Introducing Rahebeh Abedi, a Ph.D. student at Clark University’s School of Geography and our Summer Intern. Rahebeh is a recipient of the Graduate Research Fellowship from the Edna Bailey Sussman Fund, funded to work on crop type modeling with Radiant Earth. Her research focuses on the prediction of crop types in different regions of the United States using computer models and ground truth labels provided by the USDA’s Cropland Data Layer. Rahebeh will focus her research on areas with incomplete crop data to determine the minimum sample size required for accurate results. In this interview, Rahebeh discusses her research project and research interests.
Congratulations on receiving the research grant! You have a background in GIS and geomatics engineering, with a focus on environmental changes, spatial analysis, data mining, and machine learning. What inspired you to pursue this field of study?
I’ve always been fascinated by the dynamic nature of our environment and how it affects various aspects of our lives. This fascination, combined with my interest in technology and problem-solving, led me to explore the field of Remote Sensing and GIS. As I delved deeper into my studies, I realized the immense potential of spatial analysis, data mining, and machine learning techniques in understanding and addressing environmental challenges. This realization inspired me to pursue research in this interdisciplinary field, where I could combine my passion for the environment with advanced technological tools to make a meaningful impact.
Let’s talk about your upcoming research on crop type mapping. What motivated you to choose this focus, and what are your goals and expectations for this study?
Crop type mapping is an important task in estimating crop yields and ensuring food security. While unsupervised and semi-supervised models have helped reduce the reliance on costly ground truth data collection, challenges persist in this field. Evaluating predictions remains difficult, as does considering the sequential nature of data in relation to crop phenology. Moreover, the presence of clouds significantly affects the accuracy of results.
In my intended research, I am particularly interested in exploring how generative adversarial networks (GANs) can help to fill the gaps caused by clouds in time series data. GANs have shown promise in generating synthetic data that can potentially compensate for missing or obscured information due to cloud cover. By leveraging the power of GANs, I aim to enhance the accuracy and reliability of crop type mapping, especially in situations where cloud coverage hinders the analysis of temporal data.
Your research is specifically targeting areas where the Cropland Data Layer information is incomplete. What potential insights or benefits do you anticipate gaining from analyzing these areas?
The stability and large size of croplands in the United States make it an ideal area for research. The ability to generate arbitrary cloud cover facilitates the establishment of benchmark models and enables comprehensive time series analysis. These benchmark models can be used to fill the gaps in CDL and have the potential to map crop types in other regions worldwide with similar challenges of cloud cover during the growing season, which poses difficulties in generating accurate crop type maps.
Additionally, could you elaborate on the significance of mapping crop types over multiple years? What valuable information can be derived from such knowledge?
Mapping crop types over multiple years yields valuable insights into the transition of crop type patterns and the long-term temporal agricultural dynamic. This information serves as a foundation for establishing many-to-many model architectures that utilize temporal data as input and generate crop type maps, taking into account the crop rotations observed over the years. By incorporating temporal information, these models can capture the dynamic nature of agricultural systems and provide a comprehensive understanding of the changing landscape of crop types over time.
What challenges are associated with generating accurate and comprehensive crop type maps using satellite imagery and ground truth data?
Variations in crop phenology (the timing of crop growth stages) and intercropping (multiple crops grown together), spatial heterogeneity, availability and quality of ground truth data, selection and optimization of classification algorithms, data integration, and preprocessing, rapid land cover changes, and seasonality are among the key challenges. To address these challenges, it is essential to devise strategies that can consider the specific characteristics and limitations of available labels.
What are some of the challenges and opportunities of using newer models for mapping?
Newer models are increasingly focused on foundation models that are trained on large amounts of unlabeled data. These models exhibit invariance to specific tasks but demand substantial computational resources and exhibit greater complexity compared to popular current models based on convolutional neural networks (CNNs). Additionally, alternative training strategies have emerged to address the challenges posed by limited labeled data. These strategies include semisupervised and weakly supervised methods, which typically involve the use of multiple networks and complex loss functions to leverage both labeled and unlabeled samples. These approaches aim to maximize the utilization of available data while minimizing the need for extensive labeling efforts.
Lastly, considering the pressing environmental issues the world faces today, what do you believe to be the most significant, and how do you envision your research contributing to a solution?
Advanced models like GANs can address the limitations caused by cloud cover in crop type mapping. By generating synthetic data, GANs effectively fill the information gaps and improve the completeness and accuracy of crop type maps. More accurate crop type maps enhance our understanding of agricultural landscapes and provide valuable insights and tools for promoting sustainable agricultural practices.