Apple trees commonly require the removal of excessive flowers by thinning to produce marketable fruit. Estimating the flower counts and phenology is important for chemical thinning decisions. Farmers generally inspect flowers in randomly sampled trees within the orchard, which is time-consuming, labour-intensive, and unreliable. Existing algorithms for estimating flower density have proven to be of low efficacy. No published algorithms for estimating flower phenology distributions exist. This thesis first presents a novel pixel-level flower segmentation algorithm named FCNs-Edge to estimate flower density in apple orchards, which showed an improvement over State-Of-The-Art (SOTA) methods. A side-view apple flower density map is then generated for a variable rate chemical sprayer. The FCNs-Edge is then extended to multiple classes - green, pink and white flowers, demonstrating the feasibility to estimate three flower stage distributions grouped by colour. A novel method named DeepPhenology is proposed to estimate phenology distributions over eight apple flower stages by building direct relationships with real counts in field. The proposed method removes the need to label images, which overcomes difficulties in distinguishing overlapping or hidden flower clusters on 2D imagery. The proposed model was shown to outperform a SOTA object detection model for this task. This thesis finally delivers a novel data-centric analysis of on-tree fruit detection based on a SOTA object detection model using public datasets; of interest given the wide usage of deep learning in agriculture which lacks in-depth data-centric analysis. The results indicated that 2500 annotated objects are generally sufficient for single-class fruit training. A novel similarity score is also proposed to easily predict the AP score without any training. The proposed FCNs-Edge and DeepPhenology have demonstrated accurate, robust and efficient estimation of apple flower density and phenology distributions. Such methods can replace traditional human inspection, significantly reduce labour requirements and thus reduce costs. Such estimation of flower information can also provide tree-level decision support to manage the variable growth between trees. Ultimately, the data-centric analysis of on-tree fruit detection will help practitioners better understand the influence on the training accuracy from data-centric attributes, and thus prepare better quality datasets and achieve higher training accuracy.