DOST NRCP Research Proposal

farm G-E-N--: Evaluating and Improving CrossFarm Generalization of Drone Based Tree Species Classification Mon Arjay Malbog Faculty Member, P-U-P.

Tree Counting 9042 2043 A042 #063 043 8041 #045 L $ 360 metric Distance Measurement 232 metric Crop Classification Crop A Crop B Crop B Crop Health Analysis High Stress Hleas Height Trend bee Crown Spread Height & Crown Measurement Yield Analytics.

BackgroundoftheStudyDronesprovidehigh resolutionaerialimagery.AIturnsvisualpatternsintospeciespredictions—buttreeappearance,siteconditions,andimageacquisitionvaryfromfarmtofarm. Theresult:accuracycanfallsharplyonanunfamiliarfarm..

One model. A different farm. 01 02 Tree Appearance Site Plus ENVIRONMENT Can a model trained here classify species accurately there? 03 04 Drone Acquisition Limited Labels.

What the field already knows Cross site identification 7 relevant studies in preliminary review U-A-V multispectral mapping Semi supervised learning Recurring limitations Deep learning on UAV RGB Multi sensor fusion Poor transferability Limited labels Seasonal change Class imbalance.

[Audio] What the literature does not yet resolve UNSEEN farms method COMPARISON label budgets Evaluation rarely holds out truly unseen farms. Domain generalization techniques lack consistent headto head testing. The value of each additional target farm label remains unclear. The missing evidence sits at the intersection of all three..

[Audio] Research Novelty The novelty is in three areas: Cross farm generalization Comparative evaluation Limited target farm labels Testing models on farms completely excluded from training. Comparing baseline models with selected domaingeneralization methods. Measuring how different amounts of labeled data affect recovered performance..

[Audio] GENERAL OBJECTIVE Evaluate—and improve—cross farm generalization of drone based tree species classifiers when labels from the target farm are scarce..

[Audio] Four objectives, one deployment question 01 02 Evaluate performance across multiple farms Identify farm and acquisition factors linked to differences 03 04 Compare baselines with domain generalization approaches Quantify the effect of target farm label budgets.

[Audio] Three questions guide the study Q1 Q2 Q3 How does performance change across farms? Can transfer or adaptation reduce the need for labels? Which farm and image conditions explain the difference?.

PROPOSED RESEARCH FRAMEWORK Agricultural Drone Image Analysis Pipeline INPUT PROCESS → Drone-based tree imagery 14 Image Preprocessing Species annotations OUTPUT Cross-farm •ll performance measurements Comparative method effectiveness Farm characteristics Image acquisition conditions Baseline Model Training Domain-Generalization Experiments Farm-Held-Out Evaluation Target-Farm Label- Budget Experiments EFO Label-budget performance analysis Evidence-based recommendations.

[Audio] ExperimentalbydesignAquantitative,multi farmprotocol 01 02 03 04 05 06 07 AcquireVerify plus labelPrepareTrainGeneralizeHoldoutVarylabelsbaselinefarmsUnseen farmevaluationremainsindependentfromalltrainingdecisions..

[Audio] Accuracy is only one signal Robustness requires balance across species, farms, and label availability. Report together Overall accuracy Macro F1 Per species precision Per species recall Across farm variation Recovery by label budget.

[Audio] Evidence that travels beyond one farm Degradation factors Generalization evidence reliable deployment Method effectiveness Label budget insight Monitoring guidance.

[Audio] Feasible—with risks made explicit ready CONSTRAINED mitigated Multiple farms Dataset access Public data fallback Drone imagery Collection costs Consistent annotation Verified annotations Image inconsistency Essential experiments first Compute plus 1050 tools Class imbalance Independent farm testing Compute limits.

[Audio] From a model that works here— to one we can trust there. This study tests reliability across farms, identifies what breaks, compares how to improve it, and measures how many target farm labels are truly needed. Thank you · Questions?.