Artificial Neural Networks Apply To Soil Quality Observation

Soil quality, which is the most direct feedback from soil productivity, can reflect properties of soil comprehensively. It is also the most sensitively dynamic indicator of soil conditions and reveals the influence of human’s activities on soil environment. In recent years, the need of foodstuff grows persistently due to the continual increase of population and the rapid development of economy. Additionally, the degeneration of ground resource and the short of reserved ground resource tense the relationship between human and the ground. Soil quality observation and prediction play an important role on relieving human and ground’s relationship and boosting rational utilization of ground resource and realizing sustainable development of agriculture.Artificial neural networks are a kind of mathematic model aimed on information processing. They are called neural networks all because of its structure, which is similar with connection structure of brain synapses. BP neural networks are the most widely used algorithm. This thesis, taking the ecologic experiment field located at Qianyanzhou, Jiang xi as an example, develops a new method for soil quality observation, which is a model using BP neural networks for soil quality assessment and observation. We expecting this method will provide a scientifically reliable evidence on improving farmland productivity and proper use of soil resource, helping realize the sustainable development of agriculture.This thesis optimizes the BP neural networks soil quality observation model on the following points using MATLAB toolbox. First is the special training sample. This thesis builds special sample based on soil nutrient property and the classification criteria of soil nutrient, which was made in the national second soil general survey. We expand the special sample to a hundred sets of data according to some algorithm. New special sample is obtained by subdividing the classification using the special sample already built. This step accomplishes detailed research about research field’s soil quality. Secondly, we adopt the trainlm algorithm to train sample data. There are three algorithms widely used in BP networks. trainlm gets faster convergence speed. Meanwhile it has adaptive-learning capability and can justify data momentum. They have already been validated and match the error demand on the trained target. Considering the assessment rank range of the result of soil quality simulation is below one, we choose the Sigmoid function, which is continuous differentiable, as the driving function. Then we compared the output errors. Finally, optimize the choose of hidden layers and nodes. Associate with the features of input and output data, we first figure out the number of input layer nodes and output layer nodes. Then test the different numbers of nodes in the networks by using several usual formulas and choose a three-layers BP network structure with 7 nodes in the hidden layer. It presents a faster converging speed and higher accuracy in data training and simulation results. The results of the simulation show that the BP networks model this thesis builds is close with the expectation on soil quality observation and assessment. It gets higher accuracy of net errors. The model is meaningful in scientific management and proper use of soil.