A2096
Title: High-dimensional clustering and random forest regression for pixel-level crop classification and suitability in Oman
Authors: Shahad Al Siyabi - Sultan Qaboos University (Oman) [presenting]
Iman Al Hasani - Sultan Qaboos University (Oman)
Abstract: In countries that lack extensive spatial agricultural data, using satellite and climate data can be beneficial, with Oman having national self-sufficiency in agriculture as one of its goals, and the lack of a fine resolution crop map or non-fieldwork tool that can be used remotely to watch the progress of the goal and the country's targets. A high dimensional feature matrix that covers the time series 2018-2026 has been used with satellite and climate data using an unsupervised clustering approach with KMeans and dimensionality reduction using UMAP, the cropland of Oman produced 12 different agricultural systems at 100m resolution. Suitability for target crops like onion, wheat, and potato is modelled using local water, soil, and terrain data with Random Forest, and the results show the top pixels to grow the crops in with more than 90 percent accuracy, and that water availability is the most important SHAP driver, highlighting that water infrastructure is where policymakers should prioritise for these crops and Oman's agricultural goals, then an interactive dashboard was created to make the data more accessible with a dynamic map.
