A DESN-based federated learning framework for NIR spectral analysis of organic carbon in farmland soils

• DESN-based NIR sensing framework for rapid analysis of soil organic carbon. • Novel design of DESN architecture for federal learning on distributed NIR data. • Optimization of the DESN by adjustable structure with tunable hyperparameters. • Fusion of MGFA method for scoring the features based on grouped distributed data. Maintaining rapid and reliable detection of soil organic carbon (OC) across diverse types of farmlands is important for quality control of agricultural products. Nevertheless, the information associated with different farmlands can be quite varied and diverse. This paper proposes a novel framework based on deep echo state network (DESN) to build decentralized analytical models for federated learning on the near-infrared (NIR) spectral data. In experiment, the methodology is validated using the LUCAS 2009 topsoil dataset from the European Soil Data Centre, which not only exhibits the decentralized property but also is suitable for simulating the federated learning under the internet of things (IoT) environment. For chemometric studies, the framework is specifically designed with flexible DESN network structures where hyperparameters are tuned to optimize reservoir units. An iterative learning strategy is introduced to evaluate the spontaneously generated reservoir neurons in each unit. Multi-group fusion analysis (MGFA) is proposed as an innovative component analysis method to enhance the extraction of feature variables, thereby adapting to the deep federated analysis of grouped variables. Experimental results demonstrate that the fusion of DESN and MGFA yields improved prediction performance in the modeling and analysis of distributed NIR data. Soil samples from distributed farmlands can be simultaneously analyzed through federated learning to quantify their OC contents using the proposed fusion method. The methodology presented in this study is expected to be integrated with IoT systems, thereby promoting the application of NIR chemometrics in agricultural applications.