Estimation of fruit crop size using UAV images and deep learning techniques: applications from agriculture to ecology
| Autor: | Apolo Apolo, Orly Enrique, Mendoza, Irene, Ramírez-Santos, Adrián, Díaz-Delgado, Ricardo, Egea, Gregorio, Jordano, Pedro |
|---|---|
| Jazyk: | angličtina |
| Rok vydání: | 2021 |
| Předmět: | |
| Zdroj: | Digital.CSIC. Repositorio Institucional del CSIC instname |
| Popis: | Trabajo presentado en el XV Congreso Nacional de la Asociación Española de Ecología Terrestre (El valor de la Naturaleza para una sociedad global), celebrado en Plasencia del 18 al 21 de octubre de 2021. Frugivorous animals critically depend on fleshy fruits for getting the caloric income needed for their subsistence and perform beneficial seed dispersal for the plants in return. Fruit production is dependent on climatic cues and there is evidence that reproduction of plants has been affected by climate change. Being able to robustly estimate fruit crop size in fleshy-fruited species is therefore key for evaluating the availability of resources to animals depending on them and for monitoring phenological changes through time. Fruit load estimation has usually been done using manual procedures by which researchers visually estimate fruit crop size using either qualitative or quantitative records. But this procedure is costly, time-consuming, and restricted to small spatial scales. The combination of drones and artificial intelligence (AI) has successfully been used in agriculture for fruit counting and yield estimation. The procedure offers promising, and rarely explored, alternatives in natural ecosystems for automatic fruit counting. As a pilot study, we took images of Pistacia lentiscus grown in scrublands in Doñana National Park (SW Spain) using a drone at three different heights. A pre-trained object detection model (Faster R-CNN) using transfer learning was trained to develop an automated image processing methodology applied to fruit load estimations (number of infructescences). Model’s results were cross-validated with on-ground direct count values, showing reasonably good accuracy in detecting infructescences at low height. This opens new possibilities for the use of AI-based models in ecological research applications at much larger spatial scales and on landscapes of variable complexity. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|