|Project Name||Stars||Downloads||Repos Using This||Packages Using This||Most Recent Commit||Total Releases||Latest Release||Open Issues||License||Language|
|Techniques||6,093||6 hours ago||1||apache-2.0|
|Techniques for deep learning with satellite & aerial imagery|
|Awesome Satellite Imagery Datasets||2,806||8 months ago||mit|
|🛰️ List of satellite image training datasets with annotations for computer vision and deep learning|
|Robosat||1,639||3 years ago||45||mit||Python|
|Semantic segmentation on aerial and satellite imagery. Extracts features such as: buildings, parking lots, roads, water, clouds|
|Eo Learn||984||4||9||2 days ago||29||June 13, 2022||10||mit||Python|
|Earth observation processing framework for machine learning in Python|
|Open Solution Mapping Challenge||351||2 years ago||53||mit||Python|
|Open solution to the Mapping Challenge :earth_americas:|
|Bv||203||6 years ago||Python|
|Quickly view satellite imagery, hyperspectral imagery, and machine learning image outputs directly in your iTerm2 terminal.|
|Geospatial Machine Learning||196||5 years ago||mit|
|A curated list of resources focused on Machine Learning in Geospatial Data Science.|
|Delta||169||7 months ago||2||apache-2.0||Python|
|Deep Learning for Satellite Imagery|
|Awesome_satellite_benchmark_datasets||154||2 months ago||1||Jupyter Notebook|
|Supplementary material for our paper "THERE IS NO DATA LIKE MORE DATA" is provided.|
|Skeyenet||108||4 months ago||10||mit||Jupyter Notebook|
|Road and Building Segmentation in Satellite Imagery|
Deep learning has transformed the way satellite and aerial images are analyzed and interpreted. These images pose unique challenges, such as large sizes and diverse object classes, which offer opportunities for deep learning researchers. This repository offers a comprehensive overview of various deep learning techniques for analyzing satellite and aerial imagery, including architectures, models, and algorithms for tasks such as classification, segmentation, and object detection. It serves as a valuable resource for researchers, practitioners, and anyone interested in the latest advances in deep learning and its impact on computer vision and remote sensing.
How to use this repository: if you know exactly what you are looking for (e.g. you have the paper name) you can
Control+F to search for it in this page. Note that material that is suitable for getting started with a topic is tagged with
BEGINNER, which can also be searched.
The UC merced dataset is a well known classification dataset.
Classification is a fundamental task in remote sensing data analysis, where the goal is to assign a semantic label to each image, such as 'urban', 'forest', 'agricultural land', etc. The process of assigning labels to an image is known as image-level classification. However, in some cases, a single image might contain multiple different land cover types, such as a forest with a river running through it, or a city with both residential and commercial areas. In these cases, image-level classification becomes more complex and involves assigning multiple labels to a single image. This can be accomplished using a combination of feature extraction and machine learning algorithms to accurately identify the different land cover types. It is important to note that image-level classification should not be confused with pixel-level classification, also known as semantic segmentation. While image-level classification assigns a single label to an entire image, semantic segmentation assigns a label to each individual pixel in an image, resulting in a highly detailed and accurate representation of the land cover types in an image. Read A brief introduction to satellite image classification with neural networks
(left) a satellite image and (right) the semantic classes in the image.
Image segmentation is a crucial step in image analysis and computer vision, with the goal of dividing an image into semantically meaningful segments or regions. The process of image segmentation assigns a class label to each pixel in an image, effectively transforming an image from a 2D grid of pixels into a 2D grid of pixels with assigned class labels. One common application of image segmentation is road or building segmentation, where the goal is to identify and separate roads and buildings from other features within an image. To accomplish this task, single class models are often trained to differentiate between roads and background, or buildings and background. These models are designed to recognize specific features, such as color, texture, and shape, that are characteristic of roads or buildings, and use this information to assign class labels to the pixels in an image. Another common application of image segmentation is land use or crop type classification, where the goal is to identify and map different land cover types within an image. In this case, multi-class models are typically used to recognize and differentiate between multiple classes within an image, such as forests, urban areas, and agricultural land. These models are capable of recognizing complex relationships between different land cover types, allowing for a more comprehensive understanding of the image content. Read A brief introduction to satellite image segmentation with neural networks. Note that many articles which refer to 'hyperspectral land classification' are often actually describing semantic segmentation. Image source
Extracting roads is challenging due to the occlusions caused by other objects and the complex traffic environment
In instance segmentation, each individual 'instance' of a segmented area is given a unique lable. For detection of very small objects this may a good approach, but it can struggle seperating individual objects that are closely spaced.
