ABSTRACT

This work is an attempt to improve models for forecasting phonology phases, yields and productions of perenial crops, apples in particular. Models, using polynomial functions and artificial neural networks, were developed, tested and compared. The combination of the results provided by these two approaches was also performed by the use of linear programming with the purpose of minimizing prediction errors.

A spacial analyses methodology was also developed in order to characterize the farm with the objetive of discriminating cultivated areas, in terms of species and cultivars. Digital images, generated by a compact airborne spectrographic imager (CASI), and aerophotos from an aerophotogrammetric survey were used for such purpose. The area of study has 920ha and is part of Fazenda Bom Futuro, located in Fraiburgo, SC.

The models have agrometeorological basis combined with the productive potential of each cultivar. The variables used are related to the concepts of degree-days, chilling temperature units and water balance. Several forms of polynomial functions and a back-propagation type of neural network were used.

Considering the dynamics of agriculture and the convenience of visualizing spatial data, a geographic information system (SPANS - Spatial Analyses System) was used to integrate and manipulate the maps, restituted and digitized, and the alphanumeric data. Eight spectral bands of the scanner, from the visible to the near infrared (465-790nm), were used with a spatial resolution of 3.5 x 3.5 m, corresponding to six flight lanes of 512 pixels and lateral coverage of 30%. Aerophotogrammetric and spectrografic coverages were performed concomitantly, on november 05, 1992. An ARIES-380 workstation was used to carry out the supervised and automatic classification of the area of study, adopting the maximum likelihood criterion. A detailed inventory of the farm, including a map showing the orchards, the digital aerophotogrammetric restitution and the aerophotos were used as terrestrial thuth. GIS was also used to determine the areas for each class by overlapping the mosaic of classified images and the corrected map showing the orchards (25) and their subdivisions (276).

The models proposed clearly indicate that climate variables linked to energy are decisive for inducing flowering. These variables are expressed as the sum of chilling hours, the sum of degree-days, the number of hours of sunshine and the average monthy temperatures. Degree-days was the most important factor for predicting harvesting date. In general, a linear type of response was observed for most estimators used for predicting flowering and harvesting dates and yield.

Neural models and polynomial functions had a similar performance in terms of their ability to predict phenoligical events, yields and production. The number of neurons in the hidden layer 1, during the trainning phase of the network, varied 2 to 10, corresponding to 0.5 to 1.5 times the number of inputs and outputs. It became clear that the fit of the models depends more on the quality and representativity of the data than on their quantity, indicating that a good forecast can be reached using a relatively small number of replications (less than 50 crop-years). The use of linear programming to combine the results of polynomial and neural models provided a smaller relative error in terms of forecast than the individual models, Depending on the model and cultivar, a good forecast can be provided 120 days before the event. Obviously, the margin of error diminishes as this interval becomes smaller.

CASI images, under the conditions of this work, were able to discriminate four apple cultivars (Gala, Golden Delicious, Belgolden and Fuji + Granny Smith) in commercial orchards with low density (<1000 plants/ha), small areas (<3ha) and consortiated planting system. Such performance demonstrates the potential for using this type of sensor under similar field conditions.

A combination of accurate and adequately scaled (1:5000) maps with classified images from CASI, cross-analyzed using GIS, privide an objetive and safe discrimination of cultivated areas, in terms of apple cultivars.