Behind the Spottitt MF Curtains: Role of Machine Learning in Analyzing Satellite Imagery

• 
  October 23, 2023

Spottitt is known for utilising satellite data to offer infrastructure owners valuable insights into the diverse external risks affecting their assets. But how precisely do we derive these insights from satellite imagery? The answer is – machine learning, and today we’re inviting you behind the curtains of our product to unveil the whole process.

Machine Learning: Bridging Raw Satellite Data to Insights

Machine learning is like the sophisticated lens through which we decode the vast and intricate tapestry of Earth Observation (EO) data. It empowers us to uncover hidden patterns, predict future changes, and transform complex satellite images into actionable insights. 

Whether it’s predicting natural disasters, monitoring biodiversity, vegetation or, vice versa, deforestation, machine learning acts as the bridge between raw satellite data and our understanding of the conditions on the ground. 

Machine learning in the realm of Earth Observation data is like a masterful detective, meticulously identifying specific objects, segmenting vast satellite images, and categorizing distinct patterns, such as different habitats, with precision. 

For instance, imagine a vast forested region as seen from space. Machine learning can dissect that image, differentiating between water bodies, different tree species, and other habitats – a task referred to as ‘habitat classification‘. Beyond just understanding, machine learning automates this entire analytical process. This automation relieves the heavy lifting for experts, allowing them to cover larger terrains in the same or even shorter timeframes. By leveraging machine learning, we interpret, understand, and engage with it more efficiently and at an unprecedented scale.

Art and Science of Processing Satellite Data for Insightful Analysis

By carefully crafting each image through preprocessing steps, we set the stage to transform complex satellite captures into clear, actionable insights. Preparing satellite data for insightful analysis is akin to a chef preparing ingredients before cooking a dish. Several preprocessing steps ensure the data’s quality and readiness:

1. Radiometric Correction: This step adjusts for sensor irregularities and corrects any distortions due to the sensor’s sensitivity. Essentially, it ensures that the recorded brightness of an object is consistent across different satellites or sensors.
2. Geometric Correction: Satellites move fast, and the Earth’s rotation and terrain variation can cause distortions. Geometric correction rectifies these distortions so that the images align accurately with geographical locations on Earth.
3. Atmospheric Correction: Our atmosphere can interfere with satellite readings, especially in the case of clouds or atmospheric particles. This step removes those interferences to provide a clearer view of the Earth’s surface.
4. Cloud Masking: Since clouds can obscure the view of the Earth’s surface, they need to be identified and masked or removed from the analysis.
5. Image Registration: If we’re comparing multiple images over time, we need to ensure they align perfectly. Image registration does this by overlapping images to ensure consistent analysis.
6. Resampling: Different satellites might have different resolutions. Resampling adjusts these variations to make sure we’re comparing apples to apples.
7. Data Augmentation (for Machine Learning): By slightly altering or rotating images, we can increase the amount of data available for training, enhancing the robustness of our models.

With these steps in place, we not only ensure the accuracy of our satellite images but also set the stage for extracting meaningful, actionable insights through advanced analysis, including machine learning.

Mapping Earth’s Complexity: Extracting Satellite Imagery Features

Extracting features from satellite imagery is akin to finding the individual notes in a symphony. Each note, or feature, tells a part of Earth’s dynamic story. From a machine learning engineer’s perspective working on Earth Observation, here’s what we glean from the raw satellite data:

1. Spectral Features: These are unique ‘signatures’ from different surface materials on Earth. Just as our eyes detect colors, satellites detect a broader spectrum, capturing details invisible to us. From this, we can identify vegetation types, water bodies, or human-made structures.
2. Texture Features: By examining the arrangement and frequency of pixel values, we discern patterns. This helps in identifying areas with similar characteristics, like urban zones, vegetation areas or agricultural fields.
3. Temporal Features: Satellites passing over the same area capture changes over time. Tracking these changes can provide insights into phenomena like urban expansion, deforestation, or the growth and decline of ice.
4. Shape Features: The geometry and shape of objects can be crucial. For example, the specific layout of buildings and roads can tell us about urban or transportation planning and infrastructure.
5. Contextual Features: It’s not just about individual items but their relation to each other. The proximity of a forested area to a water body might provide clues about potential biodiversity hotspots.
6. Topographical Features: Derived from elevation data, these features give insights into the terrain, helping in tasks like flood prediction or identifying suitable land for construction.