Image showing the suitability of rotated bounding boxes in remote sensing.
Object detection in remote sensing involves locating and surrounding objects of interest with bounding boxes. Due to the large size of remote sensing images and the fact that objects may only comprise a few pixels, object detection can be challenging in this context. The imbalance between the area of the objects to be detected and the background, combined with the potential for objects to be easily confused with random features in the background, further complicates the task. Object detection generally performs better on larger objects, but becomes increasingly difficult as the objects become smaller and more densely packed. The accuracy of object detection models can also degrade rapidly as image resolution decreases, which is why it is common to use high resolution imagery, such as 30cm RGB, for object detection in remote sensing. A unique characteristic of aerial images is that objects can be oriented in any direction. To effectively extract measurements of the length and width of an object, it can be crucial to use rotated bounding boxes that align with the orientation of the object. This approach enables more accurate and meaningful analysis of the objects within the image. Image source
Orinted bounding boxes (OBB) are polygons representing rotated rectangles. For datasets checkout DOTA & HRSC2016
Detecting the most noticeable or important object in a scene
Oil is stored in tanks at many points between extraction and sale, and the volume of oil in storage is an important economic indicator.
A variety of techniques can be used to count animals, including object detection and instance segmentation. For convenience they are all listed here:
When the object count, but not its shape is required, U-net can be used to treat this as an image-to-image translation problem.
Regression prediction of windspeed.
Regression in remote sensing involves predicting continuous variables such as wind speed, tree height, or soil moisture from an image. Both classical machine learning and deep learning approaches can be used to accomplish this task. Classical machine learning utilizes feature engineering to extract numerical values from the input data, which are then used as input for a regression algorithm like linear regression. On the other hand, deep learning typically employs a convolutional neural network (CNN) to process the image data, followed by a fully connected neural network (FCNN) for regression. The FCNN is trained to map the input image to the desired output, providing predictions for the continuous variables of interest. Image source
(left) False colour image and (right) a cloud & shadow mask.
Clouds are a major issue in remote sensing images as they can obscure the underlying ground features. This hinders the accuracy and effectiveness of remote sensing analysis, as the obscured regions cannot be properly interpreted. In order to address this challenge, various techniques have been developed to detect clouds in remote sensing images. Both classical algorithms and deep learning approaches can be employed for cloud detection. Classical algorithms typically use threshold-based techniques and hand-crafted features to identify cloud pixels. However, these techniques can be limited in their accuracy and are sensitive to changes in image appearance and cloud structure. On the other hand, deep learning approaches leverage the power of convolutional neural networks (CNNs) to accurately detect clouds in remote sensing images. These models are trained on large datasets of remote sensing images, allowing them to learn and generalize the unique features and patterns of clouds. The generated cloud mask can be used to identify the cloud pixels and eliminate them from further analysis or, alternatively, cloud inpainting techniques can be used to fill in the gaps left by the clouds. This approach helps to improve the accuracy of remote sensing analysis and provides a clearer view of the ground, even in the presence of clouds. Image adapted from this source
(left) Initial and (middle) after some development, with (right) the change highlighted.
Change detection is a vital component of remote sensing analysis, enabling the monitoring of landscape changes over time. This technique can be applied to identify a wide range of changes, including land use changes, urban development, coastal erosion, and deforestation. Change detection can be performed on a pair of images taken at different times, or by analyzing multiple images collected over a period of time. It is important to note that while change detection is primarily used to detect changes in the landscape, it can also be influenced by the presence of clouds and shadows. These dynamic elements can alter the appearance of the image, leading to false positives in change detection results. Therefore, it is essential to consider the impact of clouds and shadows on change detection analysis, and to employ appropriate methods to mitigate their influence. Image source
Prediction of the next image in a series.
