Random forest is a classification technique widely used in remote sensing. One of its advantages is that it produces an estimation of classification accuracy based on the so called out-of-bag cross-validation method. It is usually assumed that such estimation is not biased and may be used instead of validation based on an external data-set or a cross-validation external to the algorithm.In this paper we show that this is not necessarily the case when classifying remote sensing imagery using training areas with several pixels or objects. According to our results, out-of-bag cross-validation clearly overestimates accuracy, both overall and per class. The reason is that, in a training patch, pixels or objects are not independent (from a statistical point of view) of each other; however, they are split by bootstrapping into in-bag and out-of-bag as if they were really independent. We believe that putting whole patch, rather than pixels/objects, in one or the other set would produce a less biased out-of-bag cross-validation. To deal with the problem, we propose a modification of the random forest algorithm to split training patches instead of the pixels (or objects) that compose them. This modified algorithm does not overestimate accuracy and has no lower predictive capability than the original. When its results are validated with an external data-set, the accuracy is not different from that obtained with the original algorithm.We analysed three remote sensing images with different classification approaches (pixel and object based); in the three cases reported, the modification we propose produces a less biased accuracy estimation. HighlightsRandom forest out-of-bag cross validation may be biased in RS classification.A modification of random forest is proposed to perform a different out-of-bag CV.The Out-of-bag-CV of the modified algorithm equals a leave-one-patch-out CV.The modification does not cause a loss in the predictive capability of the algorithm.

Populations of the Mediterranean fan musselPinna nobilishave progressively decreased over the last decades as a result of anthropogenic activities. The rate of decline has strongly increased since 2016, when a mass mortality event triggered by the parasiteHaplosporidium pinnaeoccurred, and evidence exists thatMycobacteriumspecies may also have played a major role in the event. Indeed, the epidemic has spread throughout the Mediterranean, although coastal lagoons seem to offer a degree of ‘resistance’ against the parasite. In the early 1980s,P. nobilisappeared in the Mar Menor lagoon and rapidly became an important component of the benthos. However, colonization of the lagoon by the fan mussel was cut short in 2016 when a massive mortality event occurred, possibly as a consequence of the environmental collapse that occurred in the lagoon, parallel to the mortality that the species suffered in the Mediterranean that same year. In this study, we estimated the spatial distribution ofP. nobilisin the Mar Menor for 3 periods: 2003-2004, 2013 and 2016. The first 2 periods use published data, and the last period uses data collected in a new campaign. The probability of occurrence for the 3 periods was estimated using random forest and random forest regression-kriging models. The main environmental variables that determined the dispersion and colonization of the bivalve in the lagoon before 2016 are also identified.

Large ensembles of climate models are increasingly available either as ensembles of opportunity or perturbed physics ensembles, providing a wealth of additional data that is potentially useful for improving adaptation strategies to climate change. In this work, we propose a framework to evaluate the predictive capacity of 11 multi-model ensemble methods (MMEs), including random forest (RF), to estimate reference evapotranspiration (ET${}_0$) using 10 AR5 models for the scenarios RCP4.5 and RCP8.5. The study was carried out in the Segura Hydrographic Demarcation (SE of Spain), a typical Mediterranean semiarid area. ET${}_0$ was estimated in the historical scenario (1970–2000) using a spatially calibrated Hargreaves model. MMEs obtained better results than any individual model for reproducing daily ET${}_0$. In validation, RF resulted more accurate than other MMEs (Kling–Gupta efficiency (KGE) $M=0.903$, $SD=0.034$ for KGE and $M=3.17$, $SD=2.97$ for absolute percent bias). A statistically significant positive trend was observed along the 21st century for RCP8.5, but this trend stabilizes in the middle of the century for RCP4.5. The observed spatial pattern shows a larger ET${}_0$ increase in headwaters and a smaller increase in the coast.

Land cover classification in semiarid areas is a difficult task that has been tackled using different strategies, such as the use of normalized indices, texture metrics, and the combination of images from different dates or different sensors. In this paper we present the results of an experiment using three sensors (Sentinel-1 SAR, Sentinel-2 MSI and LiDAR), four dates and different normalized indices and texture metrics to classify a semiarid area. Three machine learning algorithms were used: Random Forest, Support Vector Machines and Multilayer Perceptron; Maximum Likelihood was used as a baseline classifier. The synergetic use of all these sources resulted in a significant increase in accuracy, Random Forest being the model reaching the highest accuracy. However, the large amount of features (126) advises the use of feature selection to reduce this figure. After using Variance Inflation Factor and Random Forest feature importance, the amount of features was reduced to 62. The final overall accuracy obtained was 0.91 ± 0.005 ($\alpha$ = 0.05) and kappa index 0.898 ± 0.006 ($\alpha$ = 0.05). Most of the observed confusions are easily explicable and do not represent a significant difference in agronomic terms.