributed to end users. Modern Earth System Models (ESMs) face similar challenges, operating at high spatial and temporal resolutions, producing petabytes of data per simulated day. Data compression has gained relevance over the past decade, with neural compression (NC) emerging from deep learning and information theory, making EO data and ESM outputs ideal candidates due to their abundance of unlabeled data. In this review, we outline recent developments in NC applied to geospatial data. We introduce the fundamental concepts of NC including seminal works in its traditional applications to image and video compression domains with focus on lossy compression. We discuss the unique characteristics of EO and ESM data, contrasting them with "natural images", and explain the additional challenges and opportunities they present. Moreover, we review current applications of NC across various EO modalities and explore the limited efforts in ESM compression to date. The advent of self-supervised learning (SSL) and foundation models (FM) has advanced methods to efficiently distill representations from vast unlabeled data. We connect these developments to NC for EO, highlighting the similarities between the two fields and elaborate on the potential of transferring compressed feature representations for machine--to--machine communication. Based on insights drawn from this review, we devise future directions relevant to applications in EO and ESM.

Johannes Dollinger, Damien Robert, Elena Plekhanova et al.

Deep learning on climatic data holds potential for macroecological applications. However, its adoption remains limited among scientists outside the deep learning community due to storage, compute, and technical expertise barriers. To address this, we introduce Climplicit, a spatio-temporal geolocation encoder pretrained to generate implicit climatic representations anywhere on Earth. By bypassing the need to download raw climatic rasters and train feature extractors, our model uses x3500 less disk space and significantly reduces computational needs for downstream tasks. We evaluate our Climplicit embeddings on biomes classification, species distribution modeling, and plant trait regression. We find that single-layer probing our Climplicit embeddings consistently performs better or on par with training a model from scratch on downstream tasks and overall better than alternative geolocation encoding models.

Vivien Sainte Fare Garnot, Lynsay Spafford, Jelle Lever et al.

Phenology, the timing of cyclical plant life events such as leaf emergence and coloration, is crucial in the bio-climatic system. Climate change drives shifts in these phenological events, impacting ecosystems and the climate itself. Accurate phenology models are essential to predict the occurrence of these phases under changing climatic conditions. Existing methods include hypothesis-driven process models and data-driven statistical approaches. Process models account for dormancy stages and various phenology drivers, while statistical models typically rely on linear or traditional machine learning techniques. Research shows that process models often outperform statistical methods when predicting under climate conditions outside historical ranges, especially with climate change scenarios. However, deep learning approaches remain underexplored in climate phenology modeling. We introduce PhenoFormer, a neural architecture better suited than traditional statistical methods at predicting phenology under shift in climate data distribution, while also bringing significant improvements or performing on par to the best performing process-based models. Our numerical experiments on a 70-year dataset of 70,000 phenological observations from 9 woody species in Switzerland show that PhenoFormer outperforms traditional machine learning methods by an average of 13% R2 and 1.1 days RMSE for spring phenology, and 11% R2 and 0.7 days RMSE for autumn phenology, while matching or exceeding the best process-based models. Our results demonstrate that deep learning has the potential to be a valuable methodological tool for accurate climate-phenology prediction, and our PhenoFormer is a first promising step in improving phenological predictions before a complete understanding of the underlying physiological mechanisms is available.

Yuchang Jiang, Jan Dirk Wegner, Vivien Sainte Fare Garnot

Plant phenology modelling aims to predict the timing of seasonal phases, such as leaf-out or flowering, from meteorological time series. Reliable predictions are crucial for anticipating ecosystem responses to climate change. While phenology modelling has traditionally relied on mechanistic approaches, deep learning methods have recently been proposed as flexible, data-driven alternatives with often superior performance. However, mechanistic models tend to outperform deep networks when data distribution shifts are induced by climate change. Domain Adaptation (DA) techniques could help address this limitation. Yet, unlike standard DA settings, climate change induces a temporal continuum of domains and involves both a covariate and label shift, with warmer records and earlier start of spring. To tackle this challenge, we introduce Mid-feature Rank-adversarial Domain Adaptation (MIRANDA). Whereas conventional adversarial methods enforce domain invariance on final latent representations, an approach that does not explicitly address label shift, we apply adversarial regularization to intermediate features. Moreover, instead of a binary domain-classification objective, we employ a rank-based objective that enforces year-invariance in the learned meteorological representations. On a country-scale dataset spanning 70 years and comprising 67,800 phenological observations of 5 tree species, we demonstrate that, unlike conventional DA approaches, MIRANDA improves robustness to climatic distribution shifts and narrows the performance gap with mechanistic models.