The analysis of time series observations in remote sensing data has numerous applications, including enhancing the accuracy of classification models and forecasting future patterns and events. Image source. Note: since classifying crops and predicting crop yield are such prominent use case for time series data, these tasks have dedicated sections after this one.
(left) false colour image and (right) the crop map.
Crop classification in remote sensing is the identification and mapping of different crops in images or sequences of images. It aims to provide insight into the distribution and composition of crops in a specific area, with applications that include monitoring crop growth and evaluating crop damage. Both traditional machine learning methods, such as decision trees and support vector machines, and deep learning techniques, such as convolutional neural networks (CNNs), can be used to perform crop classification. The optimal method depends on the size and complexity of the dataset, the desired accuracy, and the available computational resources. However, the success of crop classification relies heavily on the quality and resolution of the input data, as well as the availability of labeled training data. Image source.
Wheat yield data. Blue vertical lines denote observation dates.
Crop yield is a crucial metric in agriculture, as it determines the productivity and profitability of a farm. It is defined as the amount of crops produced per unit area of land and is influenced by a range of factors including soil fertility, weather conditions, the type of crop grown, and pest and disease control. By utilizing time series of satellite images, it is possible to perform accurate crop type classification and take advantage of the seasonal variations specific to certain crops. This information can be used to optimize crop management practices and ultimately improve crop yield. However, to achieve accurate results, it is essential to consider the quality and resolution of the input data, as well as the availability of labeled training data. Appropriate pre-processing and feature extraction techniques must also be employed. Image source.
COVID-19 impacts on human and economic activities.
The traditional approach of collecting economic data through ground surveys is a time-consuming and resource-intensive process. However, advancements in satellite technology and machine learning offer an alternative solution. By utilizing satellite imagery and applying machine learning algorithms, it is possible to obtain accurate and current information on economic activity with greater efficiency. This shift towards satellite imagery-based forecasting not only provides cost savings but also offers a wider and more comprehensive perspective of economic activity. As a result, it is poised to become a valuable asset for both policymakers and businesses. Image source.
Detecting buildings destroyed in a disaster.
Remote sensing images are used in disaster response to identify and assess damage to an area. This imagery can be used to detect buildings that are damaged or destroyed, identify roads and road networks that are blocked, determine the size and shape of a disaster area, and identify areas that are at risk of flooding. Remote sensing images can also be used to detect and monitor the spread of forest fires and monitor vegetation health. Also checkout the sections on change detection and water/fire/building segmentation. Image source.
Super resolution using multiple low resolution images as input.
Super-resolution is a technique aimed at improving the resolution of an imaging system. This process can be applied prior to other image processing steps to increase the visibility of small objects or boundaries. Despite its potential benefits, the use of super-resolution is controversial due to the possibility of introducing artifacts that could be mistaken for real features. Super-resolution techniques are broadly categorized into two groups: single image super-resolution (SISR) and multi-image super-resolution (MISR). SISR focuses on enhancing the resolution of a single image, while MISR utilizes multiple images of the same scene to create a high-resolution output. Each approach has its own advantages and limitations, and the choice of method depends on the specific application and desired outcome. Image source.
Note that nearly all the MISR publications resulted from the PROBA-V Super Resolution competition
Pansharpening example with a resolution difference of factor 4.
Pansharpening is a data fusion method that merges the high spatial detail from a high-resolution panchromatic image with the rich spectral information from a lower-resolution multispectral image. The result is a single, high-resolution color image that retains both the sharpness of the panchromatic band and the color information of the multispectral bands. This process enhances the spatial resolution while preserving the spectral qualities of the original images. Image source
(left) Sentinel-1 SAR input, (middle) translated to RGB and (right) Sentinel-2 true RGB image for comparison.
Image-to-image translation is a crucial aspect of computer vision that utilizes machine learning models to transform an input image into a new, distinct output image. In the field of remote sensing, it plays a significant role in bridging the gap between different imaging domains, such as converting Synthetic Aperture Radar (SAR) images into RGB (Red Green Blue) images. This technology has a wide range of applications, including improving image quality, filling in missing information, and facilitating cross-domain image analysis and comparison. By leveraging deep learning algorithms, image-to-image translation has become a powerful tool in the arsenal of remote sensing researchers and practitioners. Image source
Illustration of a fusion workflow.
