ABSTRACT

Crop production is a major industry in the United States (U.S.). In 2021, the U.S. grain export accounted for over 40% share of international grain trade. Millions of U.S. farmers depend on international market for living and prosperity. However, the U.S. grain export is not only facing tough competition from other export countries, but also impacted by grain yield in import countries. In order to gain the competitive edge, stakeholders need to know as early as possible where and how many acres each type of crops that have been planted in a growing season around the world so that yield can be estimated, production and demand balance can be assessed, and grain prices can be predicted. This requires generating in-season crop maps of both U.S. and foreign countries. The classic method to generate in-season crop maps needs a large amount of verified information on crops (i.e., ground truths) to train algorithms for classifying in-season satellite remote sensing images. However, it is difficult or even impossible to obtain ground truths in foreign countries, particularly in early season. This study proposes to develop a spatiotemporally transferable machine-learning algorithm which will be trained with U.S. data and applied to in-season satellite remote sensing images of foreign countries for creating the in-season crop maps of the countries. Success of this project will make the in-season crop mapping of foreign countries possible. The project will significantly enhance the competitiveness and profitability of U.S. agriculture, increase the food security of the world, and potentially bring billions-of-dollars economic benefits to U.S. farmers.

Satellite remote sensing with ground truth tagging is the current practice for crop mapping. However, it suffers from two problems: 1) Unavailability of ground truth in foreign countries; 2) Spatiotemporal intransferability of trained classifiers. This study will design spatiotemporally transferable learning algorithm and temporal learning strategy that would maximally transfer label data and models from U.S. to foreign countries. The proposed method utilizes adversarial training and contrastive learning. Through this two-player game, the feature extractor produces domain-invariant features. A classifier trained on this domain-invariant representation can transfer its model to a new domain because the target features match those seen during training, thus bridging the gap between times and locations. The U.S. trained algorithm will be tested in Canada and Brazil to demonstrate its cross-country and cross-hemisphere transferability. Scientifically this project will advance landcover science in in-season crop mapping by offering a novel method of transfer learning, advance machine learning in unsupervised domain adaptation across both space and time, and offer new methods to derive spatiotemporally invariant features from time-series remote sensing images. Socioeconomically this project will enhance competitiveness and profitability of U.S. agriculture, increase food security of the world, and potentially bring billions-of-dollars benefits to U.S. farmers.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

Please report errors in award information by writing to: awardsearch@nsf.gov.