Olivier Dietrich, Torben Peters, Vivien Sainte Fare Garnot et al.

Access to detailed war impact assessments is crucial for humanitarian organizations to assist affected populations effectively. However, maintaining a comprehensive understanding of the situation on the ground is challenging, especially in widespread and prolonged conflicts. Here we present a scalable method for estimating building damage resulting from armed conflicts. By training a machine learning model on Synthetic Aperture Radar image time series, we generate probabilistic damage estimates at the building level, leveraging existing damage assessments and open building footprints. To allow large-scale inference and ensure accessibility, we tie our method to run on Google Earth Engine. Users can adjust confidence intervals to suit their needs, enabling rapid and flexible assessments of war-related damage across large areas. We provide two publicly accessible dashboards: a Ukraine Damage Explorer to dynamically view our precomputed estimates, and a Rapid Damage Mapping Tool to run our method and generate custom maps.

Stefano D'Aronco, Giorgio Trumpy, David Pfluger et al.

We propose the first accurate digitization and color reconstruction process for historical lenticular film that is robust to artifacts. Lenticular films emerged in the 1920s and were one of the first technologies that permitted to capture full color information in motion. The technology leverages an RGB filter and cylindrical lenticules embossed on the film surface to encode the color in the horizontal spatial dimension of the image. To project the pictures the encoding process was reversed using an appropriate analog device. In this work, we introduce an automated, fully digital pipeline to process the scan of lenticular films and colorize the image. Our method merges deep learning with a model-based approach in order to maximize the performance while making sure that the reconstructed colored images truthfully match the encoded color information. Our model employs different strategies to achieve an effective color reconstruction, in particular (i) we use data augmentation to create a robust lenticule segmentation network, (ii) we fit the lenticules raster prediction to obtain a precise vectorial lenticule localization, and (iii) we train a colorization network that predicts interpolation coefficients in order to obtain a truthful colorization. We validate the proposed method on a lenticular film dataset and compare it to other approaches. Since no colored groundtruth is available as reference, we conduct a user study to validate our method in a subjective manner. The results of the study show that the proposed method is largely preferred with respect to other existing and baseline methods.

Alexander Becker, Stefania Russo, Stefano Puliti et al.

Monitoring and managing Earth's forests in an informed manner is an important requirement for addressing challenges like biodiversity loss and climate change. While traditional in situ or aerial campaigns for forest assessments provide accurate data for analysis at regional level, scaling them to entire countries and beyond with high temporal resolution is hardly possible. In this work, we propose a method based on deep ensembles that densely estimates forest structure variables at country-scale with 10-meter resolution, using freely available satellite imagery as input. Our method jointly transforms Sentinel-2 optical images and Sentinel-1 synthetic-aperture radar images into maps of five different forest structure variables: 95th height percentile, mean height, density, Gini coefficient, and fractional cover. We train and test our model on reference data from 41 airborne laser scanning missions across Norway and demonstrate that it is able to generalize to unseen test regions, achieving normalized mean absolute errors between 11% and 15%, depending on the variable. Our work is also the first to propose a variant of so-called Bayesian deep learning to densely predict multiple forest structure variables with well-calibrated uncertainty estimates from satellite imagery. The uncertainty information increases the trustworthiness of the model and its suitability for downstream tasks that require reliable confidence estimates as a basis for decision making. We present an extensive set of experiments to validate the accuracy of the predicted maps as well as the quality of the predicted uncertainties. To demonstrate scalability, we provide Norway-wide maps for the five forest structure variables.

Nico Lang, Konrad Schindler, Jan Dirk Wegner

The increasing demand for commodities is leading to changes in land use worldwide. In the tropics, deforestation, which causes high carbon emissions and threatens biodiversity, is often linked to agricultural expansion. While the need for deforestation-free global supply chains is widely recognized, making progress in practice remains a challenge. Here, we propose an automated approach that aims to support conservation and sustainable land use planning decisions by mapping tropical landscapes at large scale and high spatial resolution following the High Carbon Stock (HCS) approach. A deep learning approach is developed that estimates canopy height for each 10 m Sentinel-2 pixel by learning from sparse GEDI LIDAR reference data, achieving an overall RMSE of 6.3 m. We show that these wall-to-wall maps of canopy top height are predictive for classifying HCS forests and degraded areas with an overall accuracy of 86 % and produce a first high carbon stock map for Indonesia, Malaysia, and the Philippines.

