MILO seminars are scientific seminars devoted to mathematics and computer science applied or applicable to Earth observation, e.g. machine learning for satellite data analysis. The seminars are held by videoconference every month or so, usually on Tuesdays between 1:30pm and 2:30pm.
The seminars are organized within the COMET TSI, a community of experts in signal and image processing initiated by CNES, the French spatial agency, in order to gather academic researchers and engineers from the space industry in France.
The abundance of unlabeled Satellite Image Time Series (SITS) from missions such as Sentinel-2 contrasts with the scarcity of expert annotations, making self-supervised learning (SSL) an appealing strategy for large-scale remote sensing. Contrastive SSL is a natural way to exploit such data, yet its application to SITS remains challenging: data augmentation, a key component of contrastive learning, is less established for time series than for natural images. We first introduce FranceCrops, a large-scale dataset of 6.3 million agricultural parcels designed for SSL on SITS, providing unlabeled data for pre-training alongside labeled subsets for finetuning and evaluation. Then, we propose two methods to make contrastive SSL effective on SITS. First, GaPP (Groups as Positive Pairs) bypasses augmentations altogether by exploiting pre-existing groupings within the data to form positive pairs. Second, a temporal resampling augmentation creates two distinct yet coherent views of the same time series. Together, these approaches reach state-of-the-art performance in low-label settings on crop classification tasks and highlight the importance of positive-pair quality in SITS representation learning.
Neural networks have been successfully applied to image restoration tasks. The standard approach directly learns to predict the restored image from the degraded one in a supervised manner. However, such methods are often unsuitable for satellite image restoration, as they lack interpretability and require retraining for each specific sensor. Alternative approaches leverage neural networks to learn a prior over target images, which is then integrated into a classical variational optimization framework for solving inverse problems. This allows for restoring images from different sources using the same trained network. We focus on a line of work that looks for the solution of the restoration task in the encoded - or latent - space of a trained generative model. We introduce a variational Bayes framework, VBLE-xz (Variational Bayes Latent Estimation), which enables approximating the posterior distribution of the inverse problem within variational autoencoders (VAEs). Furthermore, we adopt compressive VAEs, whose lightweight architecture ensures a scalable restoration process while effectively regularizing the inverse problem through their hyperprior. Experimental results on satellite images highlight the benefits of VBLE-xz over state-of-the-art deep learning baselines for satellite image restoration.
Humans perceive the world through multisensory integration, blending the information of different modalities to adapt their behavior. Contrastive learning offers an appealing solution for multimodal self-supervised learning. Indeed, by considering each modality as a different view of the same entity, it learns to align features of different modalities in a shared representation space. However, this approach is intrinsically limited as it only learns shared or redundant information between modalities, while multimodal interactions can arise in other ways. In this work, we introduce CoMM, a Contrastive Multimodal learning strategy that enables the communication between modalities in a single multimodal space. Instead of imposing cross- or intra- modality constraints, we propose to align multimodal representations by maximizing the mutual information between augmented versions of these multimodal features. Our theoretical analysis shows that shared, synergistic and unique terms of information naturally emerge from this ormulation, allowing us to estimate multimodal interactions beyond redundancy. We test CoMM both in a controlled and in a series of real-world settings: in the former, we demonstrate that CoMM effectively captures redundant, unique and synergistic information between modalities. In the latter, CoMM learns complex multimodal interactions and achieves state-of-the-art results on the six multimodal benchmarks.
Recently, Gaussian splatting has emerged as a strong alternative to NeRF, demonstrating impressive 3D modeling capabilities while requiring only a fraction of the training and rendering time. In this paper, we show how the standard Gaussian splatting framework can be adapted for remote sensing, retaining its high efficiency. This enables us to achieve state-of-the-art performance in just a few minutes, compared to the day-long optimization required by the best-performing NeRF-based Earth observation methods. The proposed framework incorporates remote-sensing improvements from EO-NeRF, such as radiometric correction and shadow modeling, while introducing novel components, including sparsity, view consistency, and opacity regularizations.
Reliably modeling the distribution of species at a large scale is critical for understanding the drivers
of biodiversity loss in our rapidly changing climate. Specifically, the task involves predicting the probability that
a species will occur in a particular location given the prevailing environmental conditions.Traditional approaches from
the field of ecology often tackle the task from a statistical perspective, fitting a per-species density function around
the known occurrences in feature space. These features are hand-crafted from environmental data, and spatial reasoning
is limited to the exact location of a known occurrence.
An expanding body of research leverages deep neural networks to learn features in an end-to-end manner
for jointly modeling the distribution of multiple species, capturing potential interactions. These models can derive
rich representations from diverse inputs, including climatic rasters, satellite imagery, and species interaction graphs.
In this talk, we will introduce the task of modeling species distributions and give a brief overview of established approaches
in the field, along with their limitations. We will then explore in more detail how the potential of deep learning can help
to overcome these limitations and arrive at a more holistic view of species distribution models.
