Machine Learning-Based Crop Drought Mapping System by UAV Remote Sensing RGB Imagery

Water stress has adverse effects on crop growth and yield, where its monitoring plays a vital role in precision crop management. This paper aims at initially exploiting the potentials of UAV aerial RGB image in crop water stress assessment by developing a simple but effective supervised learning system. Various techniques are seamlessly integrated into the system including vegetation segmentation, feature engineering, Bayesian optimization and Support Vector Machine (SVM) classifier. In particular, wheat pixels are first segmented from soil background by using the classical vegetation index thresholding. Rather than performing pixel-wise classification, pixel squares of appropriate dimension are defined as samples, from which various features for pure vegetation pixels are extracted including spectral and color index (CI) features. SVM with Bayesian optimization is adopted as the classifier. To validate the developed system, a Unmanned Aerial Vehicle (UAV) survey is performed to collect high-resolution atop canopy RGB imageries by using DJI S1000 for the experimental wheat fields of Gucheng town, Heibei Province, China. Two levels of soil moisture were designed after seedling establishment for wheat plots by using intelligent irrigation and rain shelter, where field measurements were to obtain ground soil water ratio for each wheat plot. Comparative experiments by three-fold cross-validation demonstrate that pixel-wise classification, with a high computation load, can only achieve an accuracy of 82.8% with poor F1 score of 71.7%; however, the developed system can achieve an accuracy of 89.9% with F1 score of 87.7% by using only spectral intensities, and the accuracy can be further improved to 92.8% with F1 score of 91.5% by fusing both spectral intensities and CI features. Future work is focused on incorporating more spectral information and advanced feature extraction algorithms to further improve the performance.