Devis Tuia, Ribana Roscher, Jan Dirk Wegner et al.

In the last years we have witnessed the fields of geosciences and remote sensing and artificial intelligence to become closer. Thanks to both the massive availability of observational data, improved simulations, and algorithmic advances, these disciplines have found common objectives and challenges to advance the modeling and understanding of the Earth system. Despite such great opportunities, we also observed a worrying tendency to remain in disciplinary comfort zones applying recent advances from artificial intelligence on well resolved remote sensing problems. Here we take a position on research directions where we think the interface between these fields will have the most impact and become potential game changers. In our declared agenda for AI on Earth sciences, we aim to inspire researchers, especially the younger generations, to tackle these challenges for a real advance of remote sensing and the geosciences.

Nikolai Kalischek, Jan Dirk Wegner, Konrad Schindler

Style transfer aims to render the content of a given image in the graphical/artistic style of another image. The fundamental concept underlying NeuralStyle Transfer (NST) is to interpret style as a distribution in the feature space of a Convolutional Neural Network, such that a desired style can be achieved by matching its feature distribution. We show that most current implementations of that concept have important theoretical and practical limitations, as they only partially align the feature distributions. We propose a novel approach that matches the distributions more precisely, thus reproducing the desired style more faithfully, while still being computationally efficient. Specifically, we adapt the dual form of Central Moment Discrepancy (CMD), as recently proposed for domain adaptation, to minimize the difference between the target style and the feature distribution of the output image. The dual interpretation of this metric explicitly matches all higher-order centralized moments and is therefore a natural extension of existing NST methods that only take into account the first and second moments. Our experiments confirm that the strong theoretical properties also translate to visually better style transfer, and better disentangle style from semantic image content.

Nico Lang, Nikolai Kalischek, John Armston et al.

NASA's Global Ecosystem Dynamics Investigation (GEDI) is a key climate mission whose goal is to advance our understanding of the role of forests in the global carbon cycle. While GEDI is the first space-based LIDAR explicitly optimized to measure vertical forest structure predictive of aboveground biomass, the accurate interpretation of this vast amount of waveform data across the broad range of observational and environmental conditions is challenging. Here, we present a novel supervised machine learning approach to interpret GEDI waveforms and regress canopy top height globally. We propose a probabilistic deep learning approach based on an ensemble of deep convolutional neural networks(CNN) to avoid the explicit modelling of unknown effects, such as atmospheric noise. The model learns to extract robust features that generalize to unseen geographical regions and, in addition, yields reliable estimates of predictive uncertainty. Ultimately, the global canopy top height estimates produced by our model have an expected RMSE of 2.7 m with low bias.

Yujia Liu, Stefano D'Aronco, Konrad Schindler et al.

We introduce PC2WF, the first end-to-end trainable deep network architecture to convert a 3D point cloud into a wireframe model. The network takes as input an unordered set of 3D points sampled from the surface of some object, and outputs a wireframe of that object, i.e., a sparse set of corner points linked by line segments. Recovering the wireframe is a challenging task, where the numbers of both vertices and edges are different for every instance, and a-priori unknown. Our architecture gradually builds up the model: It starts by encoding the points into feature vectors. Based on those features, it identifies a pool of candidate vertices, then prunes those candidates to a final set of corner vertices and refines their locations. Next, the corners are linked with an exhaustive set of candidate edges, which is again pruned to obtain the final wireframe. All steps are trainable, and errors can be backpropagated through the entire sequence. We validate the proposed model on a publicly available synthetic dataset, for which the ground truth wireframes are accessible, as well as on a new real-world dataset. Our model produces wireframe abstractions of good quality and outperforms several baselines.

Mehmet Ozgur Turkoglu, Stefano D'Aronco, Gregor Perich et al.

