Location: Dale Bumpers Small Farms Research CenterTitle: Towards a dynamic soil survey: Identifying and delineating soil horizons in-situ using deep learning
|JIANG, ZHUODONG - Shenyang Agricultural University|
|ZHANG, CHUN-LIANG - Shenyang Agricultural University|
|BRYE, KRISTOPHER - University Of Arkansas|
|WEINDORF, DAVID - Central Michigan University|
|SUN, ZHONGXIU - Shenyang Agricultural University|
|SUN, FUJUN - Shenyang Agricultural University|
|WANG, QUIBING - Shenyang Agricultural University|
Submitted to: Geoderma
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 7/5/2021
Publication Date: 7/15/2021
Citation: Jiang, Z., Owens, P.R., Zhang, C., Brye, K.R., Weindorf, D.C., Adhikari, K., Sun, Z., Sun, F., Wang, Q. 2021. Towards a dynamic soil survey: Identifying and delineating soil horizons in-situ using deep learning. Geoderma. 401(2021)115341. https://doi.org/10.1016/j.geoderma.2021.115341.
Interpretive Summary: Rapid information collection is changing scientific data collection. With the number of smartphones increasing, there is an opportunity to take advantage of Apps that allow for data collection and identification. This research focused on utilizing deep learning methods to identify soil properties which can allow users to take a photo and have information delivered about the vertical soil section. The testing data showed promise with greater than 70% accuracy on identifying soil horizons. This process was taken to develop into an Android Smartphone App that will allow the user to take a photo and get the soil horizon designation. These rapid data collection tools allow people mapping soils to collect and analyze the data immediately while working in the field.
Technical Abstract: Soil horizons are the basic morphologic indicators of soil formation and environmental relationship to pedogensis. Therefore, quantitively and accurately delineated soil horizons is important for soil survey and increased knowledge of soil formation. The objective of this study was to develop a deep learning-based soil profile imaging method for identifying and delineating soil horizons. A total of 160 soil profile images were collected from north China. The 160 profile images were augmented to 2400 images for the model building by the data augmentation procedure. The augmented profile image dataset was divided into a 70% training dataset, a 15% validation dataset, and a 15% test dataset. The training and validation datasets were imported into a nested U-net network for model building. The proposed deep learning (DL) model could classify the soil profiles into A, B, and C horizons. The performance of the DL model was evaluated on a 70% training dataset with a mean pixel accuracy (MPA) of 0.86, on a 15% validation dataset with an MPA of 0.82, and on a 15% test dataset with an MPA of 0.83. Results showed that the DL model was a reasonably accurate method for identifying and delineating soil master horizons from the profile images. Based on the proposed model, a smartphone application was developed to assist in capturing information in the field. The smartphone application was developed for utilization on Android system post version 5.0 with a response time of less than 5 s. This study demonstrated the developed model and smartphone application may provide a quick, simple, and accurate tool for identifying and delineating soil horizons to assist the research, soil survey and teaching activities.