A carregar...
Publicação
Optimizing the Allocation of Field Samples for Remote Sensing and Machine Learning-Based Crop classification Models by Assessing the Area of Applicability
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 3.8 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Reliable and accurate crop classification information can provide key insights for food security, production
sustainability, agronomic decisions, logistics, and yield estimations; serving several different sectors and
industries. Since decades ago, crop classification has been escalated by the use of different remote sensing
sources and machine learning techniques, creating prediction models that could retrieve information from
continuous areas and classify them according to the ground-truth data used as references to identify and
map the targeted crops. However, the process of ground-truth sampling is nowadays the most challenging
part resources-wise, since its obtaining implies gathering reference information through field surveys and
crowdsourcing. Alternatively, efforts to reduce or guide ground-truth necessity have been studied, by using
limited or transferred reference data, which implies pre-existent databases availability and state-of-the-art
data extraction workflows. This study explores the Area of Applicability (AOA) assessment for validating
crop classification model predictions by identifying areas outside the estimated prediction limits and
providing different levels of confidence for the prediction made based on the training data and the variables
presented to the model. The outcomes could provide a variant for areas that do not hold or provide open
access to historical ground-truth data collections, to help optimize ground-truth samples allocation for crop
classifications, by offering a potential guide for indicating areas that need further samples collection for
presenting reliable predictions over these extensions, and could potentially bring further insights for
validation of practices involving prediction models’ spatiotemporal transferability.
Descrição
Dissertation submitted in partial fulfilment of the requirements for the Degree of Master of Science in Geospatial Technologies
Palavras-chave
Area of Applicability spatial analysis crop classification remote sensing machine learning ground-truth samples