ABSTRACT The Caatinga biome in Brazil comprises the largest and most continuous expanse of the seasonally dry tropical forest (SDTF) worldwide; nevertheless, it is among the most threatened and least studied, despite its ecological and biogeographical importance. The spatial distribution of volumetric wood stocks in the Caatinga and the relationship with environmental factors remain unknown. Therefore, this study intends to quantify and analyze the spatial distribution of wood volume as a function of environmental variables in Caatinga vegetation in Bahia State, Brazil. Volumetric estimates were obtained at the plot and fragment level. The multiple linear regression techniques were adopted, using environmental variables in the area as predictors. Spatial modeling was performed using the geostatistical kriging approach with the model residuals. The model developed presented a reasonable fit for the volume m3 ha with r2 of 0.54 and Root Mean Square Error (RMSE) of 10.9 m3 ha–1. The kriging of ordinary residuals suggested low error estimates in unsampled locations and balance in the under and overestimates of the model. The regression kriging approach provided greater detailing of the global wood volume stock map, yielding volume estimates that ranged from 0.01 to 109 m3 ha–1. Elevation, mean annual temperature, and precipitation of the driest month are strong environmental predictors for volume estimation. This information is necessary to development action plans for sustainable management and use of the Caatinga SDTF in Bahia State, Brazil.

Resumen La selva amazónica presenta diferentes estratos forestales, debido a su estructura heterogénea. En el que estos estratos pueden variar en altos, medios y bajos. El conocimiento de los diferentes patrones de estructuras verticales que se encuentran en el bosque es extremadamente importante para comprender la dinámica de la vegetación, lo que influye en las estrategias de conservación del bosque. Con el fin de optimizar el proceso de clasificación de los diferentes tipos de estratos, el objetivo del presente trabajo fue utilizar redes neuronales artificiales (RNA) para clasificar estos estratos. Se utilizaron dos algoritmos de resilient propagation (Rprop+ y Rprop-), en cuatro configuraciones diferentes de variables de entrada. El entrenamiento y la prueba de los ocho modelos de RNAs se realizaron utilizando el software R. Los modelos se evaluaron mediante una matriz de confusión. En qué modelos con entradas: HT, DAP y QF; HT, DAP y solo HT del algoritmo Rprop + obtuvieron 100% de aciertos en la clasificación de estrato, presentando una alta tasa de aprendizaje, confiabilidad y generalización de datos.

Abstract The Amazon forest presents different forest strata, due to its heterogeneous structure. In which these strata can vary in upper, middle and lower. Knowledge about the different patterns of vertical structures found in the forest is extremely important for understanding the vegetation dynamics, influencing forest conservation strategies. In order to optimize the process of classifying the different types of strata, the objective of the present work was to use artificial neural networks (ANNs) to classify these strata. Two resilient propagation algorithms (Rprop + and Rprop-) were used, in four different configurations of input variables. The training and testing of the eight RNA models were performed using the R software. The models were evaluated using a confusion matrix. In which models with inputs: HT, DAP and QF; HT, DAP and only HT from the Rprop + algorithm obtained 100% correct answers in the classification of strata. Demonstrating a high rate of learning, reliability and generalization of data.