Application of Remote Sensing for Monitoring Crop Health and Growth

Abstract

This paper examines the use of hyperspectral remote sensing with artificial intelligence models to detect and classify various stress types of crops, including nitrogen deficiency, phosphorus deficiency, drought stress, and fungal disease, in non-homogenous agricultural systems in the initial stages. The hyperspectral and thermal images were obtained through an unmanned aerial vehicle on five types of crops (maize, wheat, soybean, potato, and tomato) during two growing seasons alongside ground-truth measurements of the chlorophyll content, biomass, and severity of the diseases. Nine machine learning and deep learning models were tested, such as Random Forest, Support Vector Machine, 1D Convolutional Neural Network, Transfer Learning CNN, Gradient Boosting, Logistic Regression, K-Nearest Neighbors, XGBoost, and LightGBM. The transfer learning CNN was the overall best performer, with a macro-average classification of 97.08, precision of 96.89-98.76, recall of 95.78-98.92, and AUROC of 0.989997 across stress classes. The capability of early detection was 9.2 days ahead of visual symptom development in case of drought stress and the average lead time was 8.18 days in all stresses. The red-edge wavelength at 710 nm was found to provide the highest spectral feature importance, with the chlorophyll index red-edge having the highest F-statistic. Intersection over Union values were greater than 89% across all stress types, and cross-crop generalization accuracy was between 96.61% (potato) and 97.53% (maize). The model was shown to have a high rate of inference (0.67 ms per sample), which validates its capability to be used in real-time. The findings confirm that hyperspectral imaging with the application of transfer learning can be used to provide accurate, early, and spatially resolved diagnoses of the physiological stress of the crops and provide a revolutionary tool in precision agriculture and sustainable management of crops.