Coffee Rust Severity Analysis in Agroforestry Systems Using Deep Learning in Peruvian Tropical Ecosystems

This article characterizes two coffee agroforestry systems in the districts of Chirinos (San Ignacio province) and San José del Alto (Jaén province), Cajamarca, northern Peru, and estimates coffee leaf rust severity using deep learning. Following a quantitative descriptive approach, four plots and ten subplots were surveyed for associated forest species (Inga edulis and Persea americana predominating), coffee varieties (Pache, Caturra, Bourbon, Geisha, Catimor) and shade cover (25.5–67.5%, measured with HabitApp and a visual template). Traditional polycultures were the most common system; no commercial polycultures were found. A total of 319 leaf photographs were collected with a professional camera, then segmented and classified with MobileNet and VGG16 transfer learning models under two severity scales — SENASA Peru (grades 0-4) and SENASICA Mexico (grades 0-6) — supported by the Leaf Doctor app. Grade 1 was the most prevalent severity level under both scales (1-5% of leaf area affected per SENASA; 0-2% per SENASICA). MobileNet achieved the best overall accuracy — 94% under SENASA and 92% under SENASICA after 50 epochs, against 91% and 90% for VGG16 — with both models performing strongly on the dominant grades 0 and 1 but only moderately on under-represented classes. The authors note the absence of meteorological analysis (prolonged leaf wetness, temperature) as a limitation and conclude that these algorithms offer a viable path toward low-cost computational tools for early rust detection and timely control and mitigation measures in coffee production.