XAI for small-data problems in remote sensing: monitoring Atlantic forests with UAVs

• DOI: 10.5194/egusphere-egu24-9192
• Publication Date: 2024-03-08T19:51:51Z
• Authors: Pranav Chandramouli; Caroline M. Gevaert; Francesco Nattino; Ou Ku; Alexandra Aguiar ...; Patricia do Prado...; Eduardo Hortal Pe...; Felipe de Ovileira
• Keywords: Atlantic forest

Despite the increased availability of UAV / drone imagery in Low- to upper  Middle-Income countries and demonstrated potential deep learning support interpretation these images for sustainable development purposes, practical operations are constrained by need sufficient labeled data-sets which often difficult obtain (especially tropical forest). This makes it train suitable networks assess whether model is performing well. One such example use drones monitor Atlantic Forest Sao Paulo, Brazil. Here, members Paulo Municipal Green Environment Secretariat (Secretaria do Verde e Meio Ambiente - SVMA) starting  identify some native invasive species their forests. Deep will quickly speed up this process, but there little training data available. refers so called ‘small-data problem’ commonly found DL remote sensing applications [1]. A workflow was designed application through a novel zero-shot technique explainable AI methods. pre-trained tree-crown detection ‘DeepForest’ [2] used identify individual tree crowns imagery. The detected tree-crowns further classified using Siamese network architecture – trained on relevant not exposed test data-set. motivated explainability models results be making administrative decision forest management. more intricate (such as image segmentation) could accurate at cost transparency/explainability. In particular, we apply variation ‘What I Know’ (WIK) method [3] provides examples from set along with sample increasing transparency understanding results. [1] Safonova, Anastasiia, et al. "Ten techniques address small problems sensing." International Journal Applied Earth Observation Geoinformation 125 (2023): 103569. Weinstein, Ben G., "DeepForest: Python package RGB crown delineation." Methods Ecology Evolution 11.12 (2020): 1743-1751. Ishikawa, Shin-nosuke, "Example-based its classification." 118 103215.