Hybrid CNN-LSTM model for predicting nitrogen, phosphorus, and potassium (NPK) fertilization requirements: Integrating satellite spectral indices with field microclimate data

Abstract

This study presents a deep learning approach to predict nitrogen (N), phosphorus (P), and potassium (K) fertilization requirements using satellite and climate data. A hybrid CNN-LSTM model was developed to combine spatial features of Sentinel-2 vegetation indices (NDVI, NDRE, MSAVI, RECI) with temporal daily climate variables, including temperature, humidity, precipitation, wind speed, and solar radiation.

The model was trained on 3,208 samples integrating spectral, climatic, and field information such as parcel size and observation dates, and tested on a fully separated five-month period. The evaluation on the normalized scale demonstrated strong performance, with test results as follows: for nitrogen, MSE $=$ 0.0208, MAE $=$ 0.1132, and $R^2=0.9542$; for phosphorus, MSE $=$ 0.0281, MAE $=$ 0.1313, and $R^2=0.9480$; and for potassium, MSE $=$ 0.0225, MAE $=$ 0.1154, and $R^2=0.9474$. The model’s stability was further confirmed by consistent predictions across four individual months. This approach effectively integrates multimodal data for robust nutrient forecasting and can assist farmers in optimizing fertilization strategies. The outcomes support improved crop management, reduced environmental impact, and increased yields, especially in regions with limited ground data.