Cross Domain Early Crop Mapping based on Time-series Remote Sensing Data

This dissertation presents a comprehensive study of advanced remote sensing
methodologies for early crop mapping, employing innovative machine learning tech-
niques to address significant challenges in agricultural monitoring. The research
encapsulates three main studies: ECMDCM (Early Crop Mapping using Dynamic
Clustering Method), CropSTGAN (Crop Spectral-temporal Generative Adversarial
Neural Network), and MultiCropGAN (Multiple Crop Mapping Generative Adversar-
ial Neural Network), each contributing uniquely to the field of precision agriculture.
The ECMDCM introduces a novel dynamic clustering approach using time-
series NDVI and EVI data to enhance the accuracy of early crop mapping across
the continental United States. By optimizing ecoregion delineations through the
elbow and silhouette methods and employing Kmeans++ for clustering, this method
demonstrates significant improvements over traditional static clustering techniques,
offering a more dynamic and precise mapping of crop types.
The CropSTGAN framework addresses the challenges of cross-domain variability
in remote sensing-based crop mapping. It incorporates a domain mapper that
effectively aligns temporal and spectral features across different geographic and
temporal scales, facilitating robust model performance even in the presence of
significant data distribution discrepancies. This framework has been validated
across diverse regions and years, showcasing superior accuracy and adaptability in
comparison to conventional approaches.
Lastly, the MultiCropGAN framework is developed to tackle domain shift and
label space discrepancies, which are prevalent in global agricultural settings. By
incorporating identity losses into the generator’s loss function, MultiCropGAN ensures
the preservation of essential characteristics in the data, enhancing the authenticity
and accuracy of crop type classification. Extensive testing across various North
American regions highlights its effectiveness, particularly in handling divergent label
spaces, thereby improving the reliability and applicability of crop mapping techniques.
Together, these studies not only demonstrate the potential of generative adver-
sarial networks and dynamic clustering in remote sensing but also pave the way for
future innovations in agricultural monitoring. This thesis aims to contribute to the
enhancement of global food security strategies through improved crop monitoring and
management, underlining the critical role of advanced remote sensing technologies in
the future of agriculture.