Daylight Calibration Method For Agricultural Remote Sensing With A Wide Angle “Sky Camera” And Deep Learning

An advancement of Unmanned Aerial Vehicle (UAV) technology accelerates airborne imaging in the agricultural sector. Various cameras, such as a multispectral camera or hyperspectral camera, are equipped with drones. It enables farmers to access several vegetation indices such as nitrogen or water stress. The image pixel value is a crucial variable of many indices calculations. However, the sunlight heavily influences image pixel values. The non-calibration data could lead to misinterpretation. The current daylight calibration method is using a white reference board made from highly reflective material. Including the white reference board provides sunlight intensity information. However, this method requires the white reference board's existence in every image. A spectrometer is a sensor that gives light spectrum intensity directly without the white reference board. This study develops a regression model to produce daylight intensity from spectrometer data. However, the result shows that the model outcome does not match with the white reference board. Although the spectrometer eliminates white reference board necessity, it cannot replace the white reference board method due to outcome incompatibility. A new daylight calibration method using sky information is introducing in this study. An RGB camera mounted with a wide-angle or fisheye lens pointing to the sky captures a sky's dynamics. A Convolutional Neural Network is trained using sky images. The model R-square is 0.997. The number of outliers, a prediction that mismatch its ground truth more than 10 percent, measures the model performance. From 921 samples, seven outliers existed or 0.8 percent. The proposed wide-angle camera solution can produce similar light intensity values as the physical white reference board method. This alternative method operation uses a single camera offering a higher practicable daylight calibration method for aerial remote sensing.