Rapid Adaptation of Earth Observation Foundation Models for Segmentation

This study investigates the efficacy of Low-Rank Adaptation (LoRA) in fine-tuning Earth Observation (EO) foundation models for flood segmentation. We hypothesize that LoRA, a parameter-efficient technique, can significantly accelerate the adaptation of large-scale EO models to this critical task while maintaining high performance. We apply LoRA to fine-tune a state-of-the-art EO foundation model pre-trained on diverse satellite imagery, using a curated dataset of flood events. Our results demonstrate that LoRA-based fine-tuning (r-256) improves F1 score by 6.66 points and IoU by 0.11 compared to a frozen encoder baseline, while significantly reducing computational costs. Notably, LoRA outperforms full fine-tuning, which proves computationally infeasible on our hardware. We further assess generalization through out-of-distribution (OOD) testing on a geographically distinct flood event. While LoRA configurations show improved OOD performance over the baseline. This work contributes to research on efficient adaptation of foundation models for specialized EO tasks, with implications for rapid response systems in disaster management. Our findings demonstrate LoRA's potential for enabling faster deployment of accurate flood segmentation models in resource-constrained, time-critical scenarios.

Uncertainty and Generalizability in Foundation Models for Earth Observation

We take the perspective in which we want to design a downstream task (such as estimating vegetation coverage) on a certain area of interest (AOI) with a limited labeling budget. By leveraging an existing Foundation Model (FM) we must decide whether we train a downstream model on a different but label-rich AOI hoping it generalizes to our AOI, or we split labels in our AOI for training and validating. In either case, we face choices concerning what FM to use, how to sample our AOI for labeling, etc. which affect both the performance and uncertainty of the results. In this work, we perform a large ablative study using eight existing FMs on either Sentinel 1 or Sentinel 2 as input data, and the classes from the ESA World Cover product as downstream tasks across eleven AOIs. We do repeated sampling and training, resulting in an ablation of some 500K simple linear regression models. Our results show both the limits of spatial generalizability across AOIs and the power of FMs where we are able to get over 0.9 correlation coefficient between predictions and targets on different chip level predictive tasks. And still, performance and uncertainty vary greatly across AOIs, tasks and FMs. We believe this is a key issue in practice, because there are many design decisions behind each FM and downstream task (input modalities, sampling, architectures, pretraining, etc.) and usually a downstream task designer is aware of and can decide upon a few of them. Through this work, we advocate for the usage of the methodology herein described (large ablations on reference global labels and simple probes), both when publishing new FMs, and to make informed decisions when designing downstream tasks to use them.

M3LEO: A Multi-Modal, Multi-Label Earth Observation Dataset Integrating Interferometric SAR and RGB Data

Satellite-based remote sensing has revolutionised the way we address global challenges in a rapidly evolving world. Huge quantities of Earth Observation (EO) data are generated by satellite sensors daily, but processing these large datasets for use in ML pipelines is technically and computationally challenging. Specifically, different types of EO data are often hosted on a variety of platforms, with differing availability for Python preprocessing tools. In addition, spatial alignment across data sources and data tiling can present significant technical hurdles for novice users. While some preprocessed EO datasets exist, their content is often limited to optical or near-optical wavelength data, which is ineffective at night or in adverse weather conditions. Synthetic Aperture Radar (SAR), an active sensing technique based on microwave length radiation, offers a viable alternative. However, the application of machine learning to SAR has been limited due to a lack of ML-ready data and pipelines, particularly for the full diversity of SAR data, including polarimetry, coherence and interferometry. We introduce M3LEO, a multi-modal, multi-label EO dataset that includes polarimetric, interferometric, and coherence SAR data derived from Sentinel-1, alongside Sentinel-2 RGB imagery and a suite of labelled tasks for model evaluation. M3LEO spans 17.5TB and contains approximately 10M data chips across six geographic regions. The dataset is complemented by a flexible PyTorch Lightning framework, with configuration management using Hydra. We provide tools to process any dataset available on popular platforms such as Google Earth Engine for integration with our framework. Initial experiments validate the utility of our data and framework, showing that SAR imagery contains information additional to that extractable from RGB data. Data at huggingface.co/M3LEO, and code at github.com/spaceml-org/M3LEO.

