DL4Cashew: Deep Learning Model For Cashew Trees Counting

---

Résumé

Tree identification and counting are processes that provide an inventory on the different plant species covering a field. In this paper, we presenté DL4Cashew, a deep learningbased identification model that can identify cashew trees on a heterogeneous agricultural area. We also proposed an algorithm that uses the results of the identification to count the number of cashew trees a threshold value of 0.8 above which the tree is not counted as a cashew. The construction of the learning model was done using RGB images collected by a UAV taken in various light conditions and altitudes. Our resolution approach consisté of labeling the raw images collected from control fields in PascalVOC format. Then, several models were investigated, and their performance was evaluated. The Faster-RCNN model to demonstrate a better performance of 93% compared to the other models on the image jet considered in this work. For the counting part, we considered the map of a control field of 11.53 hectares. From this map, we applied segmentation techniques to generate patches of dimension 1020x120 containing valid trees. An algorithm for filtering patches containing valid trees was proposed. We subsequently performed DL4Cashew model inference on each patch. Finally, we performed an aggregation of the number of feet on each patch to obtain the total number on the total area. In total, 358 cashews feet were detected by the model on the control field.

Mots-clés

Deep learning, Image segmentation, Image resolution, Filtering, Computational modeling, Vegetation, Autonomous aerial vehicles, Cashew, Deep Learning, UAV, RGB, Identification, CNN