AI for landcover classification
Course Overview
Embark on the "AI for Landcover Classification" journey, exploring core principles and the importance of remote sensing images for landcover insights. Delve into the essence of remote sensing, learning to observe from a distance. Explore benchmark datasets, honing classification skills, and master contrastive self-supervised learning for enhanced accuracy with unlabeled data. Utilize Google Earth Engine to procure tailored satellite images. Navigate transformative techniques for augmenting views in remote sensing. This course focuses on landcover classification, providing expertise for AI applications and a comprehensive exploration of geospatial datasets. In the 2nd Part, delve into Activation Space Occlusion Sensitivities (ASOS) - a novel explainable machine learning approach. Witness practical ASOS application, enhancing transparency and interpretability, and understanding the activation space. Gain a comprehensive understanding of ASOS for model interpretability.
Details
Lecturer
Overview
Embark on the "AI for Landcover Classification" journey, exploring core principles and the importance of remote sensing images for landcover insights. Delve into the essence of remote sensing, learning to observe from a distance. Explore benchmark datasets, honing classification skills, and master contrastive self-supervised learning for enhanced accuracy with unlabeled data. Utilize Google Earth Engine to procure tailored satellite images. Navigate transformative techniques for augmenting views in remote sensing. This course focuses on landcover classification, providing expertise for AI applications and a comprehensive exploration of geospatial datasets. In the 2nd Part, delve into Activation Space Occlusion Sensitivities (ASOS) - a novel explainable machine learning approach. Witness practical ASOS application, enhancing transparency and interpretability, and understanding the activation space. Gain a comprehensive understanding of ASOS for model interpretability.
Introduction to Remote Sensing
This video gives an introduction about remote sensing images and how they provide meaningful landcover information.
Introduction to Land Cover Classification
In this video, you will learn about Landcover classification with remote sensing and how you need to proceed to perform Landcover classification
Benchmark Dataset
Benchmark datasets are already processed and published which is essentially used to compare different models.
Contrastive self-supervised Learning
Self-supervised learning is a method of machine learning where model learns from unlabeled data. In this video you will learn about contrastive learning, a type of self-supervised learning.
Introduction to GEE via Code Snippets Walkthrough
Google Earth Engine is a platform tailored for remote sensing researchers. It contains lot of images and allows users to obtain images of your location at your designed timestamp from your desired satellite mission.
Introduction to Atmospheric Transformation
In this video, you will learn about transformations used to obtain augmented views of an image in the domain of remote sensing images.
Location-Based Labels
In this video, we showcase an example of location-based labels that could be employed instead of the pseudo-labels typically utilized in contrastive self-supervised learning.
The Basic Idea of ASOS
The Basic Idea of “Activation Space Occlusion Sensitivities” or short “ASOS”
The Neural Network Architecture and Training
ASOS does only work for specific neural network architectures. The combination of an encoder-decoder network and a classifier is needed. In this video, a U-Net encoder-decoder network is used as example.
Defining Occlusions Using the Activation Space
The video demonstrates how we occlude similar activations and determine their influence.
Interpretability of the Activation Space
The activation space makes the neural network more transparent and interpretable.
The Advantages of ASOS
“Activation Space Occlusion Sensitivities” or short “ASOS” has many advantages compared to other methods for saliency maps.
Datasets - AnthroProtect
The AnthroProtect dataset can be used for simple classification tasks. But it is especially interesting, to combine the classification with explainable machine learning - for example, attribution methods.
Datasets - MapInWild
The MapInWild dataset is a multi-modal large-scale benchmark dataset. It is designed for the task of wilderness mapping using remote sensing data.
Datasets - TorchGeo
TorchGeo is a Python library to integrate geospatial data into the PyTorch ecosystem. In this video, I want to give you an overview, explain the most important implementations, and show how to use them in a standard workflow.
Datasets - BigEarthNet
BigEarthNet is a large-scale multi-modal and multi-label dataset to support deep learning studies in remote sensing image classification.
Datasets - Eurosat
The EuroSAT dataset provides Sentinel-2 imagery to tackle the challenge of land use and land cover classification.
Spatial Data Split
To validate and test a machine learning model, it is essential to have your data split into three subsets. In this video you will learn how to split your remote sensing data in a better way.
