USE OF ARTIFICIAL NEURAL NETWORKS OF DIFFERENT ARCHITECTURE AND LEARNING RATE TO PREDICT SOIL HUMUS CONTENT USING NORMALIZED DIFFERENCE VEGETATION INDEX

As far as remotely sensed normalized difference vegetation index (NDVI) applications extend to more specific subjects than plants vegetation conditions assessment, studies are conducted to adopt this spatial index for evaluation of soil properties, e.g., soil nutrients and organic matter content, electrical conductivity, pH, etc. [1 – 3]. Our studies are directed to the development of mathematical model for derivation of soil humus content using normalized difference vegetation index values. Although some success has been achieved in this field by the means of regression analysis, the prediction accuracy and model fitting quality are still insufficient to provide it for practical implementation [4]. As it is known that artificial neural networks (ANN) in many cases provide much better results than traditional regression analysis, the study was performed with different ANN architecture and learning rates to establish the relationship and improve the quality of soil humus content prediction based on the values of spatial vegetation index [5]. The study was carried out in the fallow-field period of 2022-2023 for the fields of Kherson oblast. 1478 bare-soil NDVI values were collected and generalized by the major natural and agricultural districts of the region according to the zoning [6]. Average NDVI values were associated with the soil humus content of the corresponding district of Kherson oblast. The soil humus content was taken from the results of the regional soil surveys and investigations conducted in the Institute of Climate-Smart Agriculture


Lykhovyd Pavlo Volodymyrovych candidate of agricultural sciences, postdoc student Institute of Climate-Smart Agriculture of NAAS, Ukraine
As far as remotely sensed normalized difference vegetation index (NDVI) applications extend to more specific subjects than plants vegetation conditions assessment, studies are conducted to adopt this spatial index for evaluation of soil properties, e.g., soil nutrients and organic matter content, electrical conductivity, pH, etc. [1][2][3].Our studies are directed to the development of mathematical model for derivation of soil humus content using normalized difference vegetation index values.Although some success has been achieved in this field by the means of regression analysis, the prediction accuracy and model fitting quality are still insufficient to provide it for practical implementation [4].As it is known that artificial neural networks (ANN) in many cases provide much better results than traditional regression analysis, the study was performed with different ANN architecture and learning rates to establish the relationship and improve the quality of soil humus content prediction based on the values of spatial vegetation index [5].
The study was carried out in the fallow-field period of 2022-2023 for the fields of Kherson oblast.1478 bare-soil NDVI values were collected and generalized by the major natural and agricultural districts of the region according to the zoning [6].Average NDVI values were associated with the soil humus content of the corresponding district of Kherson oblast.The soil humus content was taken from the results of the regional soil surveys and investigations conducted in the Institute of Climate-Smart Agriculture (former Institute of Irrigated Agriculture).In total, 34 data pairs "NDVI (pts.)-soil humus content (%)" were created and analyzed.
Statistical analysis in the pilot study was conducted by the means of nine regression model types [4].In this study we applied ANN approach for the same dataset.Tiberius XL software was used for ANN modeling [7].The settings included different architecture of the network (with the minimum 1 and the maximum 5 neurons in the hidden layer), and different values of the learning rate (from 0.7 to 1.0).The training was conducted within 1000 epochs.The results of the modeling were analyzed using coefficients of correlation (by Pearson), coefficient of determination, and mean absolute percentage error (MAPE).The interpretation of СЕКЦІЯ XI.АГРАРНІ НАУКИ ТА ПРОДОВОЛЬСТВО MAPE values was conducted according to [8], while the coefficients values were interpreted according to [9].
The results of the modeling using different ANN are presented in the Table 1.The best fitting quality and prediction accuracy are in the variant of ANN with 5 neurons in the hidden layer and learning rate of 0.8.Less number of neurons in the hidden layer, as well as higher learning rate, resulted in statistically significant decrease of the fitting quality and accuracy.The best variant of ANN model surpassed the best regression model (cubic) both in terms of curve approximation (R 2 of 0.29 vs. 0.25) and prediction accuracy (MAPE of 12.28% vs. 13.77%)[4].In general, ANN-based algorithm (Tiberius XL uses back-propagation of the error) is superior traditional regression analysis in case of humus content prediction by the values of NDVI.However, the drawback is that we can use the results only within Tiberius XL model, as it is impossible to derive the mathematical equation out of the ANN.
As a solution, we tried a combined ANN-regression approach, that provides an opportunity of mathematical expression of the ANN-based model, however, with a bit lower accuracy.The methodology of such a combination of two mathematical techniques was successfully introduced in the study of beans yield prediction [10].In the current study, there also was slight improvement in the model quality: MAPE of 13.22%, Pearson's correlation coefficient of 0.54, and coefficient of determination -0.29.The new combined model for humus content in the soils of Kherson oblast looks as follows (1): y = 43.164− 1168.8x+ 12122x 2 − 54272x 3 + 89044x 4  (1) where: y -humus content in the soil, %; x -bare-soil NDVI value for the area.According to [9], the combined mathematical model has moderate fitting quality, while the predictive accuracy of the model with accordance to [8] is good.Further extension of the number of inputs, and inclusion of new "NDVI (pts.)-soil humus content (%)" pairs in the mathematical model may result in additional enhancement of the model quality and precision.Even in its current state the model certifies about promising prospects for implementation of spatial NDVI in soil surveys.

Table 1
Results of modeling soil humus content by the bare-soil NDVI values in ANN with different architecture and learning rate