Neural Network Approach Using PyTorch to Predict the Growth of Various Types of Plants

Abstract

In the era of rapid technological advancement, agriculture faces increasing challenges in optimizing production efficiency and managing resources sustainably. In Indonesia, various plant types are essential agricultural commodities, yet their productivity is often disrupted by erratic weather, poor land management, pest infestations, and land-use change. This study proposes a predictive model for plant growth using a neural network implemented in the PyTorch framework, integrating multiple environmental features such as temperature, humidity, soil moisture, nutrients, pH, and NPK levels. Unlike previous works that typically focus on specific crops or limited variables, this research introduces a multivariate approach combining diverse agro-environmental data to classify plant types accurately. The model architecture was tuned using GridSearchCV, resulting in optimal hyperparameters (e.g., batch size 32, learning rate 0.001, activation: tanh), achieving high performance with Area Under the Curve (AUC) values nearing 1.0 across most classes. Visualization of network weights reveals how input features are transformed through hidden layers, providing interpretability and transparency in decision-making. The proposed system demonstrates strong generalization capability, as validated on unseen data, and offers real-time prediction feasibility for deployment on edge devices such as NVIDIA Jetson Nano. This work contributes a novel, data-driven approach to smart agriculture by enabling precise growth prediction across multiple plant types, enhancing strategic planning for resource allocation and crop management. Future work includes model adaptation for time-series forecasting and validation with live sensor inputs in real-world agricultural environments.