Enhancing Orchard Classification Accuracy through Object-Level Confidence Metrics: A Case Study in Automatic Orchard Identification in Valencia, Spain

Effective agricultural management policies rely heavily on accurate crop identification across various scales. This is particularly challenging in highly fragmented agricultural landscapes, such as those found in many European regions, particularly for fruit tree orchard identification. While solutions exist, the confidence in individual orchard classification is often overlooked, despite its importance in enhancing classification accuracy (precision, recall and specificity).

Several confidence metrics at pixel level have been proposed by estimating the probabilities of a pixel to belong to each of the possible classes. The higher the probability of class membership for a given class, the greater the confidence associated with that class. In this sense, a measure of confidence can be based in the differences of probability between the first two highest values (sometimes called the distance to the second cluster).

This study introduces an innovative methodological approach to build a classification confidence metric at object (orchard) level. Once segmentation is completed, all pixels whose confidence is not above a certain threshold are masked out. Then, each orchard is initially assigned to a class by computing the mode of the unmasked pixels inside its perimeter. In a subsequent step, a confidence metric at orchard level is estimated, based on the number of mode class pixels, the total number of pixels completely inside the orchard, and the proportion of the mode pixels and unmasked pixels within the orchard. This confidence metric allows for a balance between increased precision and a reduction in the number of classified orchards (those with insufficient confidence in their classification).

The proposed method, fully implemented in Google Earth Engine, was tested in a highly fragmented area in Valencia (Spain). The system’s performance was assessed by using a Random Forest classification algorithm on Fourier coefficients of spectral indexes time-series at pixel-level plus a specific spatial cross-validation procedure. By setting a 70% orchard classification confidence level, the mean overall accuracy increased from 88.74 ± 3.03% to 93.58 ± 2.85%, and the Kappa index from 0.78 ± 0.06% to 0.87 ± 0.05%, albeit at the cost of leaving 12.60 ± 7.18 % of orchards unclassified.