Research Projects

Timely identification of fine-grained semantic information (labels: double plant, drydown, endrow, nutrient deficiency, planter skip, storm damage, water, waterway and weed cluster) of agricultural fields is vital for the farmers to take an appropriate action in order to prevent loss or improve yield. Due to the massive sizes of agricultural fields, manual identification of the semantic information at regular intervals of time in near-real time is infeasible. In literature, several research groups have attempted to automate identification of semantic information using a combination of aerial images using unmanned aerial systems, satellite imagery and/or other proximal sensors. However, the current methods do not achieve satisfactory results in obtaining fine-grained annotations. This is due to the fact that the data from the existing set of sensors is insufficient for separating the finer annotations. Also, in literature, the current evaluation measures for assessing the segmentation results are not appropriate from a economic standpoint. In this project, our aims are four-fold: 1) we plan to propose a measure for evaluating the segmentation that aligns with the economic implications of the estimated semantic labels, apply the existing methods to obtain a baseline performance on the new evaluation measure. 2) Develop a system that can collect sensor data from various types of distributed sensors such as ultrasound sensors, LIDAR, optoelectronic sensors, etc. in a seamless manner. Deploy a variety of sensors on a test agricultural land and collect signals from them using the system deployed to collect the sensor data. 3) Define a cost-optimal metric for sensor-based agriculture. Identify the cost-optimal set of sensors to be deployed alongside RGB-NIR bands such that automated segmentation models on such data allow actionable insights. 4) With respect to the cost-optimal metric in point 3, we will obtain a novel algorithm on the augmented sensor data to establish a new baseline for semantic segmentation of aerial agricultural images.