Weed Detection from Unmanned Aerial Vehicle Imagery Using Deep Learning-A Comparison between High-End and Low-Cost Multispectral Sensors.

In order to meet the increasing demand for crops under challenging climate conditions, efficient and sustainable cultivation strategies are becoming essential in agriculture. Targeted herbicide use reduces environmental pollution and effectively controls weeds as a major cause of yield reduction. The key requirement is a reliable weed detection system that is accessible to a wide range of end users. This research paper introduces a self-built, low-cost, multispectral camera system and evaluates it against the high-end MicaSense Altum system. Pixel-based weed and crop classification was performed on UAV datasets collected with both sensors in maize using a U-Net. The training and testing data were generated via an index-based thresholding approach followed by annotation. As a result, the F1-score for the weed class reached 82% on the Altum system and 76% on the low-cost system, with recall values of 75% and 68%, respectively. Misclassifications occurred on the low-cost system images for small weeds and overlaps, with minor oversegmentation. However, with a precision of 90%, the results show great potential for application in automated weed control. The proposed system thereby enables sustainable precision farming for the general public. In future research, its spectral properties, as well as its use on different crops with real-time on-board processing, should be further investigated.

Learning a Transform Base for the Multi- to Hyperspectral Sensor Network with K-SVD.

A promising low-cost solution for monitoring spectral information, e.g., on agricultural fields, is that of wireless sensor networks. In contrast to remote sensing, these can achieve more continuous monitoring due to their long-term deployment and are less impacted by the atmosphere, making them a promising solution for the calibration of satellite data. In this paper, we explore an alternative approach for processing data from such a network. Hyperspectral sensors were found to be too complex for such a network. While previous work considered fusing the data from different multispectral sensors in order to derive hyperspectral data, we shift the assessment of the hyperspectral modeling in a separate preprocessing step based on machine learning. We then use the learned data as additional input while using identical multispectral sensors, further reducing the complexity of the sensors. Despite requiring careful parametrization, the approach delivers hyperspectral data of similar and in some cases even better quality.

Systematic Approach for Remote Sensing of Historical Conflict Landscapes with UAV-Based Laserscanning.

In order to locate historical traces, drone-based Laserscanning has become increasingly popular in archaeological prospection and historical conflict landscapes research. The low resolution of aircraft-based Laserscanning is not suitable for small-scale detailed analysis so that high-resolution UAV-based LiDAR data are required. However, many of the existing studies lack a systematic approach to UAV-LiDAR data acquisition and point cloud filtering. We use this methodology to detect anthropogenic terrain anomalies. In this study, we systematically investigated different influencing factors on UAV-LiDAR data acquisition. The flight parameters speed and altitude above ground were systematically varied. In addition, different vegetation cover and seasonal acquisition times were compared, and we evaluated three different types of filter algorithms to separate ground from non-ground. It could be seen from our experiments that for the detection of subsurface anomalies in treeless open terrain, higher flight speeds like 6 m/s were feasible. Regarding the flight altitude, we recommend an altitude of 50-75 m above ground. At higher flight altitudes of 100-120 m above ground, there is the risk that terrain characteristics smaller than 50 cm will be missed. Areas covered with deciduous forest should only be surveyed during leaf-off season. In the presence of low-level vegetation (small bushes and shrubs with a height of up to 2 m), it turned out that the morphological filter was the most suitable. In tree-covered areas with total absence of near ground vegetation, however, the choice of filter algorithm plays only a subordinate role, especially during winter where the resulting ground point densities have a percentage deviation of less than 6% from each other.

UAV-Based RGB Imagery for Hokkaido Pumpkin (Cucurbita max.) Detection and Yield Estimation.

Pumpkins are economically and nutritionally valuable vegetables with increasing popularity and acreage across Europe. Successful commercialization, however, require detailed pre-harvest information about number and weight of the fruits. To get a non-destructive and cost-effective yield estimation, we developed an image processing methodology for high-resolution RGB data from Unmanned aerial vehicle (UAV) and applied this on a Hokkaido pumpkin farmer's field in North-western Germany. The methodology was implemented in the programming language Python and comprised several steps, including image pre-processing, pixel-based image classification, classification post-processing for single fruit detection, and fruit size and weight quantification. To derive the weight from two-dimensional imagery, we calculated elliptical spheroids from lengths of diameters and heights. The performance of this processes was evaluated by comparison with manually harvested ground-truth samples and cross-checked for misclassification from randomly selected test objects. Errors in classification and fruit geometry could be successfully reduced based on the described processing steps. Additionally, different lighting conditions, as well as shadows, in the image data could be compensated by the proposed methodology. The results revealed a satisfactory detection of 95% (error rate of 5%) from the field sample, as well as a reliable volume and weight estimation with Pearson's correlation coefficients of 0.83 and 0.84, respectively, from the described ellipsoid approach. The yield was estimated with 1.51 kg m-2 corresponding to an average individual fruit weight of 1100 g and an average number of 1.37 pumpkins per m2. Moreover, spatial distribution of aggregated fruit densities and weights were calculated to assess in-field optimization potential for agronomic management as demonstrated between a shaded edge compared to the rest of the field. The proposed approach provides the Hokkaido producer useful information for more targeted pre-harvest marketing strategies, since most food retailers request homogeneous lots within prescribed size or weight classes.