The aim of this paper is to map agricultural crops by classifying satellite image time series. Domain experts in agriculture work with crop type labels that are organised in a hierarchical tree structure, where coarse classes (like orchards) are subdivided into finer ones (like apples, pears, vines, etc.). We develop a crop classification method that exploits this expert knowledge and significantly improves the mapping of rare crop types. The three-level label hierarchy is encoded in a convolutional, recurrent neural network (convRNN), such that for each pixel the model predicts three labels at different level of granularity. This end-to-end trainable, hierarchical network architecture allows the model to learn joint feature representations of rare classes (e.g., apples, pears) at a coarser level (e.g., orchard), thereby boosting classification performance at the fine-grained level. Additionally, labelling at different granularity also makes it possible to adjust the output according to the classification scores; as coarser labels with high confidence are sometimes more useful for agricultural practice than fine-grained but very uncertain labels. We validate the proposed method on a new, large dataset that we make public. ZueriCrop covers an area of 50 km x 48 km in the Swiss cantons of Zurich and Thurgau with a total of 116'000 individual fields spanning 48 crop classes, and 28,000 (multi-temporal) image patches from Sentinel-2. We compare our proposed hierarchical convRNN model with several baselines, including methods designed for imbalanced class distributions. The hierarchical approach performs superior by at least 9.9 percentage points in F1-score.

Michael Kölle, Dominik Laupheimer, Stefan Schmohl et al.

Automated semantic segmentation and object detection are of great importance in geospatial data analysis. However, supervised machine learning systems such as convolutional neural networks require large corpora of annotated training data. Especially in the geospatial domain, such datasets are quite scarce. Within this paper, we aim to alleviate this issue by introducing a new annotated 3D dataset that is unique in three ways: i) The dataset consists of both an Unmanned Aerial Vehicle (UAV) laser scanning point cloud and a 3D textured mesh. ii) The point cloud features a mean point density of about 800 pts/sqm and the oblique imagery used for 3D mesh texturing realizes a ground sampling distance of about 2-3 cm. This enables the identification of fine-grained structures and represents the state of the art in UAV-based mapping. iii) Both data modalities will be published for a total of three epochs allowing applications such as change detection. The dataset depicts the village of Hessigheim (Germany), henceforth referred to as H3D. It is designed to promote research in the field of 3D data analysis on one hand and to evaluate and rank existing and emerging approaches for semantic segmentation of both data modalities on the other hand. Ultimately, we hope that H3D will become a widely used benchmark dataset in company with the well-established ISPRS Vaihingen 3D Semantic Labeling Challenge benchmark (V3D). The dataset can be downloaded from https://ifpwww.ifp.uni-stuttgart.de/benchmark/hessigheim/default.aspx.

Nando Metzger, Mehmet Ozgur Turkoglu, Stefano D'Aronco et al.

Optical satellite sensors cannot see the Earth's surface through clouds. Despite the periodic revisit cycle, image sequences acquired by Earth observation satellites are therefore irregularly sampled in time. State-of-the-art methods for crop classification (and other time series analysis tasks) rely on techniques that implicitly assume regular temporal spacing between observations, such as recurrent neural networks (RNNs). We propose to use neural ordinary differential equations (NODEs) in combination with RNNs to classify crop types in irregularly spaced image sequences. The resulting ODE-RNN models consist of two steps: an update step, where a recurrent unit assimilates new input data into the model's hidden state; and a prediction step, in which NODE propagates the hidden state until the next observation arrives. The prediction step is based on a continuous representation of the latent dynamics, which has several advantages. At the conceptual level, it is a more natural way to describe the mechanisms that govern the phenological cycle. From a practical point of view, it makes it possible to sample the system state at arbitrary points in time, such that one can integrate observations whenever they are available, and extrapolate beyond the last observation. Our experiments show that ODE-RNN indeed improves classification accuracy over common baselines such as LSTM, GRU, and temporal convolution. The gains are most prominent in the challenging scenario where only few observations are available (i.e., frequent cloud cover). Moreover, we show that the ability to extrapolate translates to better classification performance early in the season, which is important for forecasting.

Vít Růžička, Stefano D'Aronco, Jan Dirk Wegner et al.

