Detecting selective logging in tropical forests with optical satellite data: an experiment in Peru shows texture at 3 m gives the best results

Selective logging is known to be widespread in the tropics, but is currently very poorly mapped, in part because there is little quantitative data on which satellite sensor characteristics and analysis methods are best at detecting it. To improve this, we used data from the Tropical Forest Degradation Experiment (FODEX) plots in the southern Peruvian Amazon, where different numbers of trees had been removed from four plots of 1 ha each, carefully inventoried by hand and terrestrial laser scanning before and after the logging to give a range of biomass loss (∆AGB) values. We conducted a comparative study of six multispectral optical satellite sensors at 0.3–30 m spatial resolution, to find the best combination of sensor and remote sensing indicator for change detection. Spectral reflectance, the normalised difference vegetation index (NDVI) and texture parameters were extracted after radiometric calibration and image preprocessing. The strength of the relationships between the change in these values and field‐measured ∆AGB (computed in % ha−1) was analysed. The results demonstrate that: (a) texture measures correlates more with ∆AGB than simple spectral parameters; (b) the strongest correlations are achieved for those sensors with spatial resolutions in the intermediate range (1.5–10 m), with finer or coarser resolutions producing worse results, and (c) when texture is computed using a moving square window ranging between 9 and 14 m in length. Maps predicting ∆AGB showed very promising results using a NIR‐derived texture parameter for 3 m resolution PlanetScope (R2 = 0.97 and root mean square error (RMSE) = 1.91% ha−1), followed by 1.5 m SPOT‐7 (R2 = 0.76 and RMSE = 5.06% ha−1) and 10 m Sentinel‐2 (R2 = 0.79 and RMSE = 4.77% ha−1). Our findings imply that, at least for lowland Peru, low‐medium intensity disturbance can be detected best in optical wavelengths using a texture measure derived from 3 m PlanetScope data.