Prediction of ‘Gigante’ cactus pear yield by morphological characters and artificial neural networks

ABSTRACT Estimating cactus pear yield is important for the planning of small and medium rural producers, especially in environments with adverse climatic conditions, such as the Brazilian semi-arid region. The objective of this study was to evaluate the potential of artificial neural networks (ANN) for predicting yield of ‘Gigante’ cactus pear, and determine the most important morphological characters for this prediction. The experiment was conducted in the Instituto Federal Baiano, Guanambi campus, Bahia, Brazil, in 2009 to 2011. The area used is located at 14° 13’ 30” S and 42° 46’ 53” W, and its altitude is 525 m. Six vegetative agronomic characters were evaluated in 500 plants in the third production cycle. The data were subjected to ANN analysis using the R software. Ten network architectures were trained 100 times to select the one with the lowest mean square error for the validation data. The networks with five neurons in the middle layer presented the best results. Neural networks with coefficient of determination (R2) of 0.87 were adjusted for sample validation, assuring the generalization potential of the model. The morphological characters with the highest relative contribution to yield estimate were total cladode area, plant height, cladode thickness and cladode length, but all characters were important for predicting the cactus pear yield. Therefore, predicting the production of cactus pear with high precision using ANN and morphological characters is possible.

摘要：估算梨果仙人掌的产量，对中小型农村生产者的生产规划具有重要意义，尤其在巴西半干旱区这类气候逆境环境中。本研究旨在评估人工神经网络（Artificial Neural Networks, ANN）预测‘巨无霸’梨果仙人掌产量的潜力，并明确该预测任务中关键的形态性状。
试验于2009年至2011年在巴西巴伊亚州的巴伊亚联邦研究所瓜纳比校区开展。试验地块位于南纬14°13′30″、西经42°46′53″，海拔525米。研究在第三个生产周期内，对500株植株的6个营养农艺性状进行了测定。
试验数据采用R软件开展人工神经网络分析。共构建10种网络架构，每种架构均重复训练100次，以筛选出验证集均方误差最低的最优模型。其中，中间层包含5个神经元的网络表现最优。最终调优后的神经网络模型决定系数（R²）达0.87，可有效适配样本验证，保障了模型的泛化能力。
对产量估算相对贡献最高的形态性状依次为总茎节（cladode）面积、株高、茎节厚度与茎节长度，但所有性状均对梨果仙人掌的产量预测具有重要价值。综上，利用人工神经网络结合形态性状可精准实现梨果仙人掌的产量预测。