OPTIMIZING UAV DATA PROCESSING FOR PATTERN CLASSIFICATION WITH CNN ON LOW TO MODERATE-QUALITY IMAGERY
Over the last decade, UAV systems have enabled high-resolution data collection for various applications at relatively low
costs and with great flexibility in acquisition time and parameters [5]. This data can serve as a valuable reference for large-
scale space-borne applications. However, the flexibility in image acquisition presents challenges related to varying data types
and quality, which are affected by environmental conditions, sensor specifications, and radiometric calibration. Capturing
comparable reflectance values with UAV systems is particularly challenging, and many early studies relied on minimal
preprocessing or raw digital number (DN) values [4]. Given that some datasets' spectral information (reflectance/DN) may
not be directly comparable, a classifier that emphasizes generalized texture information is needed rather than relying solely
on spectral data. Among common machine learning (ML) techniques, the convolutional neural network (CNN) of deep
learning (DL) has proven to be a successful tool in classifying images of land use from remote sensing data [1, 2]. CNN
allows high-order representation based on generalized texture information already used in crop classification [7, 3, 6].
Our research explores the potential of using low-to-moderate-quality UAV data for agricultural pattern classification,
focusing on how color-balancing techniques can enhance data consistency when images are captured under variable lighting
conditions. We evaluated the performance of CNNs in classifying agricultural patterns using moderate- to low-quality, high-
resolution (0.07-meter) optical multispectral data collected from three agricultural test sites in Germany between 2019 and
- We used models trained exclusively on samples converted to reflectance values and applied them to images impacted
by different sunlight conditions, including digital number (DN) and reflectance data. The models were trained to classify
small-scale agricultural patterns, such as damaged and undamaged canopy, weed-infested and bare soil areas, across four
crop types: winter wheat, rapeseed, corn, and spring barley.
This study, funded by the German Federal Ministry for Economic Affairs and Energy (FKZ: 50EE1901), is carried out in
collaboration with CLAAS E-Systems GmbH to develop an application for crop monitoring based on Sentinel-1 data.
