Verifying a 640-Hectare Deforestation Alert in the Chure
A national forest monitoring unit compares Landsat (2019) and Sentinel-2 (2023) imagery and reports 640 ha of forest loss inside a 4,000 ha community forest in the Chure range; the map uses an NDVI threshold and a 0.1 ha minimum mapping unit. The division forest office sends a team to 25 randomly selected 'loss' polygons: 3 are genuine new clearings with stumps and charcoal kilns, 5 are old agricultural enclaves that were never forest, 14 are degraded shrubland that the classifier called forest in 2019, and 3 are cloud and shadow errors. Meanwhile the CFUG has records of 40 ha of legal thinning and no reported encroachment. The ministry wants a number for its national report in three weeks, and a journalist has already asked for the shapefile.
1. Assess the accuracy of the 640 ha figure. Compute what can be computed and explain what is still missing.
Read the worked answer
Precision of the 'loss' class: 3 genuine out of 25 checked = 12%. Most mapped loss is classification error: 14 polygons were degraded shrub misclassified as forest in 2019, 5 were long-standing enclaves, and 3 were cloud or shadow. The 640 ha cannot be presented as deforestation inside the CF; the true area is likely far smaller, though the 3 confirmed clearings are real and require enforcement.
What is missing: precision alone is not accuracy. No sample was drawn from the 'no loss' or 'stable forest' classes, so overall accuracy and recall cannot be computed, and we cannot say how much real loss was missed. The sample is also small: with 3 successes in 25, the 95% confidence interval on precision is roughly 3–31%.
Method mismatch: the NDVI threshold, the 0.1 ha minimum mapping unit and the seasonally mismatched images do not correspond to Nepal's forest definition (canopy cover of at least 10%, area of at least 0.5 ha, height of at least 5 m). Much of the apparent change is a baseline problem, not new loss.
2. Design the verification sampling and classification protocol you would insist on before any figure is published.
Read the worked answer
Use stratified random sampling with strata from the change map (loss, stable forest, degraded, non-forest), allocating samples proportional to area and expected variability. Plan at least 50–100 points per class for a usable confusion matrix, with a minimum of 30 in the loss class.
At each point: a 30 m x 30 m plot, GPS fix, photographs, canopy cover by densiometer, evidence of stumps or charcoal, and land-use history cross-checked with CFUG records.
Fix the protocol: apply Nepal's forest definition; set the minimum mapping unit at 0.5 ha; use the same season for both dates (post-monsoon, November–January); mask cloud and shadow with the quality band and add Sentinel-1 SAR for monsoon-independent confirmation; calibrate the NDVI threshold against field plots or replace it with a supervised classifier using training data; separate degradation from deforestation; and reconcile detected change with DFO records of legal thinning.
Report a confusion matrix with overall accuracy, user's and producer's accuracy, and area estimates with confidence intervals.
3. What would you report to the ministry and say to the journalist, and how would you handle the shapefile request?
Read the worked answer
To the ministry: report an interim figure with a range, not a single number. State the confirmed clearings and their area, the precision estimate and its confidence interval, and flag the degraded-shrub reclassification as the dominant error. Do not transmit 640 ha as deforestation; explain that the three-week deadline is achievable only as a preliminary estimate with stated uncertainty.
To the journalist: provide the map together with the accuracy statement and the field verification results. Explain that the method detects change in greenness, not legal forest loss, and that legal thinning and degraded shrub are included. Avoid a defensive tone; the confirmed clearings are a real enforcement issue and should be acknowledged.
Shapefile: release it with metadata, the accuracy statement and a written caveat, or release a summarized version if it is likely to be misread as a parcel-level legal finding. Do not release a parcel map that could be used to accuse individual households without field verification.
Follow-up: refer the 3 confirmed clearings to the DFO for investigation, and correct the 2019 baseline before the next national report.