Towards a crop monitoring system for Hawai‘i: Evaluating machine learning approaches for mapping smallholder agriculture across space and time

Abstract

Smallholder agriculture represents a significant portion of global food production, yet effective monitoring of these farms remains challenging, especially in regions like Hawai‘i where frequent cloud cover and year-round cultivation complicate traditional crop mapping efforts. This study investigates the potential for establishing a robust, generalizable crop monitoring system (CMS) for smallholder agriculture in Hawai‘i using high-resolution, multi-temporal 8-band PlanetScope satellite imagery. Three supervised machine learning models—random forest (RF), multi-layer perceptron (MLP), and long short-term memory (LSTM)—were evaluated across varying time series input lengths, and their ability to generalize across space and time was assessed. Results confirm that incorporating temporal dependencies significantly enhances model performance, with LSTM demonstrating superior accuracy in classifying both binary (crop presence) and multi-class (crop growth stages) tasks. This research establishes methodologies for operationalizing CMSs in Hawai‘i to address agricultural data gaps and offers insights applicable to other smallholder agricultural systems more broadly.