Up-to-date and precise mapping of the Earth's surface is critical for monitoring and mitigating the effects
of global warming. For about a decade, Earth Observation (EO) missions have been providing frequently high-resolution
multispectral imagery of the entire globe. To exploit the resulting Satellite Image Time Series (SITS),
Deep Neural Networks have become increasingly popular. Nonetheless, these methods require large amounts of labeled data
for training, limiting their scalability across diverse geographic regions and time periods. In response,
there is growing support for pre-training large deep neural networks, known as foundation models,
using self-supervised learning techniques. These techniques allow models to learn high-level representations
of input data without the need for labeled datasets. Once the model is pre-trained, it is utilized to provide
representations that serve as inputs to shallow models in many downstream tasks.
This presentation focuses on the development of easy-to-use representations of SITS for climate experts and
geoscientists. In particular, we propose a novel DNN architecture (ALISE) capable of generating aligned and
fixed-size representations from inherently irregular and unaligned SITS input data. In addition, we explore hybrid
self-supervised learning strategies to improve the extraction of high semantic features in the representations.
For the downstream applications considered, a single linear layer is trained on the representations obtained by
the pre-trained and frozen model, demonstrating the simplicity of exploiting the ALISE representations.
Structural pruning of neural networks conventionally relies on identifying and discarding less important neurons, a practice often resulting in significant accuracy loss that necessitates subsequent fine-tuning efforts. This paper introduces a novel approach named Intra-Fusion, challenging this prevailing pruning paradigm. Unlike existing methods that focus on designing meaningful neuron importance metrics, Intra-Fusion redefines the overlying pruning procedure. Through utilizing the concepts of model fusion and optimal transport, we leverage an agnostically given importance metric to arrive at a more effective sparse model representation. Notably, our approach achieves substantial accuracy recovery without the need for resource-intensive fine-tuning, making it an efficient and promising tool for neural network compression.
I present the first generative diffusion surrogate model tailored to sea-ice processes. Trained on over 20 years of coupled neXtSIM-NEMO mesoscale simulations (~12 km resolution) in the region north of Svalbard, the model predicts 12-hour sea-ice evolution. Its inherent stochastic nature enables robust ensemble forecasting, outperforming all baselines in forecast error while resolving the smoothing limitations of deterministic surrogates. Crucially, I demonstrate for the first time that a fully data-driven model can generate physically consistent forecasts akin to those from neXtSIM. While this marks a significant advancement in realistic surrogate modelling, it incurs substantial computational costs compared to traditional approaches. Hence, I will showcase efforts to scale the model to Arctic-wide applications, including leveraging generative diffusion in a learned reduced space, and training at different resolutions for an enhanced performance.
Methods inspired by Artificial Intelligence (AI) are starting to fundamentally change computational science and engineering through breakthrough performances on challenging problems. However, reliability and trustworthiness of such techniques is becoming a major concern. In inverse problems in imaging, the focus of this talk, there is increasing empirical evidence that methods may suffer from hallucinations, i.e., false, but realistic-looking artifacts; instability, i.e., sensitivity to perturbations in the data; and unpredictable generalization, i.e., excellent performance on some images, but significant deterioration on others. This talk presents a potential mathematical framework describing how and when such effects arise in arbitrary reconstruction methods, not just AI-inspired techniques. Several of our results take the form of `no free lunch' theorems. Our results trace these effects to the kernel of the forward operator whenever it is nontrivial, but also extend to the case when the forward operator is ill-conditioned. Lastly, an outlook onto the relevance and utility of these findings in Earth Observation is presented.
In this presentation, we will consider the problem of producing diverse super-resolved images corresponding to a low-resolution observation. From a probabilistic perspective, this amounts to sample from the posterior distribution of the super-resolution inverse problem. As a prior model on the solution, we propose to use pretrained hierarchical variational autoencoder (HVAE), a powerful class of deep generative model. We train a lightweight stochastic encoder to encode low-resolution images in the latent space of a pretrained HVAE. At inference, we combine the low-resolution encoder and the pretrained generative model to super-resolve an image. We demonstrate on the task of face super-resolution that our method provides an advantageous trade-off between the computational efficiency of conditional normalizing flows techniques and the sample quality of diffusion based methods.
Representation learning models gained traction as a way to leverage the vast amounts of unlabeled Earth Observation data. However, due to the multiplicity of remote sensing sources, these models should learn sensor agnostic representations, that generalize across sensor characteristics with minimal fine-tuning. Sensor agnosticism is strongly related to multi-modality: more and more models deal with different kinds of data, such as MSI, coordinates, and so on. These models should also be capable of dealing with images from different locations and different timeframes, being spatiotemporal invariant. The talk will investigate different approaches to develop these models, such as (continual) self-supervised learning, spanning different downstream tasks, such as semantic segmentation and domain adaptation.