Remote sensing for precision agriculture: Within-field spatial variability analysis and mapping with aerial digital multispectral images

Advances in remote sensing technology and biological sensors provided the motivation for this study on the applications of aerial multispectral remote sensing in precision agriculture. The feasibility of using high-resolution multispectral remote sensing for precision farming applications such as soil type delineation, identification of crop nitrogen levels, and modeling and mapping of weed density distribution and yield potential within a crop field was explored in this study. Some of the issues such as image calibration for variable lighting conditions and soil background influence were also addressed. Intensity normalization and band ratio methods were found to be adequate image calibration methods to compensate for variable illumination and soil background influence. Several within-field variability factors such as growth stage, field conditions, nutrient availability, crop cultivar, and plant population were found to be dominant in different periods.; Unsupervised clustering of color infrared (CIR) image of a field soil was able to identify soil mapping units with an average accuracy of 76%. Spectral reflectance from a crop field was highly correlated to the chlorophyll reading. A regression model developed to predict nitrogen stress in corn identified nitrogen-stressed areas from nitrogen-sufficient areas with a high accuracy (R2 = 0.93). Weed density was highly correlated to the spectral reflectance from a field. One month after planting was found to be a good time to map spatial weed density. The optimum range of resolution for weed mapping was 4 m to 4.5 m for the remote sensing system and the experimental field used in this study. Analysis of spatial yield with respect to spectral reflectance showed that the visible and NIR reflectance were negatively correlated to yield and crop population in heavily weed-infested areas. The yield potential was highly correlated to image indices, especially to normalized brightness. The ANN model developed for one of the research fields mapped spatial yield with 70% to 83% accuracy in different fields and seasons. The models at 6 m resolution performed better than the models at 3 m resolution. The best time to map yield potential of a field was after tasseling.