Assessing High-Resolution SkySat Imagery and Deep Learning for Detection of Coconut Scale Insect Outbreaks in the Philippines

Abstract. Coconut scale insect (CSI; Aspidiotus rigidus) outbreaks cause canopy chlorosis and tree mortality, posing a serious threat to coconut-based livelihoods in the Philippines. This study assesses the potential of high-resolution SkySat imagery (0.5 m) combined with the DETReg deep-learning object detection model to automatically identify CSI-affected coconut crowns. Two outbreak sites were examined: Anahawan, Southern Leyte, and the Philippine Coconut Authority–Zamboanga Research Center (PCA-ZRC) in Talisayan, Zamboanga City. SkySat ortho analytic surface reflectance scenes were processed in ArcGIS Pro and QGIS to generate false-color composites and compute the Visible Atmospherically Resistant Index (VARI). A training dataset of 424 manually digitized symptomatic crowns (64 × 64-pixel chips, augmented with rotations) was used to train a DETReg model with a ResNet-50 backbone. Inference produced 1,974 detections in Anahawan and 5,955 in Talisayan, with an average precision of ~77%. VARI stratification revealed site-dependent specificity: detections in Anahawan were dominated by High-Moderate VARI values (99.44%), suggesting overprediction of spectrally healthy crowns, while Talisayan showed a higher proportion of Low-Negative VARI values (34.14%), consistent with chlorosis from severe infestation. Findings demonstrate that DETReg can sensitively flag candidate CSI crowns from SkySat RGB imagery, but spectral limitations of four-band data and reliance on crown texture contributed to false positives and reduced transferability across landscapes. Expanding labeled samples, integrating spectral indices and temporal stacks, and applying ensemble approaches are recommended. Until validated through stratified campaigns, the method is best suited as a triage tool to prioritize ground surveys rather than for autonomous deployment.