Enhancing predictive mapping of soil carbon by incorporating vegetation growth dynamic information via deep learning

<p>The spatial distribution of soil organic carbon (SOC) serves as critical geographic information for assessing ecosystem services and guides land management for migrating carbon emissions. Digital mapping of SOC is challenging due to the complex relationships between the soil and its environmental conditions. Except for the well-known topography and climate environmental covariates, the aboveground vegetation growth, which interacts with belowground soil carbon, in&#64258;uences SOC signi&#64257;cantly over seasonal and interannual variations. Although several remote-sensing-based vegetation indices (e.g. NDVI and EVI) have been widely adopted in digital soil mapping, variables indicating long-term vegetation growth status have been less used. The vegetation phenology, an indicator of vegetation growth characteristics, can be used as a potential time series environmental covariate for SOC prediction. In this study, a CNN-RNN hybrid model was developed for SOC prediction with inputs of static and dynamic environmental variables in a study area located in Xuancheng City, China. The spatially contextual features in static variables (e.g., topographic variables) were extracted by the convolutional neural network (CNN), while the temporal features in dynamic variables (e.g., vegetation phenology over a long period) were extracted by a recurrent neural network (RNN) as represented by using a long short-term memory (LSTM) network. The ten-year phenological variables before the sampling year derived from satellite-based observations were adopted as new predictors reflecting historical temporal changes in vegetation in addition to the commonly used static variables. The random forest model was used as a reference model for comparison. Our results indicate that adding phenological variables can improve the soil carbon prediction accuracy, and demonstrate that the &#64257;ne-tuned CNN-RNN model is potentially effective and can be a powerful model for SOC predictive mapping. We conclude that the hybrid deep learning models have great potential to enhance soil prediction by simultaneously extracting spatial and temporal latent features from different types of environmental variables, and highlight that using the long-term historical vegetation phenology information can serve as a useful extra input for future applications in the predictive mapping of soil carbon.</p> <p>References</p> <p>Zhang, L., Cai, Y., Huang, H., Li, A., Yang, L., Zhou, C., 2022. A CNN-LSTM Model for Soil Organic Carbon Content Prediction with Long Time Series of MODIS-Based Phenological Variables. Remote Sensing 14, 4441<br />Yang, L., Cai, Y., Zhang, L., Guo, M., Li, A., Zhou, C., 2021. A deep learning method to predict soil organic carbon content at a regional scale using satellite-based phenology variables. International Journal of Applied Earth Observation and Geoinformation 102, 102428<br />He, X., Yang, L., Li, A., Zhang, L., Shen, F., Cai, Y., Zhou, C., 2021. Soil organic carbon prediction using phenological parameters and remote sensing variables generated from Sentinel-2 images. CATENA 205, 105442.</p>