Data fusion is a technique for combining information from different sources such as Synthetic Aperture Radar (SAR), optical imagery, and non-imagery data such as Internet of Things (IoT) sensor data. The integration of diverse data sources enables data fusion to overcome the limitations of individual sources, leading to the creation of models that are more accurate and informative than those constructed from a single source. Image source
Example generated images using a GAN.
Generative Adversarial Networks (GANs) are a type of deep learning architecture that leverages the power of competition between two neural networks. The objective of a GAN is to generate new, synthetic data that appears similar to real-world data. This is achieved by training the two networks, the generator and the discriminator, in a zero-sum game, where the generator attempts to produce data that is indistinguishable from the real data, while the discriminator tries to distinguish between the generated data and the real data. In the field of remote sensing, GANs have found numerous applications, particularly in generating synthetic data. This synthetic data can be used for a wide range of purposes, including data augmentation, data imbalance correction, and filling in missing or corrupted data. By generating realistic synthetic data, GANs can improve the performance of remote sensing algorithms and models, leading to more accurate and reliable results. Additionally, GANs can also be used for various other tasks in remote sensing, such as super-resolution, denoising, and inpainting. Image source
Example of using an autoencoder to create a low dimensional representation of hyperspectral data.
Autoencoders are a type of neural network that aim to simplify the representation of input data by compressing it into a lower dimensional form. This is achieved through a two-step process of encoding and decoding, where the encoding step compresses the data into a lower dimensional representation, and the decoding step restores the data back to its original form. The goal of this process is to reduce the data's dimensionality, making it easier to store and process, while retaining the essential information. Dimensionality reduction, as the name suggests, refers to the process of reducing the number of dimensions in a dataset. This can be achieved through various techniques such as principal component analysis (PCA) or singular value decomposition (SVD). Autoencoders are one type of neural network that can be used for dimensionality reduction. In the field of computer vision, image embeddings are vector representations of images that capture the most important features of the image. These embeddings can then be used to perform similarity searches, where images are compared based on their features to find similar images. This process can be used in a variety of applications, such as image retrieval, where images are searched based on certain criteria like color, texture, or shape. It can also be used to identify duplicate images in a dataset. Image source
Illustration of the remote sensing image retrieval process.
Image retrieval is the task of retrieving images from a collection that are similar to a query image. Image retrieval plays a vital role in remote sensing by enabling the efficient and effective search for relevant images from large image archives, and by providing a way to quantify changes in the environment over time. Image source
Example captioned image.
Image Captioning is the task of automatically generating a textual description of an image. In remote sensing, image captioning can be used to automatically generate captions for satellite or aerial images, which can be useful for a variety of purposes, such as image search and retrieval, data cataloging, and data dissemination. The generated captions can provide valuable information about the content of the images, including the location, the type of terrain or objects present, and the weather conditions, among others. This information can be used to quickly and easily understand the content of the images, without having to manually examine each image. Image source
Visual Question Answering (VQA) is the task of automatically answering a natural language question about an image. In remote sensing, VQA enables users to interact with the images and retrieve information using natural language questions. For example, a user could ask a VQA system questions such as "What is the type of land cover in this area?", "What is the dominant crop in this region?" or "What is the size of the city in this image?". The system would then analyze the image and generate an answer based on its understanding of the image content.
Mixed data learning is the process of learning from datasets that may contain an mix of images, textual and numeric data. Mixed data learning can help improve the accuracy of models by allowing them to learn from multiple sources at once and use more sophisticated methods to identify patterns and correlations.
This is a class of techniques which attempt to make predictions for classes with few, one or even zero examples provided during training. In zero shot learning (ZSL) the model is assisted by the provision of auxiliary information which typically consists of descriptions/semantic attributes/word embeddings for both the seen and unseen classes at train time (ref). These approaches are particularly relevant to remote sensing, where there may be many examples of common classes, but few or even zero examples for other classes of interest.