By meticulously extracting these features, we convert raw, often overwhelming satellite imagery into a rich tapestry of information. This is where our clients get handy insights for the monitoring of their assets. 

1. Building Detection: Satellite images are teeming with structural details. We pinpoint specific patterns and shapes, often rectangular or grid-like, to identify buildings amidst the natural landscape. Coupled with shadow analysis, this allows us to locate and categorize structures in both urban and remote settings.
2. Tree & Vegetation Detection: By harnessing spectral features, we can recognize vegetation signatures. Different vegetation types reflect light in unique ways. This capability allows us to distinguish between tree canopies, shrubs, crops, and more, enabling accurate mapping of parks, and agricultural lands.
3. Water Detection: Water bodies have distinctive spectral signatures, especially in infrared wavelengths. By capturing these, we can precisely delineate lakes, rivers, and reservoirs, and even detect moisture levels in soil.
4. Change Detection: One of the marvels of satellite imagery is its ability to capture the Earth’s evolution. By comparing images over time, we identify differences – be it the urban sprawl of a growing city, a flood changing its course, or forests being cleared.
5. Height Estimation: Shadows and multi-angle imagery can be interpreted to gauge the height of structures and natural features. This is crucial for tasks such as vegetation height analysis, skyscraper mapping in cities or mountain range analysis.
6. Ground Movement Analysis: By observing shifts in the Earth’s features over a sequence of images, we can detect ground movements. This becomes indispensable in monitoring scenarios like landslides, subsidence, or movement of assets located on this land.

Every pixel in a satellite image carries a tale, a hint of the world below. With machine learning, we’re not just capturing photographs from space; we’re weaving an evolving narrative of our planet and objects on it, ensuring we remain informed stewards of the future.

Machine Learning Algorithms: Master Tools in Decoding Satellite Imagery

In the grand theatre of Earth Observation, using satellite imagery, machine learning algorithms play the starring roles. Choosing the right algorithm is much like selecting the right tool for a specific task. The vastness and complexity of satellite data demand that we match our machine-learning tools precisely with the problems at hand.

1. Classification: If you envision satellite images as intricate puzzles, classification is about sorting out each piece correctly. For simpler classification tasks, such as distinguishing between water bodies and land, traditional algorithms like Decision Trees can be incredibly effective. They provide a clear, hierarchical method of making decisions based on specific features.
2. Object Detection: When it comes to identifying and locating specific objects within satellite images, such as buildings, vehicles, or even specific tree types, Convolutional Neural Networks (CNNs) are our go-to.
3. Anomaly Detection: When searching for unusual patterns or changes within satellite images, simpler models might sometimes be the better pick. For instance, Decision Trees or ensemble methods like Random Forest can be adept at flagging areas of interest based on established criteria, especially when we have prior knowledge of what to look for.

The key takeaway? Machine learning in EO isn’t a one-size-fits-all domain. It’s about pairing the right algorithm with the task at hand, ensuring this perfect match is what leads us to insights that are not just accurate, but truly illuminating.

From Pixels to Understanding: The Art of Training Earth Observation Models

In the exhilarating world of EO, training a machine learning model is akin to teaching it the language of our planet. It’s a journey of transforming a machine into an expert interpreter of satellite images. Let’s journey through this transformative process step by step:

1. Data Collection: Our story begins with gathering satellite images. These images are our raw material, offering a bird’s-eye view of the Earth’s varied landscapes, from bustling cities to serene forests.
2. Data Preprocessing: Like a sculptor chiseling a block of marble, we refine and clean these images. This might involve enhancing their quality, removing atmospheric disturbances, or even segmenting them into meaningful patches.
3. Labeling: Here, the real teaching begins. We annotate specific features in the images—labeling forests, rivers, buildings, and so on. It’s a way of telling our model: ‘This is what a river looks like’ or ‘Here’s how an urban area appears from space.’
4. Model Selection: Think of this as choosing the best tutor for our apprentice. Depending on the task at hand, be it detecting specific objects or classifying terrains, we select an appropriate machine learning algorithm.
5. Training: With our data prepared and a tutor (model) chosen, the learning begins. Our model goes through the satellite images, learning patterns, shapes, and colors associated with each label. The goal? To recognize similar features on its own in the future.
6. Validation: No learning is complete without tests. We present our model with new, unseen images to assess its accuracy. It’s our way of ensuring that the model has not just memorized the training data but truly understands it.
7. Fine-Tuning: Based on the validation results, we might make a few tweaks here and there, refining our model to perfection. It’s all about iterative improvement.
8. Deployment: Once trained to our satisfaction, our model is ready for the real world! It can now autonomously analyze vast swathes of satellite data, extracting meaningful insights with remarkable accuracy.

Training a machine learning model for Earth Observation is not a one-off task; it’s an ongoing journey of exploration and refinement. Just as our planet is dynamic and ever-evolving, our models too must adapt and evolve to reflect these changes.

• Continuous Improvement: Just like any seasoned professional, our models benefit from continuous learning. As we acquire more satellite data and as landscapes change over time, adding this new information allows our models to stay updated. This is crucial because as the Earth changes – be it through urban development, natural disasters, or climate variations – our models need to recognize and adapt to these changes.
• Quality Enhancement: It’s not just about quantity, but quality too. Even if our models have been trained on a vast amount of data, there might be scenarios where they falter. Perhaps they misclassify a newly developed urban region or miss subtle signs of deforestation. In such cases, we revisit and refine. We dissect where they went wrong, and by adding more targeted data or tweaking the algorithms, we ensure they perform better next time.
• Incorporating Feedback: As with any learning process, feedback is invaluable. Whether it’s from experts analyzing the results or from automated systems highlighting potential errors, this feedback is the compass that guides our model improvements.
• Embracing Innovation: The field of machine learning is dynamic, with new methodologies and techniques emerging regularly. By staying abreast of these developments and incorporating the latest and greatest in our models, we ensure that our Earth Observation tools are not just current but cutting-edge.

In summary, while the initial training of our machine learning models sets the foundation, it’s the persistent refinement and thirst for excellence that transforms them into unparalleled tools, helping our clients monitor their critical infrastructure more effectively than ever before.

Ensuring Precision in EO: Validating Machine Learning Models

In the intricate world of Earth Observation, ensuring the accuracy of our machine-learning models is paramount, as they influence decisions that impact the safety, reliability, and efficiency of assets, individuals, and the supply chain. So, how can we validate the precision of these tools? 

1. Training and Testing Data Split: At the outset, we divide our dataset into two: one part for training our model and the other for testing it. This ensures that our model is evaluated on data it hasn’t seen before, offering a realistic representation of its real-world efficacy.
2. Ground Truth Comparison: By contrasting our model’s predictions with ground truth data — which are actual, verified data collected from direct observations or expert analyses — we can pinpoint the accuracy of our machine learning models.
3. Confusion Matrices and Classification Reports: These tools provide a deeper dive into how our models are performing. Not just overall accuracy, but insights into where they excel and where they might be faltering, be it false positives, false negatives, or other nuances.
4. Visual Inspections: Sometimes, there’s no substitute for the human eye. We visually inspect a subset of our model’s outputs against the actual satellite images, ensuring that our algorithms are aligned with real-world patterns.
5. Continuous Feedback Loop: Once models are deployed, the journey doesn’t end. We foster a feedback loop, refining and recalibrating models based on newer data and emergent patterns.
6. Benchmarking: By comparing our models with recognized benchmarks or state-of-the-art models, we ensure we’re at the forefront of technology and methodology in Earth Observation.

In conclusion, validating the accuracy of our ML models is a blend of science, cutting-edge technology, and human expertise. With these rigorous checks in place, we ensure that our insights from the skies are as accurate, actionable and grounded in reality as they can be.

The work done is part of the project co-financed by NCBR.

Teodor Niżyński

Machine Learning Engineer, Spottitt

Author: EARSC