Satellite image deep learning

In this episode I sat down with Isaac to discuss RF-DETR, a new state-of-the-art family of real-time object detection and segmentation models from Roboflow. We cover the motivation for building models that are not just accurate but also fast, cost-efficient, and deployable across diverse hardware and data regimes, and why moving beyond fixed architectures is key to achieving that. Isaac explains how RF-DETR combines strong foundation backbones like DINOv2 with efficient neural architecture search to unlock novel speed–accuracy trade-offs, including dropping decoder layers and queries after training. We also discuss the model’s strong transfer performance on domains far from COCO, the introduction of a memory-efficient instance segmentation head, and the team’s unusually rigorous benchmarking approach, before closing on the challenges of open-source research and upcoming improvements to inference and platform integration.
* 👤 Isaac on LinkedIn
* 🖥️ RF-DETR on Github
* 📖 Paper
* 📺 Video of this conversation on YouTube
Bio: Isaac Robinson is a Machine Learning Research Engineer at Roboflow. He’s worked across the field of computer vision, from real-time stereo depth estimation on household robots to biomedical research at the NIH to founding a zero shot computer vision infrastructure startup. Isaac focusses on the intersection of low latency and high performance, with the goal of helping people unlock new capabilities through vision.


This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.satellite-image-deep-learning.com

In this episode I caught up with Sadiq Jaffer and Frank Feng to discuss Tessera, a large-scale foundation model for Earth observation that produces annual, pixel-level temporal embeddings from multi-sensor satellite data. They explain why moving beyond single-date imagery is essential for understanding phenology, land cover, and environmental change, and how aggregating a full year of Sentinel-1 and Sentinel-2 observations enables far richer representations of the Earth’s surface. We dive into the unique engineering challenges behind Tessera, including its unusual cost profile where inference is more expensive than training, the need to ingest petabyte-scale archives, and the design choices required to scale a pixel-based model without representation collapse. Frank walks through their self-supervised training strategy based on redundancy reduction (Barlow Twins), while Sadiq highlights how downstream evaluations—from wildfire analysis to land-cover mapping—demonstrate that the embeddings already encode meaningful temporal and semantic structure. We also discuss the practical impact for ecology and conservation, where Tessera dramatically accelerates research workflows and reduces label requirements, and look ahead to Tessera v2, which will incorporate Landsat data to extend embeddings back to the 1970s and unlock new capabilities in change detection and forecasting.
* 📺 This conversation on YouTube
* 🖥️ Tessera on Github
* 📖 Paper
* 🖥️ Franks website
* 🖥️ Sadiqs website
Slides discussed in the episode



This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.satellite-image-deep-learning.com

In this episode I caught up with Roberto del Prete to learn about his work on AutoML for in-orbit model deployment, and how it enables satellites to run highly efficient AI models under severe power and hardware constraints. Roberto explains why traditional computer-vision architectures—optimised for ImageNet or COCO—are a poor fit for narrow, mission-specific tasks like wildfire or vessel detection, and why models must be co-designed with the actual edge devices flying in space. He describes PyNAS, his neural architecture search framework, in which a genetic algorithm drives the optimisation process, evolving compact, hardware-aware neural networks and profiling them directly on representative onboard processors such as Intel Myriad and NVIDIA Jetson. We discuss the multiobjective challenge of balancing accuracy and latency, the domain gap between training data and new sensor imagery, and how lightweight models make post-launch fine-tuning and updates far more practical. Roberto also outlines the rapidly changing ecosystem of spaceborne AI hardware and why efficient optimisation will remain central to future AI-enabled satellite constellations.
* 🖥️ PyNAS on Github
* 📖 Nature paper
* 📺 Video of this conversation on YouTube
* 👤 Roberto on LinkedIn
Bio
Roberto is an Internal Research Fellow at ESA Φ-lab specialising in deep learning and edge computing for remote sensing. He focuses on improving time-critical decision-making through advanced AI solutions for space missions and Earth monitoring. He holds a Ph.D. at the University of Naples Federico II, where he also earned his Master’s and Bachelor’s degrees in Aerospace Engineering. His notable work includes the development of “FederNet,” a terrain relative navigation system. Del Prete’s professional experience includes roles as a Visiting Researcher at the European Space Agency’s Φ-Lab and SmartSat CRC in Australia. He has contributed to key projects like Kanyini Mission, and developed AI algorithms for real-time maritime monitoring and thermal anomaly detection. He co-developed the award-winning P³ANDA project, a compact AI-powered imaging system, earning the 2024 Telespazio Technology Contest prototype prize. Co-author of more than 30 scientific publications, Del Prete is dedicated to leveraging advanced technologies to address global challenges in remote sensing and AI.


This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.satellite-image-deep-learning.com

In this episode I caught up with Julia Wąsala to learn about methane plume detection using AutoML, and how her research bridges atmospheric science and machine learning. Julia explains the unique challenges of working with TROPOMI data—extremely coarse spatial resolution, single-channel methane measurements, and complex auxiliary fields that sometimes create plume-like artefacts leading to false detections. She walks through how her approach generalises a traditional two-stage detection pipeline to multiple gases using AutoMergeNet, a neural architecture search framework that automatically designs multimodal CNNs tailored to different atmospheric gases. We discuss why methane matters, how model performance shifts dramatically between curated test sets and real-world global data, and the ongoing effort to understand sampling bias and improve operational precision.
* 📖 AutoMergeNet paper
* 🖥️ Code on Github
* 🖥️ Julia’s homepage
* 📺 Recording of this conversation on YouTube
Bio: Julia Wąsala is currently working toward the Ph.D. degree in automated machine learning for Earth observation with the Leiden Institute for Advanced Computer Science, Leiden University, Leiden, The Netherlands, and with Space Research Organisation Netherlands, Leiden, The Netherlands. Her research focuses on the field of automated machine learning for earth observation focuses on designing new methods and validating them in real-world applications, such as atmospheric plume detection.


This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.satellite-image-deep-learning.com

In this episode, I caught up with Ibrahim Salihu Yusuf from InstaDeep’s AI for Social Good team to hear the story behind InstaGeo, an open-source geospatial machine learning framework built to make multispectral satellite data easy to use for real-world applications. Ibrahim explains how the 2019–2020 locust outbreak exposed a gap between freely available satellite imagery, existing machine learning models, and the lack of tools to turn raw data into model-ready inputs. He walks through how InstaGeo bridges this gap - fetching, processing, and preparing multispectral data; fine-tuning models such as NASA IBM’s Prithvi; and delivering end-to-end inference and visualisation in a unified app. The conversation also covers practical use cases, from locust breeding ground detection to damage assessment, air quality, and biomass estimation, as well as the team’s efforts to partner with field organisations to drive on-the-ground impact.
* 👤 Ibrahim on LinkedIn
* 🖥️ InstaGeo on Github
* 📖 Paper on InstaGeo
* 📺 Video of this conversation on YouTube
* 📺 Demo of InstaGeo on YouTube
Bio: Ibrahim is a Senior Research Engineer and Technical Lead of the AI for Social Good team at InstaDeep’s Kigali office, where he applies artificial intelligence to address real-world challenges and drive social impact across Africa and beyond. With expertise spanning geospatial machine learning, computer vision, and computational biology, he has led high-impact projects in food security, disaster response, and immunology research. He also leads the development of InstaGeo, a platform designed to democratise access to AI-powered insights from open-source satellite imagery, reflecting his commitment to using cutting-edge AI for meaningful societal benefit.


This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.satellite-image-deep-learning.com