Soil Texture Prediction with Automated Deep Convolutional Neural Networks and Population Based Learning

Context: Convolutional neural networks (CNNs) for spectroscopic modelling are usually designed and tuned manually and decisions regarding the CNN components are made heuristically. This can lead to poorly designed CNN models that are not fully optimised, or to models that take a long time to train in order to derive the optimal model. Although recent work has shown the importance of tuning the CNN hyperparameters, few studies have looked to tune the CNN components and hyperparameters concurrently. We propose such an approach when designing CNNs for soil spectroscopy, and test its effectiveness with the Lucas soil library. This approach has two phases. First, we automate the process of building the fully connected network. This first stage involves automating the selection of the different types of layers and the number of neurons per layer. In the second stage, we automate the selection of the hyperparameters for each layer. For the second stage, we experimented with two types of strategies for selecting the best combination of hyperparameters for each layer. The first strategy adapts the population based training (PBT) technique by replacing the random search used in PBT with a Bayesian Optimisation method, which we call adapted-PBT. In the second strategy, we employ the Bayesian Optimisation method. Our study reveals that our adapted-PBT was able to achieve high model performance when compared to previous studies. Our approach simplified the design of one dimensional CNNs (1D CNN) and was applied to predict soil texture properties of the LUCAS soil library. Results from our modelling were compared with two recent CNN studies conducted using the model efficiency coefficient (MEC), where the outcomes reveal improvements (5% to 26%) in performance for all three soil properties (sand, silt and clay). These findings provide new insights that could help advance the use of CNNs in soil spectroscopy modelling by concurrently combining the building of the CNN components with tuning the hyperparameters