We investigate active learning in the context of deep neural network models for change detection and map updating. Active learning is a natural choice for a number of remote sensing tasks, including the detection of local surface changes: changes are on the one hand rare and on the other hand their appearance is varied and diffuse, making it hard to collect a representative training set in advance. In the active learning setting, one starts from a minimal set of training examples and progressively chooses informative samples that are annotated by a user and added to the training set. Hence, a core component of an active learning system is a mechanism to estimate model uncertainty, which is then used to pick uncertain, informative samples. We study different mechanisms to capture and quantify this uncertainty when working with deep networks, based on the variance or entropy across explicit or implicit model ensembles. We show that active learning successfully finds highly informative samples and automatically balances the training distribution, and reaches the same performance as a model supervised with a large, pre-annotated training set, with $\approx$99% fewer annotated samples.

Andres C. Rodriguez, Stefano D'Aronco, Konrad Schindler et al.

We propose a scheme for supervised image classification that uses privileged information, in the form of keypoint annotations for the training data, to learn strong models from small and/or biased training sets. Our main motivation is the recognition of animal species for ecological applications such as biodiversity modelling, which is challenging because of long-tailed species distributions due to rare species, and strong dataset biases such as repetitive scene background in camera traps. To counteract these challenges, we propose a visual attention mechanism that is supervised via keypoint annotations that highlight important object parts. This privileged information, implemented as a novel privileged pooling operation, is only required during training and helps the model to focus on regions that are discriminative. In experiments with three different animal species datasets, we show that deep networks with privileged pooling can use small training sets more efficiently and generalize better.

Daniel Laumer, Nico Lang, Natalie van Doorn et al.

We introduce an approach for updating older tree inventories with geographic coordinates using street-level panorama images and a global optimization framework for tree instance matching. Geolocations of trees in inventories until the early 2000s where recorded using street addresses whereas newer inventories use GPS. Our method retrofits older inventories with geographic coordinates to allow connecting them with newer inventories to facilitate long-term studies on tree mortality etc. What makes this problem challenging is the different number of trees per street address, the heterogeneous appearance of different tree instances in the images, ambiguous tree positions if viewed from multiple images and occlusions. To solve this assignment problem, we (i) detect trees in Google street-view panoramas using deep learning, (ii) combine multi-view detections per tree into a single representation, (iii) and match detected trees with given trees per street address with a global optimization approach. Experiments for > 50000 trees in 5 cities in California, USA, show that we are able to assign geographic coordinates to 38 % of the street trees, which is a good starting point for long-term studies on the ecosystem services value of street trees at large scale.

Kishan Sharma, Moritz Gold, Christian Zurbruegg et al.

We propose to predict histograms of object sizes in crowded scenes directly without any explicit object instance segmentation. What makes this task challenging is the high density of objects (of the same category), which makes instance identification hard. Instead of explicitly segmenting object instances, we show that directly learning histograms of object sizes improves accuracy while using drastically less parameters. This is very useful for application scenarios where explicit, pixel-accurate instance segmentation is not needed, but there lies interest in the overall distribution of instance sizes. Our core applications are in biology, where we estimate the size distribution of soldier fly larvae, and medicine, where we estimate the size distribution of cancer cells as an intermediate step to calculate the tumor cellularity score. Given an image with hundreds of small object instances, we output the total count and the size histogram. We also provide a new data set for this task, the FlyLarvae data set, which consists of 11,000 larvae instances labeled pixel-wise. Our method results in an overall improvement in the count and size distribution prediction as compared to state-of-the-art instance segmentation method Mask R-CNN.

Mehmet Ozgur Turkoglu, Stefano D'Aronco, Jan Dirk Wegner et al.

We propose a new STAckable Recurrent cell (STAR) for recurrent neural networks (RNNs), which has fewer parameters than widely used LSTM and GRU while being more robust against vanishing or exploding gradients. Stacking recurrent units into deep architectures suffers from two major limitations: (i) many recurrent cells (e.g., LSTMs) are costly in terms of parameters and computation resources; and (ii) deep RNNs are prone to vanishing or exploding gradients during training. We investigate the training of multi-layer RNNs and examine the magnitude of the gradients as they propagate through the network in the "vertical" direction. We show that, depending on the structure of the basic recurrent unit, the gradients are systematically attenuated or amplified. Based on our analysis we design a new type of gated cell that better preserves gradient magnitude. We validate our design on a large number of sequence modelling tasks and demonstrate that the proposed STAR cell allows to build and train deeper recurrent architectures, ultimately leading to improved performance while being computationally more efficient.

Steve Branson, Jan Dirk Wegner, David Hall et al.

