Tomato farming plays a major role in Ghana’s agricultural sector by contributing to food supply and serving as a source of income for many smallholder farmers. However, tomato plants are easily affected by leaf diseases that can spread quickly and reduce crop yield if not detected early. In many farming communities, disease detection is still done manually, which is often slow, inconsistent, and affected by human error. This study developed a deep learning-based system for the early detection of tomato leaf diseases using a Convolutional Neural Network (CNN) based on the VGG16 architecture. A total of 16,211 tomato leaf images, comprising healthy leaves and nine disease classes, were used for the study. The images were resized, normalized, and enhanced through data augmentation techniques such as rotation, flipping, shifting, and zooming. The model achieved its highest validation accuracy of 96.8% at epoch 16, while the final validation accuracy at epoch 20 was 94.30%, demonstrating strong performance in tomato disease classification.

Conventional cyclone risk assessments in Small Island Developing States (SIDS) systematically fail to predict food system collapse locations, leading to misallocated adaptation investments. This study quantifies a fundamental exposure-impact disconnect across Vanuatu's 71 area councils: cyclone exposure explains only 14% of food security impact variance (R² = 0.14, r = 0.38), with topographic, agro-ecological, and infrastructure mediators determining the remaining 86%. Using data from five Category 4–5 cyclones (2015–2023), we constructed a Hazard Exposure Index (HEI; proximity 40%, intensity 35%, frequency 25%) and an Impact Severity Index (ISI; four equally weighted dimensions: agriculture, infrastructure, food insecurity, recovery). Regression with interaction terms reveals three mediating mechanisms. First, topographic elevation reduces ISI by 40–60% at highland area councils (>400 m) relative to coastal equivalents at comparable HEI (r = −0.54, p < 0.001). Second, crop diversity buffers medium-exposure zones (HEI 0.3–0.6; β = −0.082, p = 0.018) but provides no significant protection under extreme cyclone loading (HEI > 0.7), consistent with resilience theory stability thresholds. Third, infrastructure capacity generates "hidden hotspots": Torba province (low exposure, ISI = 0.34) required a median 18 days to reach 50% of affected households versus 6 days in Shefa (high exposure, ISI = 0.61) — a threefold differential driven solely by maritime connectivity deficits. High-exposure provinces with strong infrastructure systematically outperform low-exposure provinces with distribution network weaknesses. A Food Security Resilience Index (FSRe; range: 46–61 on a 0–100 scale) reveals stark provincial heterogeneity — Sanma as buffered regional hub (FSRe = 61) versus Tafea and Shefa as critical/fragile (FSRe 46–47) — enabling a four-category spatial typology for targeted adaptation investments under SIDS fiscal constraints. Findings demonstrate that spatially explicit integration of topography, agro-ecology, and infrastructure is essential for efficient climate adaptation planning in archipelago SIDS, with direct implications for Sendai Framework Target E national DRR strategy development and Green Climate Fund adaptation investment prioritization.