Application of principles of microservice architecture and big data processing for automation of analysis and forecasting of soil fertility indicators in agricultural lands

Abstract

The study addresses the gap between the demand for operational fertility management and the inert, labor-intensive practices of traditional agrochemical surveys. The objective is the development and validation of a hardware-software system that integrates principles of microservice architecture and big data processing to automate the analysis and forecasting of soil parameters. The scientific novelty lies in a scalable, fault-tolerant platform with asynchronous communication and elastic scaling, enabling analysis of heterogeneous data sources in near real time. The practical significance is the support of precision agriculture, acceleration of decision-making, and reduction of costs at the level of large agricultural enterprises during sowing campaigns. The methodological foundation includes decomposition into ingest data, data processing, ml analysis, forecasting engine, and api gateway services deployed in containers with orchestration and a message broker. A dataset for 2015-2024 was compiled: over 50,000 agrochemical points (humus, mobile forms of P and K, pH), multispectral images from Landsat 8 and Sentinel 2 with NDVI/EVI, SRTM, meteorological parameters, and agrotechnological operations. Gradient boosting and random forest were applied for current parameter estimation, LSTM for dynamics; validation was carried out using k-fold methodology with R², MAE, RMSE metrics, alongside performance testing on a 1 TB data package. The results demonstrated advantages over the monolithic approach: a 35,15% reduction in average processing time, a 93,37% increase in throughput, and a 49,5% reduction in the 95th percentile latency, with comparable peak CPU usage. Model accuracy is high: for humus R²=0.917, MAE=0,12%; for P2O5 R²=0.862, MAE=11,4 mg/kg. LSTM forecasts for K2O show R²=0.881 (1 year), 0.795 (3 years), and 0.713 (5 years) at p 0.001, providing short- and medium-term planning value. The discussion confirms that microservices eliminate bottlenecks through parallelism, isolation, and flexible scaling, while asynchronous patterns stabilize response under load. Integration of remote sensing, relief, and field data enables operational supply maps and predictive planning. Economic modeling for 10,000 ha indicates a 75% reduction in analysis costs, 18,7% fertilizer savings, a 6,4% increase in yield, and a 28,4% increase in profit. Future prospects include the integration of weather forecasts, MLOps, and prescriptive recommendations.