Up-to-date catalogs of the urban tree population are important for municipalities to monitor and improve quality of life in cities. Despite much research on automation of tree mapping, mainly relying on dedicated airborne LiDAR or hyperspectral campaigns, trees are still mostly mapped manually in practice. We present a fully automated tree detection and species recognition pipeline to process thousands of trees within a few hours using publicly available aerial and street view images of Google MapsTM. These data provide rich information (viewpoints, scales) from global tree shapes to bark textures. Our work-flow is built around a supervised classification that automatically learns the most discriminative features from thousands of trees and corresponding, public tree inventory data. In addition, we introduce a change tracker to keep urban tree inventories up-to-date. Changes of individual trees are recognized at city-scale by comparing street-level images of the same tree location at two different times. Drawing on recent advances in computer vision and machine learning, we apply convolutional neural networks (CNN) for all classification tasks. We propose the following pipeline: download all available panoramas and overhead images of an area of interest, detect trees per image and combine multi-view detections in a probabilistic framework, adding prior knowledge; recognize fine-grained species of detected trees. In a later, separate module, track trees over time and identify the type of change. We believe this is the first work to exploit publicly available image data for fine-grained tree mapping at city-scale, respectively over many thousands of trees. Experiments in the city of Pasadena, California, USA show that we can detect > 70% of the street trees, assign correct species to > 80% for 40 different species, and correctly detect and classify changes in > 90% of the cases.

Nico Lang, Konrad Schindler, Jan Dirk Wegner

Sentinel-2 multi-spectral images collected over periods of several months were used to estimate vegetation height for Gabon and Switzerland. A deep convolutional neural network (CNN) was trained to extract suitable spectral and textural features from reflectance images and to regress per-pixel vegetation height. In Gabon, reference heights for training and validation were derived from airborne LiDAR measurements. In Switzerland, reference heights were taken from an existing canopy height model derived via photogrammetric surface reconstruction. The resulting maps have a mean absolute error (MAE) of 1.7 m in Switzerland and 4.3 m in Gabon (a root mean square error (RMSE) of 3.4 m and 5.6 m, respectively), and correctly estimate vegetation heights up to >50 m. They also show good qualitative agreement with existing vegetation height maps. Our work demonstrates that, given a moderate amount of reference data (i.e., 2000 km$^2$ in Gabon and $\approx$5800 km$^2$ in Switzerland), high-resolution vegetation height maps with 10 m ground sampling distance (GSD) can be derived at country scale from Sentinel-2 imagery.

Riccardo de Lutio, Stefano D'Aronco, Jan Dirk Wegner et al.

Guided super-resolution is a unifying framework for several computer vision tasks where the inputs are a low-resolution source image of some target quantity (e.g., perspective depth acquired with a time-of-flight camera) and a high-resolution guide image from a different domain (e.g., a grey-scale image from a conventional camera); and the target output is a high-resolution version of the source (in our example, a high-res depth map). The standard way of looking at this problem is to formulate it as a super-resolution task, i.e., the source image is upsampled to the target resolution, while transferring the missing high-frequency details from the guide. Here, we propose to turn that interpretation on its head and instead see it as a pixel-to-pixel mapping of the guide image to the domain of the source image. The pixel-wise mapping is parametrised as a multi-layer perceptron, whose weights are learned by minimising the discrepancies between the source image and the downsampled target image. Importantly, our formulation makes it possible to regularise only the mapping function, while avoiding regularisation of the outputs; thus producing crisp, natural-looking images. The proposed method is unsupervised, using only the specific source and guide images to fit the mapping. We evaluate our method on two different tasks, super-resolution of depth maps and of tree height maps. In both cases, we clearly outperform recent baselines in quantitative comparisons, while delivering visually much sharper outputs.

Zuoyue Li, Jan Dirk Wegner, Aurélien Lucchi

We propose a new approach, named PolyMapper, to circumvent the conventional pixel-wise segmentation of (aerial) images and predict objects in a vector representation directly. PolyMapper directly extracts the topological map of a city from overhead images as collections of building footprints and road networks. In order to unify the shape representation for different types of objects, we also propose a novel sequentialization method that reformulates a graph structure as closed polygons. Experiments are conducted on both existing and self-collected large-scale datasets of several cities. Our empirical results demonstrate that our end-to-end learnable model is capable of drawing polygons of building footprints and road networks that very closely approximate the structure of existing online map services, in a fully automated manner. Quantitative and qualitative comparison to the state-of-the-art also shows that our approach achieves good levels of performance. To the best of our knowledge, the automatic extraction of large-scale topological maps is a novel contribution in the remote sensing community that we believe will help develop models with more informed geometrical constraints.

