Using Convolutional Neural Networks and Transfer Learning for Soil Organic Carbon and Plant Available Water Modeling in Data-Sparse Regions

Abstract Submission: Accurate prediction of soil properties, such as Soil Organic Carbon (SOC) and Plant Available Water (PAW), is crucial for sustainable land management and environmental conservation. However, in data-sparse regions, where field observations are limited, traditional modeling approaches often struggle to deliver reliable results. To address this challenge, we build upon our previous research, where SOC and PAW maps were generated for CONUS. These maps serve as a foundation for training Convolutional Neural Networks (CNNs) combined with transfer learning to model SOC and PAW across heterogeneous landscapes.
CNN is initially trained in data-rich regions using the previously generated SOC and PAW maps, leveraging spatial patterns between soil properties and environmental covariates, such as climate, topography, and land use. This pre-trained model is then fine-tuned using a limited number of field samples from data-sparse regions, allowing the model to adapt to local conditions while reducing the need for extensive field data collection.
Our methodology involves the extraction of spatially consistent patches from raster datasets, representing both soil properties and associated covariates. These patches are used to train the CNN, after which transfer, learning is employed to refine the model for application in regions with limited data. We anticipate this approach will significantly enhance the predictive capability of the model in data-scarce areas, while maintaining computational efficiency.
By building on the SOC and PAW maps generated in earlier research, our model benefits from a robust initial dataset, allowing for more accurate predictions even in challenging regions with limited field observations. This research highlights the potential of deep learning techniques to improve soil property predictions in data-constrained environments, providing a valuable tool for informed decision-making in land management and agricultural practices.

Learning Objectives/Expected Outcome (Optional) :