Dissertation Defense - Hui Li - Department of Geography and Geoinformation Science

PhD Candidate: Hui Li

PhD of Science, Earth Systems and Geoinformation Science

Department of Geography and Geoinformation Science

Dissertation Title: In-Season Crop Mapping of Sugarcane, Corn, and Soybean Across the U.S., Canada, and Brazil Using Transfer Learning of Time-Invariant Phenological Features

Dissertation Director: Liping Di, Ph.D. (GMU, GGS, CSISS)

Committee Members:

Dr. Ruixin Yang  (GMU, Geography and Geoinformation Science)

Dr. John J. Qu  (GMU, Geography and Geoinformation Science)

Dr. Daniel Q. Tong (GMU, Atmospheric, Oceanic & Earth Sciences Department)

Abstract:

Sugarcane, corn, and soybean are significant crop types for food and industrial material worldwide. Timely sugarcane, corn, and soybean cultivation mapping will provide important spatial-temporal information to support multiple decision-making. However, traditional time-series crop-type mapping with machine learning and remote sensing encounters challenges due to limited training data and heterogeneous planting dates. We developed a phenology-based transfer learning algorithm to generate in-season sugarcane, corn, and soybean maps across the U.S., Canada, and Brazil using only historical U.S. training samples. At the core of this algorithm is Linear Cosine Regression (LCR), which extracts time-invariant phenological features from time-series multispectral satellite data. High-confidence labels were derived from the burning fields and the U.S. Cropland Data Layer (CDL), incorporating phenology features to train a transferable classification model for identifying these crops across different times and locations.

Our topic consists of three key studies. We first validated our methodology by mapping in-season sugarcane across different regions and times within the U.S. We extracted high-confidence training labels from burning sugarcane fields using Sentinel-2 imagery in Palm Beach County, Florida (2021). These labels' time-invariant phenology features were used to train a One-Class Support Vector Machine (OCSVM) classifier, generating sugarcane maps from June to November 2022 for Palm Beach County and Lafourche Parish, Louisiana. Validation against ground truth data showed that the September 2022 sugarcane maps achieved an overall accuracy of 0.95, outperforming the 2022 CDL. These findings were published in the IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS) in 2024.

Secondly, in-season sugarcane mapping across the U.S. and Brazil.  We leveraged historical CDL data to extract high-confidence sugarcane and other crop-type labels. These labels’ time-invariant phenology features were used to develop transferable Random Forest classifiers. In the U.S., four-month (June–September 2023) sugarcane maps were produced for Lafourche Parish, Louisiana, achieving an overall accuracy of 0.938, surpassing CDL 2023’s 0.894. Five-month (March–July 2022) sugarcane maps were generated for São Paulo State of Brazil, achieving an overall accuracy of 88.1% in May 2022, as validated by ground truth data. These results are currently under review in the ISPRS Journal of Photogrammetry and Remote Sensing.

Thirdly, we applied the algorithm to map in-season corn and soybean across the U.S., Canada, and Brazil. Similarly, high-confidence crop-type labels were selected from historical U.S. CDL data.  Their time-invariant phenology features built multiple transferable random forest classifiers for these three nations. In the U.S., three monthly in-season corn-soybean maps (June–August 2023) for the Corn Belt were generated, achieving an overall accuracy of 0.963 in July, closely matching CDL 2023’s 0.981. In Canada, two monthly growing-season maps (July–August 2023) for Ontario were produced, achieving a relative overall accuracy of 0.835 in August, compared to the Annual Crop Inventory (ACI). In Brazil, two monthly corn-soybean maps (November–December 2021) were produced for Rio Grande do Sul and Mato Grosso do Sul States, achieving a maximum overall accuracy of 0.811 in December 2021, validated by ground truth data.

These studies successfully map three major crops across three countries, overcoming challenges associated with limited training data and heterogeneous planting dates. Our approach demonstrates the effectiveness, stability, generalizability, and adaptability of time-invariant phenological features for in-season crop types mapping across diverse periods and geographic regions, without requiring local and current training samples. We present an innovative in-season crop-type mapping technique that enables reliable crop monitoring across distant locations and periods, offering valuable insights for agricultural and environmental decision-making.