Fine-tuning land cover models efficiently with transfer learning: a tool and case studies

Abstract:

Land cover mapping is not a single problem, but a distribution of problems over dimensions such as: source imagery, spatial resolution, temporal resolution, land cover class definitions, and final product presentation. As such, machine learning approaches that can solve one instance of a land cover mapping problem may not transfer easily to other instances. Implementing machine learning approaches requires large amounts of labeled data, yet it is not scalable to collect a new labeled dataset and train a model from scratch for every land cover mapping project instance or when requirements evolve. To alleviate this problem, we are creating a browser-based tool that allows users to fine-tune existing land cover models to fit the parameters of their own problem instances (i.e. facilitates transfer learning). We provide a brief overview of the tool, and 4 case studies of using the tool in the Microsoft AI For Good Research Lab:

Our tool is under active development, and currently provides the following features: visualization and interactive segmentation of raster-based data, point based fine-tuning of the underlying segmentation model, checkpointing, and downloading of model prediction over arbitrary polygons. To use the tool, users must provide: a raster data source, an accompanying basemap (RGB visualization of the raster), and an initial model. See https://github.com/microsoft/landcover for more details.

We have used this tool in four case studies:

1. Mapping land cover in Ho Chi Minh City from Sentinel 2 imagery. Starting from an unsupervised baseline model, we create a land cover model by providing few point examples of the classes required by our partners.
2. Mapping solar panels in India from Sentinel 2 imagery. The goal of this project is to segment solar installations, but only point labels are provided. We start from an unsupervised baseline that we use to create a large-scale dataset of noisy segmentation labels through the tool.
3. Mapping land cover in Colombia from Landsat 8 imagery. Starting from a model trained using historical labels, we use the tool to tune the model for applications to present-day data and newly relevant classes.
4. Mapping roof materials in flood-prone regions in India from Maxar imagery. We start with a model trained on limited data samples and use the tool to tune it to new regions.