Monitoring crop health, growth and its stand count attributes using UAV based precision agriculture: a study in tropical farmland of Thailand

Unmanned aerial vehicle (UAV) equipped with multispectral sensor has become an active research topic for crop health monitoring and has been widely used across many regions. However, the high cost associated with multispectral sensors suggests us to shift to a cheaper alternative in order to be implemented by average farmers from developing countries. We evaluated the feasibility of a lightweight (38 gm) mobius action camera with wide FOV in monitoring crop health through the removal of IR filter, and replacing it with Wratten 25A red filter. The research implements Structure from motion (SfM) for creating orthomosaic, and computes Normalized Difference Vegetation Index (NDVI). Finally, the results of NDVI from the modified camera was validated with the ground measurement of LAI carried using LI-COR LAI 2000, with a linear correlation that resulted in coefficient of determination of (R2) 0.843. The result demonstrates that the modified camera is potential in agricultural health monitoring.Likewise, this research also implemented SfM algorithm using UAV imagery supported with global positioning system (GPS) to generate multitemporal crop surface models, which assessed crop growth throughout the growth season. Banana plantation which normally takes 9-12 months from sowing to harvesting the fruit, had been sowed on the first week of November, 2016 and harvested on the mid of September, 2017. The field data acquisition was performed three times, particularly on 25th January, 26th April & 16th September 2017 in an area of 0.186 sq.km, with the area’s centroid coordinates at N 14O 15.133| E 100O 53.393| and Z 10 meters (WGS84). This study followed a methodology based on SfM to create a multitemporal surface models (DTM & DSM), followed by the difference method to generate canopy height model (CHM) which was used to assess crop growth. The growth ranged between 2.31-4.89 m for the period between Jan-Apr while the period between Apr-Sept demonstrated negative growth as a result of harvesting carried on September 10. The methodological framework adopted in this study will enable the spatial analysis of crop growth within banana plantation, enabling wide range of applications in the improvement of crop management.Furthermore, several studies has been performed for object detection from ground view perspective and has been the key topic of interest for computer vision communities, however very less has been explored in detecting objects in an aerial imagery. The convolutional neural network implemented in this study was based open source tensorflow implementation of the darknet framework named, Darkflow, which has been modified to a near real-time multi-scale detector implementing YOLOv2 object detection model to improve the performance on aerial imagery. To detect the palm trees, the YOLO v2. Neural Net was modified and fine-tuned on our dataset consisting 255 images of palm trees, each of 4000 x 3000 resolution taken at 70m above the ground. The images were manually annotated which consisted of 595 annotations representing validation dataset, whereas 1000 annotations representing training dataset required for training and accuracy assessment of the ConvNet. The annotations were created in xml format, using a python script which takes the manual input from the user regarding the bounding box of the object to be detected. Likewise, we applied pre-trained weights and configuration files for the PASCAL VOC datasets, which was modified by changing the no. of classes to 1 and the no. of filters in the last convolutional layer was modified to 30, which fits our purpose of detecting palm trees.Likewise, the batch size was set to 64, subdivisions to 8, learning rate to 0.0001 and the datasets were initially trained until 6500 iterations on GPU server consisting of 32 GB of NVidia Tesla P100-SXM2. Finally, the precision and recall for our object detection model was observed to be 45.59% and 65.87%, making our total accuracy to 55.23%, which requires further improvements before it could be directly applied for palm tree counting purpose. These initial results demonstrate that provided a large training dataset (approx. 5 times more than current) with good quality labeled images and intensive training time, YOLO v2 net can accurately detect palm trees in our project area. As for now, the training spends more time per epoch resizing than training due to large resolution images, therefore, to increase performance the future prospects of our work would focus on tiling the images into multiple sections of 666 x 500 pixel blocks using OpenCV, before actually feeding the images for training which is expected to optimize the network.