Self-supervised, unsupervised & contrastive learning are all methods of machine learning that use unlabeled data to train algorithms. Self-supervised learning uses labeled data to create an artificial supervisor, while unsupervised learning uses only the data itself to identify patterns and similarities. Contrastive learning uses pairs of data points to learn representations of data, usually for classification tasks.
Weakly & semi-supervised learning are two methods of machine learning that use both labeled and unlabeled data for training. Weakly supervised learning uses weakly labeled data, which may be incomplete or inaccurate, while semi-supervised learning uses both labeled and unlabeled data. Weakly supervised learning is typically used in situations where labeled data is scarce and unlabeled data is abundant. Semi-supervised learning is typically used in situations where labeled data is abundant but also contains some noise or errors. Both techniques can be used to improve the accuracy of machine learning models by making use of additional data sources.
Supervised deep learning techniques typically require a huge number of annotated/labelled examples to provide a training dataset. However labelling at scale take significant time, expertise and resources. Active learning techniques aim to reduce the total amount of annotation that needs to be performed by selecting the most useful images to label from a large pool of unlabelled images, thus reducing the time to generate useful training datasets. These processes may be referred to as Human-in-the-Loop Machine Learning
Federated learning is an approach to distributed machine learning where a central processor coordinates the training of an individual model in each of its clients. It is a type of distributed ML which means that the data is distributed among different devices or locations and the model is trained on all of them. The central processor aggregates the model updates from all the clients and then sends the global model parameters back to the clients. This is done to protect the privacy of data, as the data remains on the local device and only the global model parameters are shared with the central processor. This technique can be used to train models with large datasets that cannot be stored in a single device, as well as to enable certain privacy-preserving applications.
Vision transformers are state-of-the-art models for vision tasks such as image classification and object detection. They differ from CNNs as they use self-attention instead of convolution to learn global relations between all pixels in the image. Vision transformers employ a transformer encoder architecture, composed of multi-layer blocks with multi-head self-attention and feed-forward layers, enabling the capture of rich contextual information for more accurate predictions.
Efforts to detect falsified images & deepfakes
Image registration is the process of registering one or more images onto another (typically well georeferenced) image. Traditionally this is performed manually by identifying control points (tie-points) in the images, for example using QGIS. This section lists approaches which mostly aim to automate this manual process. There is some overlap with the data fusion section but the distinction I make is that image registration is performed as a prerequisite to downstream processes which will use the registered data as an input.
Measure surface contours & locate 3D points in space from 2D images. NeRF stands for Neural Radiance Fields and is the term used in deep learning communities to describe a model that generates views of complex 3D scenes based on a partial set of 2D images
Thermal infrared remote sensing is a technique used to detect and measure thermal radiation emitted from the Earth’s surface. This technique can be used to measure the temperature of the ground and any objects on it and can detect the presence of different materials. Thermal infrared remote sensing is used to assess land cover, detect land-use changes, and monitor urban heat islands, as well as to measure the temperature of the ground during nighttime or in areas of limited visibility.
SAR (synthetic aperture radar) is used to detect and measure the properties of objects and surfaces on the Earth's surface. SAR can be used to detect changes in terrain, features, and objects over time, as well as to measure the size, shape, and composition of objects and surfaces. SAR can also be used to measure moisture levels in soil and vegetation, or to detect and monitor changes in land use.
Normalized Difference Vegetation Index (NDVI) is an index used to measure the amount of healthy vegetation in a given area. It is calculated by taking the difference between the near-infrared (NIR) and red (red) bands of a satellite image, and dividing by the sum of the two bands. NDVI can be used to identify areas of healthy vegetation and to assess the health of vegetation in a given area.
ndvi = np.true_divide((ir - r), (ir + r))but challenging due to the size of the imagery
Image quality describes the degree of accuracy with which an image can represent the original object. Image quality is typically measured by the amount of detail, sharpness, and contrast that an image contains. Factors that contribute to image quality include the resolution, format, and compression of the image.
Training data can be hard to acquire, particularly for rare events such as change detection after disasters, or imagery of rare classes of objects. In these situations, generating synthetic training data might be the only option. This has become quite sophisticated, with 3D models being use with open source games engines such as Unreal.