Exploring DINO: Emergent Properties and Limitations for Synthetic Aperture Radar Imagery

Self-supervised learning (SSL) models have recently demonstrated remarkable performance across various tasks, including image segmentation. This study delves into the emergent characteristics of the Self-Distillation with No Labels (DINO) algorithm and its application to Synthetic Aperture Radar (SAR) imagery. We pre-train a vision transformer (ViT)-based DINO model using unlabeled SAR data, and later fine-tune the model to predict high-resolution land cover maps. We rigorously evaluate the utility of attention maps generated by the ViT backbone, and compare them with the model's token embedding space. We observe a small improvement in model performance with pre-training compared to training from scratch, and discuss the limitations and opportunities of SSL for remote sensing and land cover segmentation. Beyond small performance increases, we show that ViT attention maps hold great intrinsic value for remote sensing, and could provide useful inputs to other algorithms. With this, our work lays the ground-work for bigger and better SSL models for Earth Observation.

Exploring Generalisability of Self-Distillation with No Labels for SAR-Based Vegetation Prediction

In this work we pre-train a DINO-ViT based model using two Synthetic Aperture Radar datasets (S1GRD or GSSIC) across three regions (China, Conus, Europe). We fine-tune the models on smaller labeled datasets to predict vegetation percentage, and empirically study the connection between the embedding space of the models and their ability to generalize across diverse geographic regions and to unseen data. For S1GRD, embedding spaces of different regions are clearly separated, while GSSIC's overlaps. Positional patterns remain during fine-tuning, and greater distances in embeddings often result in higher errors for unfamiliar regions. With this, our work increases our understanding of generalizability for self-supervised models applied to remote sensing.

Large Scale Masked Autoencoding for Reducing Label Requirements on SAR Data

Satellite-based remote sensing is instrumental in the monitoring and mitigation of the effects of anthropogenic climate change. Large scale, high resolution data derived from these sensors can be used to inform intervention and policy decision making, but the timeliness and accuracy of these interventions is limited by use of optical data, which cannot operate at night and is affected by adverse weather conditions. Synthetic Aperture Radar (SAR) offers a robust alternative to optical data, but its associated complexities limit the scope of labelled data generation for traditional deep learning. In this work, we apply a self-supervised pretraining scheme, masked autoencoding, to SAR amplitude data covering 8.7\% of the Earth's land surface area, and tune the pretrained weights on two downstream tasks crucial to monitoring climate change - vegetation cover prediction and land cover classification. We show that the use of this pretraining scheme reduces labelling requirements for the downstream tasks by more than an order of magnitude, and that this pretraining generalises geographically, with the performance gain increasing when tuned downstream on regions outside the pretraining set. Our findings significantly advance climate change mitigation by facilitating the development of task and region-specific SAR models, allowing local communities and organizations to deploy tailored solutions for rapid, accurate monitoring of climate change effects.

Fewshot learning on global multimodal embeddings for earth observation tasks

In this work we pretrain a CLIP/ViT based model using three different modalities of satellite imagery across five AOIs covering over ~10\% of the earth total landmass, namely Sentinel 2 RGB optical imagery, Sentinel 1 SAR amplitude and Sentinel 1 SAR interferometric coherence. This model uses $\sim 250$ M parameters. Then, we use the embeddings produced for each modality with a classical machine learning method to attempt different downstream tasks for earth observation related to vegetation, built up surface, croplands and permanent water. We consistently show how we reduce the need for labeled data by 99\%, so that with ~200-500 randomly selected labeled examples (around 4K-10K km$^2$) we reach performance levels analogous to those achieved with the full labeled datasets (about 150K image chips or 3M km$^2$ in each AOI) on all modalities, AOIs and downstream tasks. This leads us to think that the model has captured significant earth features useful in a wide variety of scenarios. To enhance our model's usability in practice, its architecture allows inference in contexts with missing modalities and even missing channels within each modality. Additionally, we visually show that this embedding space, obtained with no labels, is sensible to the different earth features represented by the labelled datasets we selected.

