Search for a command to run...
The Varroa destructor mite is a leading cause of colony losses, yet most hyperspectral studies end at pixel-level detection and do not indicate when treatment is economically justified. The leakage-controlled testing with spatially blocked and Leave-One-Tile-Out (LOTO) protocols demonstrates strong within-scene robustness under conservative testing, without explicitly stating that cross-scene generalization is the case in view of the current annotation scope. The spectral-to-economic decision framework used here is validated with respect to internal consistency as well as operational feasibility, with a transition to true leave-one-scene-out evaluation postponed for future work upon obtaining additional annotations. Instead of proposing a new spectral backbone, the present work presents a decision-oriented hyperspectral framework based on strong infestation risk estimation while explicitly economically optimizing to ensure interpretability and reliability under limited annotation. Pixel-level probabilities produced by a support vector machine and a random forest are aggregated into a scene-level Hyperspectral Infestation Risk Index, which is then mapped to expected monetary loss. The operating rule is chosen by maximizing net benefit, with a closed-form threshold τ ⋆ ≈ <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</i>/(λ<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">py</i>) based on treatment cost c, loss cap λ, honey price <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</i>, and expected yield <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">y</i>. Numerical validation shows that, on the BUT-HS1 dataset, the models achieve a macro area under the precision–recall curve of 1.00 and a macro area under the receiver operating characteristic curve of 1.00 under standard splits. Under a spatially conservative leave-one-tile-out protocol with buffered regions and hard negative mining, the weighted accuracy is 0.896 with only localized failures at tile boundaries, confirming genuine spectral separability. Decision-curve analysis indicates that, at 2023 United States price and cost levels, τ ⋆ = 0 (treat all) maximizes value; when per-hive treatment cost rises to 8 dollars, τ ⋆ ≈ 0.09 defers about 16–17% of treatments while retaining about 98% of avoidable loss, reduces profit-at-risk by $18–$26 per scene, and saves$2–$8 per colony per application.