Prédiction des rendements agricoles du maïs et du soya, et du déficit en azote du maïs à l'aide d'images aéroportées et d'un réseau de neurones à rétropropagation

Abstract

This thesis explores the suitability of the multi-layer perceptron (MLP) neural network for prediction of corn and soybean yield and for classification of nitrogen corn stressed vegetation. The results show while using a combination of vegetation and texture indices, and topographic data, we can successfully predict corn yield with MLP. The regression coefficient between observed and predicted yield for corn and soybean were respectively 92% and 43% for the middle season image only. This research also indicates that the information from an image is more important when it is acquired during the period of maximum crop biological activity. Thus, MLP developed on the middle season predicts the corn and soybean yield with respective R[superscript 2] of 89% and 40%. The degree of relationship between the inputs and the outputs and the quality of the data are very important, because they determine the capabilities of the MLP. When the generalization capabilities of the MLP models were tested on a non-corrected crop yield dataset, the results drop to 45% and 36% for corn and soybean respectively. The MLP demonstrated also its capability to discriminate corn nitrogen status during the growing season. MLP models based on image spectral and textural indices reach a Kappa coefficient of 72% for middle season image and 81% for the combined multidate images. Image texture features also provide useful complementary information for the discrimination of different nitrogen stress levels. The best MLP for predicting corn nitrogen status was the one that integrates image features and topographic parameters. It outperforms the first one with an accuracy of 90% for middle season image and 95% for all two images. Evidence that topographic data are a critical discriminatory information source is both obvious and strong. In consequence, crop physiological status monitoring systems require taking into account the combined effects of soil background and canopy architecture. This thesis have highlighted that MLP has a strong potential for detecting corn nitrogen stressed vegetation and that they would help farmers to better manage crop status during the growing season, when there is still time to respond to problems. The other contributions of this study are the development of crop yield data correction and filtering algorithm and the identification of discriminant analysis as the best method for the MLP inputs selection"--Résumé abrégé par UMI.