Dimitrios Marmanis, Wei Yao, Fathalrahman Adam et al.

Very High Spatial Resolution (VHSR) large-scale SAR image databases are still an unresolved issue in the Remote Sensing field. In this work, we propose such a dataset and use it to explore patch-based classification in urban and periurban areas, considering 7 distinct semantic classes. In this context, we investigate the accuracy of large CNN classification models and pre-trained networks for SAR imaging systems. Furthermore, we propose a Generative Adversarial Network (GAN) for SAR image generation and test, whether the synthetic data can actually improve classification accuracy.

Pascal Kaiser, Jan Dirk Wegner, Aurelien Lucchi et al.

This study deals with semantic segmentation of high-resolution (aerial) images where a semantic class label is assigned to each pixel via supervised classification as a basis for automatic map generation. Recently, deep convolutional neural networks (CNNs) have shown impressive performance and have quickly become the de-facto standard for semantic segmentation, with the added benefit that task-specific feature design is no longer necessary. However, a major downside of deep learning methods is that they are extremely data-hungry, thus aggravating the perennial bottleneck of supervised classification, to obtain enough annotated training data. On the other hand, it has been observed that they are rather robust against noise in the training labels. This opens up the intriguing possibility to avoid annotating huge amounts of training data, and instead train the classifier from existing legacy data or crowd-sourced maps which can exhibit high levels of noise. The question addressed in this paper is: can training with large-scale, publicly available labels replace a substantial part of the manual labeling effort and still achieve sufficient performance? Such data will inevitably contain a significant portion of errors, but in return virtually unlimited quantities of it are available in larger parts of the world. We adapt a state-of-the-art CNN architecture for semantic segmentation of buildings and roads in aerial images, and compare its performance when using different training data sets, ranging from manually labeled, pixel-accurate ground truth of the same city to automatic training data derived from OpenStreetMap data from distant locations. We report our results that indicate that satisfying performance can be obtained with significantly less manual annotation effort, by exploiting noisy large-scale training data.

Sébastien Lefèvre, Devis Tuia, Jan Dirk Wegner et al.

In this paper, we discuss and review how combined multi-view imagery from satellite to street-level can benefit scene analysis. Numerous works exist that merge information from remote sensing and images acquired from the ground for tasks like land cover mapping, object detection, or scene understanding. What makes the combination of overhead and street-level images challenging, is the strongly varying viewpoint, different scale, illumination, sensor modality and time of acquisition. Direct (dense) matching of images on a per-pixel basis is thus often impossible, and one has to resort to alternative strategies that will be discussed in this paper. We review recent works that attempt to combine images taken from the ground and overhead views for purposes like scene registration, reconstruction, or classification. Three methods that represent the wide range of potential methods and applications (change detection, image orientation, and tree cataloging) are described in detail. We show that cross-fertilization between remote sensing, computer vision and machine learning is very valuable to make the best of geographic data available from Earth Observation sensors and ground imagery. Despite its challenges, we believe that integrating these complementary data sources will lead to major breakthroughs in Big GeoData.

Dimitrios Marmanis, Konrad Schindler, Jan Dirk Wegner et al.

We present an end-to-end trainable deep convolutional neural network (DCNN) for semantic segmentation with built-in awareness of semantically meaningful boundaries. Semantic segmentation is a fundamental remote sensing task, and most state-of-the-art methods rely on DCNNs as their workhorse. A major reason for their success is that deep networks learn to accumulate contextual information over very large windows (receptive fields). However, this success comes at a cost, since the associated loss of effecive spatial resolution washes out high-frequency details and leads to blurry object boundaries. Here, we propose to counter this effect by combining semantic segmentation with semantically informed edge detection, thus making class-boundaries explicit in the model, First, we construct a comparatively simple, memory-efficient model by adding boundary detection to the Segnet encoder-decoder architecture. Second, we also include boundary detection in FCN-type models and set up a high-end classifier ensemble. We show that boundary detection significantly improves semantic segmentation with CNNs. Our high-end ensemble achieves > 90% overall accuracy on the ISPRS Vaihingen benchmark.