A CNN-LSTM Model for Soil Organic Carbon Content Prediction with Long Time Series of MODIS-Based Phenological Variables

Abstract

The spatial distribution of soil organic carbon (SOC) serves as critical geographic information for assessing ecosystem services, climate change mitigation, and optimal agriculture management. Digital mapping of SOC is challenging due to the complex relationships between the soil and its environment. Except for the well-known terrain and climate environmental covariates, vegetation that interacts with soils influences SOC significantly over long periods. Although several remote-sensing-based vegetation indices have been widely adopted in digital soil mapping, variables indicating long term vegetation growth have been less used. Vegetation phenology, an indicator of vegetation growth characteristics, can be used as a potential time series environmental covariate for SOC prediction. A CNN-LSTM model was developed for SOC prediction with inputs of static and dynamic environmental variables 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 of time) were extracted by a long short-term memory (LSTM) network. The ten-year phenological variables derived from moderate-resolution imaging spectroradiometer (MODIS) observations were adopted as predictors with historical temporal changes in vegetation in addition to the commonly used static variables. The random forest (RF) model was used as a reference model for comparison. Our results indicate that adding phenological variables can produce a more accurate map, as tested by the five-fold cross-validation, and demonstrate that CNN-LSTM is a potentially effective model for predicting SOC at a regional spatial scale with long-term historical vegetation phenology information as an extra input. We highlight the great potential of hybrid deep learning models, which can simultaneously extract spatial and temporal features from different types of environmental variables, for future applications in digital soil mapping.

A multi-target prediction model for dam seepage field

“…The MPM model consists of two different parallel networks, CNN and LSTM (the framework is shown in Figure 2). (1) Data collection and preprocessing: abnormal data were eliminated, and data format conventions and normalization were carried out for environmental variables and target values (Yang D. S. et al, 2020) (3) Feature fusion layer: There are two main feature fusion methods in hybrid neural networks: addition (Yang J et al, 2020) and concatenation (Zhang et al, 2022). Feature addition requires the same dimension of feature vectors, while concatenation is more flexible.…”

“…Feature addition requires the same dimension of feature vectors, while concatenation is more flexible. Moreover, feature fusion through concatenation has the effect of reinforcing and interconnecting features (Zhang et al, 2017;Zhang et al, 2022). Therefore, concatenation is used to fuse feature vectors in MPM.…”

Assessing the Feasibility of Wave Energy Harvesting through Coastal Wave Modeling and Forecasting with Long Short-Term Memory Optimization using Univariate Time Series Approach

“…An LSTM layer exists in the LSTM neural network. The surrounding neurons in the same layer are also affected along with the output layer [43,51].…”

“…LSTM (Long Short-Term Memory) is a type of recurrent neural network (RNN) designed explicitly for processing sequential data, such as time series. LSTM has been shown to be very effective in univariate time series forecasting and, in many cases, outperforms traditional statistical methods such as ARIMA and exponential smoothing [43,[49][50][51].…”

Monitoring soil organic carbon under coastal restoration using time series Sentinel‐1

“…Analytical techniques have been used to assess the importance of predictor variables, with mathematical methods (e.g., correlation analysis, recursive feature elimination) being the most commonly used (Wadoux et al, 2020). Although statistically important variables are efficient for model calibration, recent studies have suggested that retaining predictor variables with physical meaning would make predictive models more interpretable and improve prediction accuracy (Yang et al, 2021; Zhang et al, 2022). For example, Yang et al (2020) selected remote sensing‐based phenological parameters that were physically related to SOC dynamics in croplands, and they showed that models with phenological parameters exhibited higher accuracy than models with variables lacking physical meaning for predicting topsoil organic C content (Yang et al, 2020).…”