Introduction
The global food system is currently navigating a precarious "double bind." To sustain a growing population, global food production must increase by nearly 30% by 2050. However, the intensive farming methods currently in use—characterized by heavy reliance on chemical fertilizers and pesticides—are driving significant environmental degradation. This position paper explores how Predictive Process Monitoring (PPM) can bridge the gap between productivity and sustainability by optimizing one of agriculture's oldest tools: crop rotation.
Beyond Static Predictions
While modern agriculture has embraced technology, most applications focus on static outcomes. Traditional Machine Learning (ML) models are frequently deployed to predict end-of-season yields or soil salinity. While useful, these models often treat the farming cycle as a "black box," failing to account for the dynamic, sequential nature of agricultural activities.
The authors argue for a paradigm shift: viewing the agricultural cycle not as a set of isolated data points, but as a continuous process. By applying PPM, farmers can move from reactive harvesting to proactive, real-time decision-making, treating the farm as a workflow in execution.
A Three-Tiered Methodological Framework
The paper categorizes the current technological landscape into three distinct tiers based on how data is utilized:
1. Process Mining (PM)
Originally developed for business process management, PM is applied here to analyze how farmers interact with Decision Support Systems (DSS). By discovering the actual workflows of users, developers can refine software to better align with real-world farming needs.
2. Standard Prediction
This tier involves using established algorithms to forecast specific agricultural outcomes. Common methods include Regression, ARIMA, and ensemble techniques like Random Forests and XGBoost. These are typically used for point-in-time predictions such as crop yield or identifying deviations in growth patterns.
3. Predictive Process Monitoring (PPM)
PPM represents the cutting edge of this research. It treats the agricultural cycle as a sequence of events recorded in an "event log." This allows for more sophisticated modeling of crop rotations over multiple years. Key techniques include:
- Directly Follow Graphs (DFGs) and Markov Chains: Useful for visualizing and predicting the immediate next step in a crop sequence.
- Deep Learning (LSTM): Long Short-Term Memory networks are employed in sequence-to-sequence architectures to predict complex, multi-year rotation scenarios.
- Spatio-temporal Integration: This involves merging spatial data (location) with time-series data (timing) to provide a holistic view of farm progress.
The "Event Log" Challenge
A significant hurdle in implementing PPM is data transformation. Agricultural data is notoriously messy and siloed, originating from satellite imagery (NDVI), IoT sensors (LoRa WSN), and historical weather records. Converting these disparate sources into a standardized "event log" format—a requirement for PPM—remains a primary technical challenge.
Key Insights and Future Directions
The core contribution of this work is the conceptualization of a farm as a "business process." This perspective enables "conformance checking"—determining if a crop is developing according to plan—and allows for mid-stream interventions that can save a season's harvest.
While early results using LSTMs for sequence prediction are promising, the authors identify critical areas for future work:
- Data Integration: Standardizing the merger of soil, weather, and satellite data into unified logs.
- Interpretability: Farmers require "explainable AI" to trust and act upon model suggestions regarding their livelihood.
- Impact Assessment: Future research must determine which specific tasks offer the highest return on investment for long-term soil health and sustainable practices.
By treating farming as an evolving workflow rather than a series of static events, PPM offers a path toward a more resilient and transparent food production system.
Summary updated