A collection of good days cannot plan the next season.
Choose a real comparator
Compare with the actual local alternative: ground equipment, manual work or hired application. Record terrain, crop stage, weather and materials, using controls and replication where feasible. Agree sample size with the research lead beforehand.
Measure quality, workload and cost
Measure deposition, efficacy, crop damage, workload, setup, travel and rework alongside coverage. Pest outcomes and yield take time; footage immediately after flight cannot establish them. Define results that would cause rejection before testing.
Make failures reusable
Record versions of aircraft software, prescription and map. Preserve uncompleted jobs, missing data, repairs and operator changes. Report ranges and exceptions, then state the conditions under which adoption is justified.
Evaluate predictions on an unseen year
A Korean rice study compares regression approaches for yield and protein prediction using multispectral observations from 2018–2020 and cross-year evaluation. We reviewed the university-hosted abstract; it does not guarantee performance in other crops or regions. Our proposed evaluation reserves a new year or previously unused fields rather than relying on familiar training sites. Randomly dividing neighboring pixels from one flight can leave evaluation data too similar to training data. Define what will actually be new at deployment and split evaluation data at that level.
Define the cost of missed problems and false alarms
In an illustrative example, only five of 100 plots have a problem. Calling every plot normal gives 95% accuracy while missing every affected plot. Report the share of real problems detected, the share of flagged locations that truly have problems, and staff time spent checking. Estimate the missed treatment opportunities caused by false negatives and extra visits or flights caused by false positives. Compare settings favoring detection with settings reducing inspection workload instead of accepting an analyst-selected threshold without discussion. Ultimately evaluate whether limited staff can make useful checks and act appropriately.
Judge flight efficiency, application quality and crop outcomes separately
Keep three layers separate. First, operational completion: which fields were finished and how much time and labor were used. Second, application quality: delivery to the target, uniformity and off-target effects. Third, delayed agronomic outcomes such as pests, growth, yield and quality. Assign dates and accountable reviewers to each layer, and label unmeasured outcomes explicitly. Better deposition in a year with little disease does not establish a yield gain. Conversely, unchanged yield can coexist with lower operator workload or more reliable completion within the crop window; evaluate those benefits separately.
Predefine adoption, conditional continuation and rejection
Before testing, agree required quality, acceptable workload, operating conditions and budget limits. Discuss tolerable peak-day delays, long replenishment cycles and weeks with repeated faults instead of relying on averages alone. Conditions absent from a small pilot are unassessed, not proven safe or reliable. Decisions can be conditional: adopt on level fields, test slopes further and reject a particular task. Include abandoned jobs and rework in economics, and specify what must be rechecked after operator or software changes. This turns a completed pilot into an ongoing quality process.