A Contrastive Method Based on Elevation Data for Remote Sensing with Scarce and High Level Semantic Labels

This work proposes a hybrid unsupervised/supervised learning method to pretrain models applied in earth observation downstream tasks where only a handful of labels denoting very general semantic concepts are available. We combine a contrastive approach to pretrain models with a pretext task to predict spatially coarse elevation maps which are commonly available worldwide. The intuition behind is that there is generally some correlation between the elevation and targets in many remote sensing tasks, allowing the model to pre-learn useful representations. We assess the performance of our approach on a segmentation downstream task on labels gathering many possible subclasses (pixel level classification of farmlands vs. other) and an image binary classification task derived from the former, on a dataset on the north-east of Colombia. On both cases we pretrain our models with 39K unlabeled images, fine tune the downstream task only with 80 labeled images and test it with 2944 labeled images. Our experiments show that our methods, GLCNet+Elevation for segmentation and SimCLR+Elevation for classification, outperform their counterparts without the elevation pretext task in terms of accuracy and macro-average F1, which supports the notion that including additional information correlated to targets in downstream tasks can lead to improved performance.

Deep learning based landslide density estimation on SAR data for rapid response

This work aims to produce landslide density estimates using Synthetic Aperture Radar (SAR) satellite imageries to prioritise emergency resources for rapid response. We use the United States Geological Survey (USGS) Landslide Inventory data annotated by experts after Hurricane Mar\'ia in Puerto Rico on Sept 20, 2017, and their subsequent susceptibility study which uses extensive additional information such as precipitation, soil moisture, geological terrain features, closeness to waterways and roads, etc. Since such data might not be available during other events or regions, we aimed to produce a landslide density map using only elevation and SAR data to be useful to decision-makers in rapid response scenarios. The USGS Landslide Inventory contains the coordinates of 71,431 landslide heads (not their full extent) and was obtained by manual inspection of aerial and satellite imagery. It is estimated that around 45\% of the landslides are smaller than a Sentinel-1 typical pixel which is 10m $\times$ 10m, although many are long and thin, probably leaving traces across several pixels. Our method obtains 0.814 AUC in predicting the correct density estimation class at the chip level (128$\times$128 pixels, at Sentinel-1 resolution) using only elevation data and up to three SAR acquisitions pre- and post-hurricane, thus enabling rapid assessment after a disaster. The USGS Susceptibility Study reports a 0.87 AUC, but it is measured at the landslide level and uses additional information sources (such as proximity to fluvial channels, roads, precipitation, etc.) which might not regularly be available in an rapid response emergency scenario.

SAR-based landslide classification pretraining leads to better segmentation

Rapid assessment after a natural disaster is key for prioritizing emergency resources. In the case of landslides, rapid assessment involves determining the extent of the area affected and measuring the size and location of individual landslides. Synthetic Aperture Radar (SAR) is an active remote sensing technique that is unaffected by weather conditions. Deep Learning algorithms can be applied to SAR data, but training them requires large labeled datasets. In the case of landslides, these datasets are laborious to produce for segmentation, and often they are not available for the specific region in which the event occurred. Here, we study how deep learning algorithms for landslide segmentation on SAR products can benefit from pretraining on a simpler task and from data from different regions. The method we explore consists of two training stages. First, we learn the task of identifying whether a SAR image contains any landslides or not. Then, we learn to segment in a sparsely labeled scenario where half of the data do not contain landslides. We test whether the inclusion of feature embeddings derived from stage-1 helps with landslide detection in stage-2. We find that it leads to minor improvements in the Area Under the Precision-Recall Curve, but also to a significantly lower false positive rate in areas without landslides and an improved estimate of the average number of landslide pixels in a chip. A more accurate pixel count allows to identify the most affected areas with higher confidence. This could be valuable in rapid response scenarios where prioritization of resources at a global scale is important. We make our code publicly available at https://github.com/VMBoehm/SAR-landslide-detection-pretraining.