Minerals play a critical role in the advanced energy technologies necessary for decarbonization, but characterizing mineral deposits hidden underground remains costly and challenging. Inspired by recent progress in generative modeling, we develop a learning method which infers the locations of minerals by masking and infilling geospatial maps of resource availability. We demonstrate this technique using mineral data for the conterminous United States, and train performant models, with the best achieving Dice coefficients of 0.31 ± 0.01 and recalls of 0.22 ± 0.02 on test data at 1×1 sq mi spatial resolution. One major advantage of our approach is that it can easily incorporate auxiliary data sources for prediction which may be more abundant than mineral data. We highlight the capabilities of our model by adding input layers derived from geophysical sources, along with a nation-wide ground survey of soils originally intended for agronomic purposes. We find that employing such auxiliary features can improve inference performance, while also enabling model evaluation in regions with no recorded minerals.

Accurate crop type maps are an essential source of information for monitoring yield progress at scale, projecting global crop production, and planning effective policies. To date, however, crop type maps remain challenging to create in low- and middle-income countries due to a lack of ground truth labels for training machine learning models. Field surveys are the gold standard in terms of accuracy but require an often-prohibitively large amount of time, money, and statistical capacity. In recent years, street-level imagery, such as Google Street View, KartaView, and Mapillary, has become available around the world. Such imagery contains rich information about crop types grown at particular locations and times. In this work, we develop an automated system to generate crop type ground references using deep learning and Google Street View imagery. The method efficiently curates a set of street-view images containing crop fields, trains a model to predict crop types using either weakly-labeled images from disparate out-of-domain sources or zero-shot labeled street view images with GPT-4V, and combines the predicted labels with remote sensing time series to create a wall-to-wall crop type map. We show that, in Thailand, the resulting country-wide map of rice, cassava, maize, and sugarcane achieves an accuracy of 93%. We publicly release the first-ever crop type map for all of Thailand 2022 at 10m-resolution with no gaps. To our knowledge, this is the first time a 10m-resolution, multi-crop map has been created for any smallholder country. As the availability of roadside imagery expands, our pipeline provides a way to map crop types at scale around the globe, especially in underserved smallholder regions.

Learning to predict accurately from a few data samples is a central challenge in modern data-hungry machine learning. On natural images, human vision typically outperforms deep learning approaches on few-shot learning. However, we hypothesize that aerial and satellite images are more challenging to the human eye. This applies particularly when the image resolution is comparatively low, as with the 10m ground sampling distance of Sentinel-2. In this study, we benchmark model-agnostic meta-learning (MAML) algorithms against human participants on few-shot land cover classification with Sentinel-2 imagery on the Sen12MS dataset. We find that categorization of land cover from globally distributed regions is a difficult task for the participants, who classified the given images less accurately than the MAML-trained model and with a highly variable success rate. This suggest that hand-labeling land cover directly on Sentinel-2 imagery is not optimal when tackling a new land cover classification problem. Labeling only a few images and employing a trained meta-learning model to this task may lead to more accurate and consistent solutions compared to hand labeling by multiple individuals.

The open Landsat archive provides a consistent view of the Earth's surface for much longer than most currently available agricultural datasets throughout the world. Here we present a summary of recent work to extend pixel-level (30 m) maps of corn and soy areas and yields back to 1999 across the entire Corn Belt of the United States. We find consistent performance back in time, as judged by comparison with county level statistics. The approaches presented here can be readily extended to other regions and to incorporate other sensors.

Geospatial analysis lacks methods like the word vector representations and pre-trained networks that significantly boost performance across a wide range of natural language and computer vision tasks. To fill this gap, we introduce Tile2Vec, an unsupervised representation learning algorithm that extends the distributional hypothesis from natural language — words appearing in similar contexts tend to have similar meanings — to spatially distributed data. We demonstrate empirically that Tile2Vec learns semantically meaningful representations for both image and non-image datasets. Our learned representations significantly improve performance in downstream classification tasks and, similarly to word vectors, allow visual analogies to be obtained via simple arithmetic in the latent space.